Six months later, one has changed how they work. The other has collected a folder full of clever prompts.

That difference is becoming the real AI gap.

It isn’t an access gap. Access still matters, especially inside organizations with strict security rules or limited budgets. But access alone is becoming less useful as an explanation for why one person gets durable value from AI and another doesn’t.

It isn’t an interest gap either. Plenty of people are interested. They read the newsletters. They watch the demos. They can name the new models before the rest of the meeting knows there was a new model.

Then Monday arrives, and their work looks exactly the same.

I think of them as AI tourists.

That isn’t an insult. Tourists can be curious, thoughtful, and well informed. They visit the territory. They notice what’s new. They take pictures. They go home with stories.

Operators live there.

They know which road floods after a storm. They know which shortcut saves time and which one ends at a locked gate. They know which parts of the map are wrong because they’ve had to find their way back in the dark.

That kind of knowledge doesn’t come from access. It comes from reps.

And that’s why the next meaningful divide in AI will be a practice gap.

Awareness got cheap. Changed behavior didn’t.

The first phase of professional AI adoption was dominated by awareness.

Have you tried ChatGPT? Did you see the new image model? Do you know what an agent is? Which model is best? Can it connect to your files?

Those were reasonable questions when most professionals had barely touched the tools.

They’re weak questions now.

Gallup’s Q1 2026 survey of 23,717 employed U.S. adults makes the gap visible. Half of all employed adults said they used AI at work at least a few times a year. Narrow to a different group, employees inside organizations that had already adopted AI, and only about one in ten strongly agreed that it had transformed how work gets done. That doesn’t prove practice is the missing cause. It does show why access and activity are weak stand-ins for changed work. Source: Gallup, April 12, 2026

The stronger question is: what changed in your work?

Not what did you test once. Not what impressed you in a demo. Not what subscription appears on your expense report.

What do you do differently because AI is now part of the work?

Do you start from a clearer brief because you know vague intent produces expensive revision?

Do you compare outputs against an explicit standard instead of accepting the one that sounds polished?

Do you save the context that made a task work so you don’t rebuild the same explanation next week?

Do you know which tasks are worth bringing to AI and which ones cost more to supervise than they return?

Do you catch different mistakes now than you caught three months ago?

Those are signs of practice. The tool has moved from something you visit to something your working method has absorbed.

The audience signals from this newsletter keep pointing in this direction. A Note asking where junior professionals will get the reps that build senior judgment became one of the strongest recent discussion starters. Another Note about the difference between people interested in AI and people changing how they work pulled the same thread from the other end.

Those engagement numbers don’t prove subscriber growth, and they don’t constitute a scientific study. They show recognition. Readers saw a professional tension they already felt but hadn’t named.

Access is not adoption.

Interest is not absorption.

Usage is not fluency.

The missing variable is practice.

Most AI use doesn’t count as a useful rep

Here’s where the argument gets uncomfortable.

Using AI more often doesn’t automatically make you better at using AI.

A year ago I argued in this newsletter that the edge came from consistent daily use. That was wrong: the edge isn’t daily reps, it’s reviewed reps.

You can repeat a weak behavior one hundred times and become efficient at doing the wrong thing. Anyone who has ever developed a terrible golf swing can confirm the general principle. Repetition is generous that way. It will strengthen whatever you feed it.

AI produces the same problem at professional scale.

A lot of activity looks like practice without containing the parts that create learning.

False rep 1: chatting without a standard

You ask a model for ideas. It gives you ten. A few sound smart. You copy one into a document and move on.

What did you learn?

Maybe something. But unless you knew what a good answer needed to do, compared the options, and noticed why one survived, the interaction taught you very little about judgment.

It taught you AI can produce options.

You already knew that.

The rep begins when you can explain why option three fits the decision and option seven only sounds impressive.

False rep 2: collecting outputs

AI makes artifacts appear so quickly that artifact production feels like progress.

A strategy document exists. A spreadsheet has formulas. A deck has twenty slides. A workflow diagram has arrows pointing confidently at other arrows.

The existence of an output tells you almost nothing about whether the underlying work improved.

This is one of AI’s stranger effects. It reduces the effort needed to create the visible evidence of work. The visible evidence used to be a rough proxy for thought, because producing it took time. Now it can appear before the thinking is done.

If you judge the rep by what was generated, you’ll overestimate what was learned.

False rep 3: automating work you never understood

Automation is seductive because it creates a clean before-and-after story.

This took two hours. Now it takes ten minutes. Look at the savings.

Sometimes that’s exactly right.

Other times the two hours contained the part where you noticed exceptions, questioned assumptions, and learned what the work was doing. Remove the task too early and you remove the feedback that would have built your judgment.

The best example is junior work.

Junior assignments are often repetitive. They’re also where people learn what normal looks like. They see fifty contracts and start noticing the unusual clause. They build twenty forecasts and learn which assumption quietly controls the whole model. They sit through the first drafts, the corrections, and the slightly painful explanation of why the senior person rejected the version that looked fine.

The production wasn’t the only output. Judgment was another output.

If AI absorbs the production, the learning doesn’t automatically move somewhere else. Teams have to move it on purpose.

False rep 4: tool hopping

Every new model creates a tiny reset.

New interface. New claims. New benchmark charts. New thread explaining the secret way to use it. For a few days, testing the model feels like the work.

Then another model arrives.

This can create the illusion of rapid learning while preventing depth. You get good at first impressions and never stay long enough to understand failure patterns.

The scarce skill isn’t forming an opinion about a model after twenty minutes. It’s building a working method that survives when the model changes.

That requires a different kind of rep.

A real rep has four parts

A useful AI rep isn’t complicated, but it’s more demanding than sending a prompt.

It contains four things:

1. A consequential task

2. An explicit standard

3. A comparison against reality

4. A change to the next attempt

Remove any one of them and the learning gets weaker.

1. A consequential task

You learn faster when the work matters enough to expose a mistake.

This doesn’t mean giving AI control over the highest-risk task you have. It means choosing real work with a visible finish line.

Draft the client brief you need, not a pretend brief about opening a coffee shop.

Summarize the meeting whose decisions you must carry into next week, not a random transcript you found online.

Analyze the project status you’re responsible for, not a clean sample dataset with no consequences.

Real work contains friction that demos remove. Missing context. Conflicting goals. Odd exceptions. Stakeholders who use the same word to mean three different things.

That friction isn’t evidence that the tool failed. It’s the material practice works on.

2. An explicit standard

Before you ask AI to do the work, write down what good must mean.

Not “make it better.”

Better for whom? Better in what way? Better at what cost?

A usable standard can be simple:

• The recommendation must name one decision, not provide a menu.

• Every number that affects the decision needs a source.

• The summary must separate what the team decided from what remains open.

• The draft must sound direct without creating unnecessary conflict.

• The analysis must surface the strongest reason the preferred option could fail.

This standard does two jobs.

First, it gives the model a clearer target.

Second, it gives you something to inspect. Without a standard, polished output wins by default. With a standard, polish has to earn its place.

3. A comparison against reality

The fastest way to learn from an AI-assisted task is to compare what you expected, what the model produced, and what happened next.

Did the recipient understand the recommendation?

Did the data check out?

Did the meeting recap preserve the actual disagreement or smooth it into fake consensus?

Did the workflow save time after review, or only before review?

Did the output fail in a way you could have predicted from the missing context?

This is where judgment gets built. Not in the generation, but in the gap between the output and the world.

The gap tells you what the system didn’t know. Sometimes it lacked facts. Sometimes it lacked a standard. Sometimes the task itself was badly framed. Sometimes you delegated a decision you hadn’t made.

Each failure can become context for the next attempt, if you preserve it.

4. A change to the next attempt

A rep is complete when it changes the next rep.

You update the brief.

You save a good example.

You add a review question.

You narrow the task.

You decide the task shouldn’t go to AI at all.

That last outcome counts. Fluency includes restraint.

If every failed interaction ends with “the model isn’t good enough yet,” nothing accumulates. The tool becomes the sole explanation, even when the real failure came from missing context, weak standards, or a task with no stable finish line.

A practice loop converts experience into a better working system.

Task. Standard. Comparison. Change.

That’s a rep.

Practice changes shape as you get better

The first useful reps with AI tend to be obvious. You learn how much context a task needs. You stop asking for “a strategy” and start naming the decision, audience, constraints, and finish line. You learn that a long answer isn’t necessarily a complete one. You discover that the model’s first response is often best treated as material for inspection, not an answer wearing a suit.

Those are setup reps. They matter, but they’re only the first layer. The second layer is evaluation.

You stop focusing on whether the model can produce something and start asking whether you can judge what it produced. That’s a much higher bar.

If you ask AI to draft an email, you probably know whether the email sounds wrong. If you ask it to interpret a legal clause, recommend a medical decision, or analyze a technical system outside your expertise, the output can sound excellent while you have no reliable way to inspect it.

Fluency isn’t the confidence to use AI on harder tasks. It’s the accuracy to know where your own review ability runs out.

That recognition changes what you delegate.

You may use AI to organize the questions for an expert without asking it to replace the expert. You may use it to identify claims that need verification without trusting it to perform the final verification. You may ask it to generate alternatives while keeping the decision with the person who carries the consequence.

Those are authority reps.

The third layer is system design.

At this point, the question is no longer “How do I get a better answer?” It becomes “How should this work happen every time?”

You preserve the context that mattered. You define which sources count. You decide what the system may change and what it can only recommend. You create a review point where a failure is still cheap to catch. You record what happened so the next person doesn’t start from zero.

This is where individual fluency becomes organizational capability. One person with a good prompt is useful. A team with a shared standard, clear decision rights, and a learning loop can improve.

The final layer is restraint.

After enough reps, you get faster at saying no.

No, this task isn’t stable enough to automate.

No, we can’t verify this output at the speed it’s being produced.

No, this junior assignment contains learning we haven’t replaced.

Restraint isn’t a retreat from AI. It’s what happens when familiarity stops being confused with judgment.

Setup. Evaluation. Authority. System design. Restraint.

The shape of practice changes, but the loop stays the same. Real work creates evidence. Evidence changes the next decision.

Subscribe now

This is a free article. Subscribe to get more articles and resources like this.

The apprenticeship problem isn’t hypothetical work nostalgia

The hardest version of the practice gap appears when AI takes over junior work.

An August 2026 revision of a Stanford working paper used payroll data through June 2026 to examine early-career employment in occupations exposed to generative AI. Employment for workers ages 22 to 25 in the most exposed occupations stood 19% below the path implied by their less-exposed peers. The authors report that the divergence came mainly through reduced hiring, not increased separations. They also found no evidence of widespread economy-wide job displacement and describe the young-worker result as early descriptive evidence, not a causal estimate. The pattern weakens under some controls and includes signs of divergence that began before generative AI. Source: Stanford Digital Economy Lab, revised August 12, 2026

The counter evidence matters just as much. A 2023 NBER working paper studied 5,179 customer-support agents and found that AI assistance increased productivity by 14% on average and 34% for novice and lower-skilled workers. The authors also found suggestive evidence that the system helped newer workers learn the practices of stronger performers. AI can weaken apprenticeship when it removes the rep, or strengthen it when the rep becomes coached comparison. The outcome depends on the learning system around the tool. Source: Brynjolfsson, Li, and Raymond, NBER, 2023

There’s a lazy response to this concern: nobody should mourn repetitive work, so let AI remove the tedious parts and people can move to higher-value work. The first half is right. The second half hides the problem, because you can’t move someone to higher-value work if they haven’t built the judgment the work requires.

Senior judgment isn’t downloaded with the promotion. It’s compressed from exposure, correction, comparison, and consequence.

The old route was imperfect. It often confused unnecessary suffering with development. Plenty of junior work was just administrative tax wearing a character-building costume.

But some of it contained real learning. The challenge isn’t to preserve every tedious task. It’s to identify the rep inside the task and keep that part when production moves to AI.

If AI writes the first draft, the junior person can still own the comparison.

Give them three drafts, including the AI version, and ask which one they would defend.

If AI reviews the contract, the junior person can still investigate the exceptions.

Ask what clause the model missed and why that miss matters here.

If AI builds the forecast, the junior person can still test the assumptions.

Ask which input could flip the recommendation and what evidence would settle it.

The rep moves from production to judgment.

That can be a better learning system than the old one. But only if someone designs it.

Otherwise the organization gets faster output today and weaker judgment tomorrow. It celebrates the saved hours while quietly spending the apprenticeship system that created its senior people.

That’s a bad trade.

What teams should measure instead of logins

Organizations love metrics that are easy to collect.

How many employees activated the tool? How many prompts did they send? How many hours did the vendor estimate they saved? How many people completed the training?

These numbers can describe activity. They can’t tell you whether work changed for the better.

A stronger adoption review asks different questions:

Which recurring tasks changed?

Name the workflow, not the aspiration.

“The marketing team uses AI” isn’t an operating fact.

“The marketing team now produces the first campaign brief from a standard context pack, then reviews it against five documented criteria” is.

Which standards became explicit?

AI often forces teams to explain what they previously carried in their heads.

That’s one of its most valuable side effects.

If the team can now name why a good recommendation is good, the adoption created an asset beyond the output. It made judgment more transferable.

Which failures became easier to detect?

Mature use isn’t the absence of errors. It’s a better detection system.

Can the team recognize when an answer is plausible but unsupported? Can it identify stale context? Does it know when the task exceeded the model’s authority? Is there a clear escalation point?

What changed after the last review?

If the same failure happens every week, the team is using AI but not learning with it.

Practice leaves a trail. Updated instructions. Better examples. Narrower authority. Clearer finish lines. Retired workflows that did not earn their cost.

That trail is adoption.

A seven-day practice protocol

You don’t need a course, a new tool, or a thirty-day challenge to close your own practice gap.

You need one recurring task and a week of attention.

Day 1: Pick the task

Choose work you will do at least twice this week.

It should matter, but failure should be reversible. A meeting recap, weekly analysis, project brief, research synthesis, or first-pass review can work well.

Avoid a task you do once a quarter. You need another rep soon enough to apply what you learn. Map the Day 3 and Day 6 reps to two real occurrences of the task, so an event-driven task doesn’t leave you waiting for the literal day.

Day 2: Write the standard

Before using AI, define three to five conditions the result must meet.

Keep them observable.

“High quality” isn’t observable.

“Names the decision, cites the figures, preserves disagreement, and ends with an owner and due date” is.

Day 3: Run the task and mark where you stepped in

Use AI, but note where you had to step in.

What did you have to explain twice?

What looked finished but wasn’t?

Where did your own judgment matter most?

Don’t fix the process yet. Capture the friction first.

Day 4: Compare the result with reality

Check the output against the standard and against what happened after you used it.

Separate three failure types:

• Missing information

• Missing judgment

• Bad task design

This distinction matters. More context can fix the first. Examples and decision rules can help with the second. The third may require a different task boundary or no AI at all.

Day 5: Change one thing

Don’t rebuild the entire setup.

Add one example. Rewrite one instruction. Create one review question. Remove one unnecessary output. Narrow one permission.

One change makes it easier to see whether your diagnosis was right.

Day 6: Run the second rep

Repeat the task with the revised setup.

Did the named failure improve?

If yes, preserve the change.

If no, your explanation was wrong. That’s useful information too.

Day 7: Decide what earned permanence

At the end of the week, make one of four calls:

1. Keep the workflow.

2. Keep it with a stronger review step.

3. Narrow what AI is allowed to do.

4. Stop using AI for this task.

All four can be signs of growing fluency.

The goal isn’t to keep the tool busy. It’s to make the work better.

The risk: practice can make bad systems feel normal

There’s a serious objection to everything I’ve argued.

Practice doesn’t guarantee improvement.

People can normalize weak outputs. Teams can lower standards to match what the tool produces. Frequent users can become skilled at rationalizing errors because the system is now familiar. A workflow can feel mature because it has documentation, even when the documentation protects a bad premise.

That’s why the standard and comparison matter more than repetition.

A CHI 2025 study of 319 knowledge workers, based on 936 first-hand examples of AI-assisted work, found that higher confidence in generative AI was associated with less reported critical-thinking effort. Higher confidence in one’s own ability was associated with more. The study is correlational and self-reported, so it can’t prove the tool caused weaker thinking. It does show why familiarity can’t be the standard. Confidence needs an external check. Source: Lee et al., CHI 2025

A real practice loop keeps contact with reality. The client response. The corrected number. The missed exception. The decision that held up. The downstream work that did or did not improve.

If the loop only compares the model with itself, confidence grows faster than competence.

The point isn’t to become comfortable with AI but to become accurate about what it does well, what it does badly, and what your work requires from both of you.

The gap is built one reviewed rep at a time

There’s another practical consequence: a practice system needs a memory.

Not a transcript archive. A short record of what failed, what changed, and whether the change worked.

Without that record, the same lesson gets rediscovered by different people. With it, one person’s reviewed rep can improve the next person’s starting point.

That’s how individual practice begins to become team capability.

The most AI-fluent person in the room may not be the person who knows the most model names.

It may be the person who can say:

This task works well because we gave the system the right examples.

That task still needs human ownership because the failure is hard to detect.

This instruction is stale.

That output looks polished but misses the decision.

We tried this workflow twice, and the review cost erased the speed gain.

We moved this junior rep from production to comparison so the learning did not disappear.

That is not tool knowledge.

It is operating judgment.

And operating judgment is built the old-fashioned way: through real work, clear standards, honest comparison, and another attempt that is better because the last one happened.

Access opens the door.

Practice is what you build once you are through it.

If this gave you a better way to think about AI adoption, share it with someone else who needs to rethink the way they are approaching AI.

Share

Subscribe now

AI helped write this essay.

It interviewed me. It pushed back on my argument. It organized the research, compared sources, tested titles, and drafted the piece you’re reading.

Substack will now let you scan it and get a number. Go ahead. I’ll wait.

When that number comes back, you’ll know one thing about this essay, and you’ll still be missing two.

So: do you trust it less now? Does it deserve less of your attention? Did it become slop the moment I told you?

Whatever you just felt, hold onto it. We’re going to take it apart. Not to talk you out of it, but to find out which of three completely different questions you were actually answering.

Because AI slop is real, and I’m not here to defend it.

You’ve read those posts. Five points, same polished rhythm, nothing observed, no difficult choice made, no weak claim checked. The sentences look finished even when the thinking never started.

You’re right to protect your attention. It’s the only thing you own that doesn’t scale.

What you’re being handed now is a number that promises to sort the thinking from the filler before you have to read either one.

That isn’t what the number measures. And I’ll go further than that, because I think most of the people defending this feature and most of the people attacking it are having the wrong argument.

Substack shouldn’t have shipped it. Not because the detector is bad. I’m going to spend a good chunk of this essay showing you it’s better than its critics claim. Because a working detector, dropped into a social system, does something a broken one couldn’t: it converts a narrow technical fact into a verdict about a person, quietly, with nobody accountable for the result.

That’s a design decision, not a detection problem. And it’s the same failure this essay is about.

Subscribe now

The case for the scan is better than the jokes about it

On July 21, Substack launched a Pangram-powered AI scan. You can request a scan on posts, Notes, comments, and replies over 100 words, published from that date forward. The result is private from other readers, which is a narrower thing than private. The score goes to whoever asked for it. The text goes to a third-party vendor. Writers can attach a statement explaining how they work, run Pangram on their own drafts before publishing, and report and remove scans on their own work they believe are wrong. Two days after launch, following complaints, Substack added a way to switch scanning off for a given post without having to scan it first.

Substack isn’t claiming that all AI use is bad. The announcement says the opposite: not all AI work is slop, and not all slop uses AI.

The concern is expectation. You show up looking for a person’s thinking and spend your evening on text nobody thought about. Substack calls it “Claudefishing,” and says it isn’t going to wait until the app turns into LinkedIn before doing something.

And yes, the obvious joke is that Substack’s answer to too much AI was more AI. I laughed too. It still doesn’t settle anything. You shouldn’t have to guess whether the process matched what you thought you were getting.

But it took about a day to go from “people should know” to “now we know what it’s worth.”

, who writes The Pragmatic Engineer, called the launch a great initiative in a post on X. Then he said why, in a post Substack later included in its roundup of early reactions:

“When I know that something is AI-written, I just don’t take time to read it, because whoever produced did not take time to write it.”

He isn’t scolding anybody. He’s describing how he spends a finite day, and if you’ve ever closed a tab three sentences in, you recognize the impulse.

But look at what the sentence is standing on. When I know. Everything after those three words depends on the knowing, and the knowing is exactly what the scan is now offering to supply.

Nobody in this fight is being stupid

Spend an hour in the replies and you’ll find at least five arguments running at once. Every one of them is protecting something worth protecting.

The relationship. People subscribe to encounter a particular mind. Even excellent synthetic prose can break the thing the reader thought they were entering. Orosz is defending a version of this, and so is , who put it as “If you let AI do your writing, I will come to your house and kill you.” (I better lock my doors.)

The feed. Cheap generation can flood a community faster than humans can write for it. Disclosure gives readers a choice and protects the people doing slower work. That’s Substack’s case, and it’s a real one.

Due process. A probabilistic score turns into an accusation the moment somebody screenshots it, and the tool has documented failure modes. has the sharpest version of this, and we’ll get to his numbers.

Access. For a lot of writers, AI is the editor, the translator, the interface between a thought and a sentence that won’t come easily. writes about this from the inside, and I’ll get to her. A score says nothing about whether a person was there.

Consent. Even private, reader-initiated scanning changes the norms of a place, and it does something more concrete than that. points out that a reader’s click hands your text to a third-party vendor, which can override an AI-training opt-out you already made. Her word for that is that it isn’t informed consent. Writers are allowed to notice and not love it.

I’ll grant every one of those, and I’m going to end up arguing that Substack still got this wrong. Hold that.

Because look at what just happened. Those aren’t five answers to one question.

They’re five answers to three different questions, and the fight only looks unresolvable because nobody has separated them.

One number, three different questions

The fight keeps collapsing three questions into one:

1. Provenance. What role did AI play in making the text?

2. Value. Did the work give you anything worth having?

3. Accountable authorship. Who made, checked, and owns the consequential decisions?

A detector can help with the first. Only you can settle the second. The third depends on evidence the finished prose doesn’t carry.

Substack says as much itself. Its launch announcement concedes the tool can tell you AI was involved and nothing at all about the care behind it.

So the same reading can come back for two pieces when one started as a button press and the other started with an idea, ten interview questions, curated research, a redirection that threw out the first full draft, a claim that died on its own evidence, and a line-by-line read.

Both come back AI-assisted. Only one of them has an author.

Pangram earns a narrower job

The easy move here would be to say Pangram doesn’t work. The evidence doesn’t support that, and I’m not going to pretend otherwise.

An independent NBER working paper tested Pangram and three other detectors against 1,992 verified human passages across six genres, matched with text from four frontier models. Pangram produced near-zero errors on medium and long passages inside that benchmark, and it was the strongest of the four detectors the paper tested.

A peer-reviewed study in the International Journal for Educational Integrity went further in June. Across 160 constructed documents and 1,163 real master’s theses, Pangram correctly identified all 40 of the fully human papers, with no false positives on that set, and placed 92.5% of the hybrid and deliberately humanized documents in the right range. Turnitin classified every one of the fully AI papers as human. Pangram’s own weakest showing in that study was on those same fully AI papers, at 65% under the strict scoring.

I’d rather you hear the strongest version of that from me than from somebody accusing me of hiding it.

So the tool is good at the thing it does. What it does is narrower than the number implies.

Pangram’s own model card sorts passages into human-written, lightly AI-assisted, moderately AI-assisted, and AI-generated. The output is a band for how much AI assistance the passage resembles, not a reconstruction of who typed which words. Pangram’s own explanation of the method describes it as author identification, asking who a passage sounds like rather than counting anything. It’s closer to forensic stylometry than to a plagiarism checker.

Which is not how it gets described in public, including by Substack. Its own launch post said you’d be able to see an estimate of how much of the text “was written by hand or with AI assistance.” By hand. That reads like a word count. It isn’t one. The slippage starts at the announcement.

put numbers on the gap, in a piece published two days before Substack’s launch. He took a human paragraph he wrote in 2017, 239 words, and appended 71 words of ChatGPT. By word count that’s 23% machine. Pangram returned 100% AI written, high confidence.

Separately, a roughly 300-word section of an old post of his came back 100% AI with high confidence. He submitted the full 5,000-word essay it was taken from. That came back 100% human, with high confidence. Same words, opposite verdicts, depending on where you cut.

His conclusion is his title: he wouldn’t say Pangram is broken, he’d say it’s brittle. Useful as part of a broader effort. Never the last word on anyone. Though he does call one thing broken outright, and it’s the number itself.

Worth knowing who’s saying this. deBoer is not a detector skeptic. He hates AI writing, says so in the piece at length, and went in wanting the tool to work.

And notice what his test does to the number. If a document that’s 23% machine comes back “100% AI,” then the percentage isn’t measuring what almost every reader assumes it’s measuring.

That’s a writer stress-testing a tool, some of it while defending himself. Here’s what it looks like when it lands on someone who wasn’t looking for it.

The Atlantic reported that the journalist was accused on X of using AI for a Vanity Fair story. She denied it. Pangram’s CEO went and investigated, then said publicly that his own tool had gotten it wrong. Her reaction afterward: “Thank god for edit history,” and “I’m so paranoid.”

Now put that next to what she actually thinks of the tool. Lorenz ran an analysis of Substack’s top newsletters through Pangram’s API in April, describes it as a partnership, and went on Substack’s own podcast to say “I’m a huge Pangram fan”. She thinks it’s accurate. She’s not a skeptic, and she isn’t my witness.

That’s the best case, then. The tool’s most visible champion, a masthead behind her, an accusation made in the open, and a vendor that checked its own work and said out loud that it had erred. She still says she’s paranoid. Almost nobody who eats a bad read gets any of that.

Now assume it works perfectly

Here’s where I part company with almost everyone arguing about this.

The reliability debate is a trap and both sides are stuck in it. One camp says the detector is broken. The other camp points at the benchmarks. I’ve just spent a thousand words telling you the benchmarks are good, and I meant every word of it.

So let’s hand Substack the strongest possible version of its own product. Assume Pangram is right every time. Zero false positives, forever, on everything anyone ever feeds it. Assume deBoer’s brittleness gets patched next quarter.

The feature is still a mistake.

Not because the number is wrong. Because of what happens to a correct number the moment it leaves the lab.

You want to know whether something is worth your time. That’s question two. The scan answers question one, accurately, privately, instantly, and in a format that looks exactly like an answer to question two.

shut this door during launch week in nine words: “Even a perfect score would answer the wrong question.” That’s the pivot, and it’s her sentence, not mine. What I want to add is what the perfect score does instead of answering, which is that it doesn’t sit there neutrally. In her words, “the writing has been handed a grade.”

Every design choice in that product points you at the wrong question and then hands you a confident figure about it.

A tool that is right about the wrong question is not safer than a tool that is wrong. It’s more persuasive.

Matteo Wong, from The Atlantic, who reported the most thorough account of this tool anyone has, wrote something that sounds like my line and isn’t. His version: a mostly reliable detector might be more dangerous than a broken one, because people trust away the small amount of error still in it. That’s a claim about error. Mine is a claim about the question. Fix every error and his worry disappears. Fix every error and mine gets worse, because now the scan is perfectly accurate about something you weren’t asking.

Wong also thinks the deeper problem isn’t the detectors, it’s what they’re trying to detect. The phenomenon really is hard to pin down. But turning something that hard into a percentage and setting it beside the writing was a decision somebody made, and it could have gone the other way.

They wrote the disclaimer themselves

Go back and read Substack’s announcement.

Not everything made with AI is slop, and not all slop is made with AI. Pangram can only detect whether AI was used to make the text, not whether great human care went into creating it, nor whether AI tools were used as a source.

That isn’t me criticizing the feature. That’s the company accurately describing, in public, on launch day, everything its own product cannot tell you.

And then shipping it anyway.

It goes further than the launch post. , Substack’s CEO, gave two interviews the day after, and in both of them he defined the thing this feature exists to fight the same way. Slop isn’t something made with AI. It’s content nobody believes in.

I agree with him completely. That’s the argument of this entire essay, and it’s the one thing his product cannot measure.

In the second of those conversations, with , he went further and described the exact writer I’ve been describing. Somebody who runs seventeen drafts, argues with the model, pushes back on it, and still ends up with a final text that Pangram reads as generated. The hard part, he said, is that this has become difficult to tell apart from the person who typed “write me 10,000 plausible sounding AI takes.” The thing that would actually separate the two he called a sci-fi technology that doesn’t exist yet.

He’s right that it doesn’t exist. He shipped the number anyway, and the number is now doing the job he said it can’t do.

I keep coming back to that. They documented exactly why the number can’t answer the question readers will inevitably use it to answer and put it in the app in the same breath. Every objection you might raise, they’d already written down and published.

Knowing better and shipping anyway isn’t a bug in the rollout. It’s the whole rollout.

Which is where I have to stop calling it a mistake. “Mistake” is the charitable reading, and the disclaimer is the thing that rules it out. A company that publishes the limits and ships anyway hasn’t misunderstood its product. It has priced the cost and decided who pays.

So, run the ledger. Substack gets the LinkedIn contrast it named itself, writers who stay because this place still feels different from the rest of the internet, and a launch week. Pangram gets distribution at platform scale, plus a public “never the ending arbiter” that moves the risk downstream. What’s left over is a reader holding a number nobody will stand behind, and a writer holding the outcome.

Nobody in that chain has to be wrong for the writer to lose. That’s what it means to convert a technical fact into a verdict with nobody accountable for the result. It isn’t a complaint about the design. It’s a description of it.

The part where “it’s private” makes it worse

The standard defense is that nobody gets publicly labeled. You request the scan, you see the result, nothing appears on anyone’s post. No badge. No mob.

I think that’s exactly backwards.

A public accusation is at least a thing you can answer. Somebody says your essay was AI-written, and you can show your drafts, walk through your process, take the argument on in front of the same people who saw the charge. It’s unpleasant. It’s survivable.

A private verdict isn’t survivable, because there’s nothing to survive. Someone runs a scan, gets a number, decides what it means, and closes the tab.

You never learn it happened. You can’t correct it, contextualize it, or point to the interview transcript and the source list. You just get slightly less audience, permanently, for reasons nobody will ever tell you.

Substack didn’t build a scarlet letter. It built one you can’t see and can’t take off.

People are calling it a witch hunt. I went looking for daylight between that frame and mine, expecting to find some, and there isn’t any.

has the sharpest version, and his Salem isn’t the town square. It’s the private examination. “The reader becomes an examiner. The writer becomes a suspect who must, at any moment, prove a human wrote each sentence.” He says plainly that nobody hangs over a Pangram score and that he isn’t pretending they do.

That’s the same machine I’m describing. The accusation never gets made out loud, so it never has to be defended, so it never has to be right, and quiet is what makes all three possible.

And the cost doesn’t land evenly. deBoer engineered a false positive in about fifteen minutes. The peer-reviewed work on detectors as a category, which I’ll come back to, found the sharpest disparities fall on English-language learners. So the people most likely to eat a bad read are the ones with the least standing to object, or as puts it, the least standing to argue about it. And they’ll never know there was anything to argue with.

By its own definition

Let me put this plainly, using the standard I’ve been arguing for the entire essay.

Substack had a real problem. Readers can’t easily tell whether the thing in front of them was actually thought about. Solving that takes taste, context, and somebody willing to make a call and be wrong in public.

They handed it to a classifier.

Slop is not the presence of AI; it is what happens when the human leaves the important decisions to the average.

I don’t think Substack shipped a detector. I think it shipped a way for readers to stop making a judgment they were already fully capable of making, and for nobody to be responsible for the outcome.

That’s the exact shape of the thing I’ve been arguing against this whole time. It just usually shows up as a blog post instead of a product.

And we’ve done this before, never once because the number was better. A loan used to mean a person who had to look at you, decide, and answer for it when the decision went bad. The credit score didn’t only make lending faster. It made the decision unattributable. Nobody turns you down anymore. You just don’t qualify.

That’s the trade every time. We take a worse answer to a narrower question in exchange for not being the one who gave it.

That definition isn’t mine alone. published it independently on July 22, the day after the launch: “Slop begins where human care ends, regardless of whether AI was involved.” She named the category, and she modeled the manners for it too, going looking for her own coinage “fully expecting someone to have used it before me. They had.”

What this essay adds is why the fight over it stays stuck, which is three questions collapsed into one and five camps answering different ones, plus the evidence to settle the parts that can be settled. It also adds one test she doesn’t have, which is whether the work cost the writer a claim they wanted. You’re about to watch that happen to me.

Pangram’s own CEO told The Atlantic the tool should “never be the ending arbiter”, only a starting point for a more thorough investigation. When the vendor is the one telling you to hold the number loosely, that’s worth noticing. It’s also a place to stand later, when somebody holds it tightly and a writer gets hurt.

The trouble starts after the number. “AI-assisted” becomes “AI-written.” “AI-written” becomes “lazy.” The explanation stays home while the accusation travels.

That isn’t hypothetical. In April, James Taranto of the Wall Street Journal editorial page published a piece calling the AI detector “a defamation machine” after Pangram flagged three of his writers as AI-generated. When he checked, one of them didn’t use AI at all and the others had used it to revise rather than to write. None of that distinction survived the label. The label was the part that traveled.

The claim I wanted to be true

This next part cost me something.

While I was working on it, a story was going around: Pangram was flagging old human writing as AI. If that were true, this piece would have written itself. I could have run the easy argument, the one where the detector is broken, the platform is credulous, and everybody worried about slop is chasing a number that doesn’t mean anything.

I wanted it to be true. So I went looking for the evidence.

What I found was a test by a writer publishing as VeryFinePrint, who did something considerably more careful than anyone arguing about it online. They went hunting for old books that had never been digitized, nothing on Amazon, nothing on archive.org, which meant the text could not have been in anyone’s training data. Some of them came off eBay. They scanned forty-five, then ran the whole corpus through Pangram. Just under 2.9 million words, a little over 8,000 segments.

Out of those 8,000 segments, Pangram flagged four.

Then the tester went back and checked them against the actual books. The AI-looking text wasn’t in the books. They’d used Mistral’s OCR to turn the scanned pages into text, and it had hallucinated entire tables onto pages that were blank. In their words, all of the “AI” grades that Pangram gave came from Mistral hallucinations.

Read that again. The machine hired to read the pages wrote some of the sentences it was supposed to be reading, and the detector caught it.

Pangram didn’t fail that test. It passed one nobody was administering. It found machine-written text in a corpus its own tester believed was entirely human, and the tester’s conclusion was that the results gave them confidence the tool is not memorizing its training set.

The false positives were real. They just weren’t Pangram’s.

Now the honest bounding, because this essay doesn’t get to skip it: that’s one test, by one person, publishing under a pen name, and it doesn’t establish Pangram’s error rate in general. I’m not going to pretend it does. It was only enough to kill my argument, which is a much lower bar and the one that mattered here.

And it had been sitting there in public the whole time. It went up six days before Substack shipped the scan, while everyone argued past it.

But sit with the shape of it, because the shape is the whole essay.

The fast take was wrong. The careful answer required tracking down forty-five books that had never been digitized, scanning them, and running three million words through a classifier, which is not a thing anyone does by hand on a Tuesday afternoon. Getting to the accurate conclusion took dramatically more machine assistance than getting to the sloppy one.

The people who used less automation were the ones who got it wrong.

The first draft may be smoother. The second may use more AI.

The difference is not the amount of machine-written prose. It is where the important decisions came from.

You can remove every em dash from an AI-generated article and still publish slop.

Slop is not the presence of AI; it is what happens when the human leaves the important decisions to the average.

Subscribe now

So don’t take my word for any of it

You have no way to verify that story about myself. I could have made the whole thing up. That’s the obvious hole in everything I’ve argued so far: if accountable authorship is the real test, and accountable authorship is invisible in the finished text, then my standard collapses into “trust me,” and every slop merchant on the internet can recite the same receipt.

So don’t take the receipt. Take the evidence of it.

You can’t audit my process, but you can audit the output. A piece where somebody actually did the thinking leaves marks, and you already know how to find them. Three of them, specifically.

Something it cost the writer. Not the confession, the damage. Anyone can type “I wanted this to be true.” What you check is whether the piece still leans on the argument it says it gave up. A real concession changes what comes after it. A decorative one doesn’t.

An easier argument, declined. There is almost always a cheaper version of any piece sitting right next to it. You don’t need to know what it was. Count whether the writer concedes anything that costs them.

Specificity where a machine would generalize. Generated prose gets vaguer exactly where the work would have gotten harder. The detail is the tell, and it runs the opposite direction from the one everyone is watching.

That’s a ninety-second read, not a scan, and you were always going to have to do it yourself.

Which is the part nobody wants to hear. There is no number coming to rescue you from reading.

The effort objection gets time backward

I get why Orosz’s line landed. The internet is filling up with people who can produce in seconds what they used to have to think about for hours, and you’re paying for it in filtering, fact-checking, and time you don’t get back.

But “AI was involved” doesn’t mean “nobody put in the time.”

Sometimes it means the time moved.

I’ve got a full-time job, a family, and other work I care about. Writing isn’t the only claim on my week. The system that lets me publish anyway didn’t show up because I typed a clever prompt. It took a long time: learning how the models actually behave, building the orchestration, curating which sources they’re allowed to trust, turning my own published writing into voice rules, designing the research checks, and fixing the process every time it broke.

The judgment in this piece wasn’t produced this week. Most of it was installed long before the draft existed.

That doesn’t entitle me to your attention. Nobody owes a writer credit for effort.

It does make the inference wrong. We don’t usually judge a tool by whether it preserved every hour of the process it replaced. Nobody refuses to wear a shirt because a machine sewed it. We judge what the tool makes possible, what it quietly destroys, and whether the person holding it still answers for what comes out.

And the inference fails much harder for people who aren’t me.

If you write in your second language, AI isn’t a shortcut. It’s the distance between the sentence you meant and the sentence you could manage alone. If something sits between your thinking and your typing, whether that’s dyslexia or ADHD or a tremor, same thing. There are writers for whom this technology is the reason the work exists at all, and “they didn’t take time to write it” is a strange thing to say to someone who took longer than you did.

That’s me arguing on their behalf, which is the weaker version of this. Here’s one of them arguing for herself. , a non-native English writer with a neurodivergent brain, says AI “helps me reach thoughts I can feel before I can explain them.” Read that again and then try to hear it as a shortcut.

The peer-reviewed work on detectors as a category isn’t reassuring here. An ACL 2026 study tested 16 detector systems on essays by American students in grades six through twelve and found real disparities affecting English-language learners. Pangram wasn’t one of the 16. Neither was any other commercial tool, so I can’t pin that on Pangram and I’m not going to, whatever it would do for my argument. But if we’re building reader reflexes around detector scores, it’s worth knowing which writers the category has historically been worst to.

I’m a guy with a full-time job and a content system. If the effort inference is merely wrong about me, it’s something worse than wrong about them.

That handles the effort claim. It doesn’t touch the harder objection, which is that you might want a human on the other end regardless of any of this.

There’s also a fairer version of the complaint, and it isn’t about me at all. Scaled production means more work arriving in front of you, and more of it arriving fast. Even if every individual piece is good, the volume itself costs you something.

That’s a real problem and I don’t have a clean answer for it. (Anybody who tells you they do is selling something.)

But here’s the version that would actually land on me. The fair criticism isn’t that I saved time. It’s that I might have saved time by making you carry the thinking I refused to do.

That’s exactly the failure I want an anti-slop standard to catch.

Human presence isn’t the same as human purity

The harder objection doesn’t depend on my process at all.

You don’t read only for information. You read for contact with another mind. You might want prose written entirely by a person because the human act is part of what you came for.

You’re not buying the information. You’re buying the encounter.

I can’t argue you out of that, and I shouldn’t try. If that’s what you want, my disclosure should help you choose, and you may decide this piece isn’t for you. That’s a relational choice. It isn’t a finding that the work is slop.

But human presence survives AI assistance more often than the binary allows. It shows up in what a writer notices, which tension they refuse to resolve too neatly, what evidence changed their mind, and whether they’ll stand behind the result.

It shows up at three moments:

• Before the draft, someone decides what the piece should say, what evidence it needs, and which tension can’t be flattened.

• During the draft, someone supplies context, direction, voice boundaries, and a standard for what good looks like.

• After the draft, someone checks the claims, rejects weak reasoning, cuts what doesn’t belong, and accepts responsibility for the result.

None of that requires a purity ritual around typing every sentence. All of it requires ownership.

My rule is simple, and it’s a floor rather than a boast. I read this one several times, but the bar I’d hold anyone to is a single honest read of every line before their name goes on it. (Yes, that is a low bar. It is also cleared less often than you’d hope.)

Reading it doesn’t make the work good. It’s the minimum act that makes the work yours.

If you’re not willing to give your own writing that much attention, why should anyone else?

The slop meter keeps moving

Orosz had a third line in that post. He said it’s still pretty easy to notice AI-written text, and that it frustrates him to no end.

He’s right. Today.

Still is the word carrying that sentence, and it has a short shelf life.

Surface tells do matter right now. If a phrase or a punctuation mark makes a reader stop thinking about your idea and start litigating whether a machine wrote it, your style is competing with your point. That’s why you won’t find an em-dash in this essay.

But here’s the part almost nobody says out loud about the em-dash.

The model uses them because we did. It read an enormous amount of human writing, and human writers loved that punctuation mark. The tell isn’t a machine fingerprint. It’s our own habit, averaged and handed back to us at volume.

We taught it to write like that. It learned. Now we’re offended.

Which is exactly why the rules expire. Models change, writers adapt, and the meter moves. People will start adding awkward phrasing, stray typos, and deliberate mistakes to prove a person was involved. If enough of us adopt the same approved imperfections, we’ll have built another template.

That isn’t a prediction, incidentally. deBoer spent about fifteen minutes deliberately writing a hundred words of human prose engineered to come back as AI, it worked, and he says he could do it again. The arms race is already running in both directions, and neither direction has anything to do with whether the writing is any good.

The other direction has a number on it, and the number is the vendor’s own. Pangram’s current model card publishes per-domain false-positive rates between zero and 0.49%, and its CEO has claimed something closer to one in 10,000 on academic essays. Asked about the reverse case, he pointed The Atlantic to a test where the false-negative rate came out closer to one in 70. Wong then ran a commercial humanizer over machine-written text himself, and Pangram called the output human-written every time he tried it.

So by its maker’s own account, the tool is far better at not accusing you falsely than at catching what it’s hunting for. Which is the opposite of the way a reader will instinctively read the score, and it’s the direction nobody is arguing about.

A style guide for where to place your convincing typo would be a strange way to recover a voice.

So stop grading punctuation. Anything still standing after the next model release has to be about how you think, not how you type. The average can always change clothes.

They already built the right feature

The part that actually bothers me is that the good version is already in the product.

Substack lets a writer attach a statement explaining how they make their work. That’s the thing. That’s authorship, disclosed by the person accountable for it, in their own words, where you can weigh it against what you just read and decide what you think of them.

It answers question three, which is the one that matters. It puts the burden on the writer, where it belongs. And it can’t be screenshotted into an accusation, because it isn’t a score. It’s a person talking.

Yes, a writer can lie in that statement. That’s the objection, and it’s real. But a lie you can read and weigh is a better problem than a number you can’t argue with. The statement can be wrong. The score can be right and still send you to the wrong conclusion, which is worse, because you’ll believe it.

There’s a harder objection than that one, and made it. “A box for explaining yourself only exists once somebody has decided you have something to explain.”

She’s right about the sequence. The suspicion got installed first and the box arrived behind it, so calling the box a favor is a little rich.

But there’s a difference inside that burden, and it’s the difference this whole essay turns on. A burden you author, in your own words, on your own terms, is not the same instrument as a burden a classifier authors about you. One of them you can answer. The other one gets answered for you.

Doyal’s own response is to leave, more or less. Build things you own instead of renting your standing from a platform that keeps changing the terms. That’s a legitimate answer and I won’t pretend it isn’t. She also caught something nobody else did: the same company putting a number on the AI in your prose will happily generate a synthetic voice to read that prose aloud. Which suggests the problem was never really AI.

Then they buried it under a button that answers question one, and made the button the headline.

Ship the statement. Skip the score. The score quietly teaches a few hundred thousand people that accounting for a writer is something a model can do.

It’s not hard to see why they went the other way. A statement generates no scan volume, no launch coverage, and no contrast with LinkedIn.

If Substack wants to protect what makes this place different from LinkedIn, that difference was never going to be enforced by a classifier. It was always going to be enforced by readers with taste, which is the thing they just built a machine to replace.

Scan it

I meant the invitation at the top.

Run this piece through the scan. Whatever comes back is true, and it answers exactly one of the three questions: what role the machine played in producing these sentences. That’s real information and you’re entitled to it.

It won’t tell you whether the argument holds. It won’t tell you whether the VeryFinePrint test checks out, and you should go read it yourself, because I did and it changed this essay. It won’t tell you that a claim I wanted got killed by its own evidence, or that I threw out an entire draft, or that I read this one several times before my name went on it.

It can’t tell you those things because they didn’t happen in the prose. They happened before it, and around it, and they’re the only part that was ever mine.

So take the number. It’s free, it’s accurate, and it’s nearly useless for the decision you’re actually making.

Then do the thing the number can’t do for you, which is the same thing you were doing before anyone built it. Read the work. Decide if it’s any good. Notice who signed it.

I don’t need Substack to protect me from slop, and I don’t think you do either. You’ve been able to close a bad tab your whole life.

I still have a brain. I’m assuming you do too.

The rules beneath the work

Your team probably has a rule for what counts as urgent. It almost certainly isn’t written anywhere. The right person knows when to break the normal process, who needs a heads-up, and which number can’t be trusted without context.

Then an AI agent enters the workflow. It doesn’t have the shared history, the hallway correction, or the colleague who leans over and says, “Not that way.” It’s working from the instruction and whatever your systems made visible.

This week kept showing what happens next: unwritten rules don’t disappear. They harden into defaults.

Agents don’t remove ambiguity. They turn it into behavior.

The format is simple.

Three signals exposed the same hidden layer from different directions. A workplace metric allegedly lost the context around protected leave. A study of multi-agent systems found failures gathering around design, coordination, and verification. A file-organizing agent didn’t follow the approval behavior its user expected.

The common denominator wasn’t the model but the missing protocol.

Subscribe now

The week

The past two weeks were a blur of launches: new models, new desktop apps, an announcement cadence that barely paused for breath. This issue isn’t here to rank any of it, because the releases weren’t the story. What agents exposed about how work actually runs was.

Most organizations didn’t begin the AI era with a clean operating manual. They started with years of habits, exceptions, workarounds, and people who know which rule matters when two rules collide.

We’re surprisingly good at carrying that invisible layer. A manager sees a low activity score and remembers the person spent the quarter mentoring half the team. A teammate receives a thin brief and knows which three questions to ask. An experienced operator sees “approval required” and still makes a backup before trusting it.

Agents don’t share that history. They read what’s available, follow what’s explicit, and reveal everything the team was quietly supplying for free.

That sounds like a limitation of AI. It’s also a diagnostic for the organization. Every awkward failure points back to a rule, judgment call, or exception path that existed in someone’s head but nowhere the system could use it.

This week’s real story wasn’t that agents need more intelligence. It was that work needs more legibility.

3 Signals

Signal 1: The proxy became the decision

What happened

On July 13, a group of 26 Meta employees filed a federal lawsuit alleging that the company used a set of internal AI systems, including activity-monitoring data and algorithm-assisted performance rankings, when selecting employees for layoffs.

The employees allege those measures disadvantaged people on medical, parental, or family leave because they couldn’t accumulate the same visible activity while away from work. Meta disputes the claim. The company told the Associated Press that workforce decisions were made by people, not AI, and that the allegations lack merit.

That distinction matters. This is an allegation being tested in court, not a settled finding. It’s still worth examining because the underlying design problem exists far beyond one company.

What this reveals

A proxy is a visible measure used to stand in for something harder to see. Activity can stand in for contribution. AI usage can stand in for adoption. That’s how a tidy score starts standing in for judgment.

The trouble starts when the proxy loses the context that made it useful. A person on protected leave can have less visible activity without creating less value. A manager who knows the person and the work can catch that mismatch. A system that only sees the score can’t, unless someone designed an exception path into it.

That exception path is an invisible protocol. It’s the unwritten rule that says, “This number normally helps, except when this condition is true.” Teams use those sentences all day. They don’t treat them as infrastructure.

AI makes that omission consequential. Once a score can feed a recommendation, ranking, or action, you can’t leave the context around the score outside the system.

Sources: Doe v. Meta Platforms complaint, July 13, 2026 · Associated Press, July 14, 2026

Signal 2: The handoff ate the context

What happened

A March 2025 research paper resurfaced in builder conversations this week because its findings don’t look academic anymore. They look like a normal Tuesday.

The MAST study examined more than 1,600 annotated traces from seven multi-agent systems. The researchers identified 14 failure modes across three broad categories: specification and system design, misalignment between agents, and verification or termination.

In plain language, the systems often failed because roles weren’t clear, information didn’t travel cleanly, steps repeated or stopped at the wrong time, and nobody reliably checked the final work. The researchers found that both the underlying model and the system’s design shaped the failure pattern. Better instructions helped in some cases, but no single prompt repaired the whole problem.

What this reveals

The most fragile moment in a workflow is often the handoff. It’s the instant when one person believes they transferred the task and the next person discovers that the context stayed behind.

Human teams compensate. Someone asks a follow-up question. Someone recognizes the customer name. Someone knows that “final” means legally approved, not visually polished. Those corrections happen so naturally that we don’t count them as part of the process.

An agent makes the seam easier to see. If the input doesn’t specify the role, the source of truth, the finish line, and the stop condition, the system has to guess. A smarter model can guess better. It’s still guessing.

Stop treating the handoff as the empty space between two boxes on a workflow diagram. The handoff is work. It carries meaning, authority, state, and the conditions for saying “done.”

Source: Cemri et al., “Why Do Multi-Agent LLM Systems Fail?”, March 17, 2025

Signal 3: The setting was not the contract

What happened

ZDNET writer David Gewirtz gave ChatGPT Work and Claude Cowork the same kind of task: organize a copied folder containing 447 PDF files.

The copy was the smart part. It made the test reversible before either agent touched anything.

In Gewirtz’s test, ChatGPT Work moved and renamed hundreds of files without presenting the approval prompts he expected, even though the approval setting was enabled. Claude Cowork had paused for approval during his earlier test. Gewirtz called the missing permission prompts the biggest deal breaker.

This was one controlled test by one reviewer. It wasn’t proof that every session behaves the same way. It still exposes the gap between configuring a permission and verifying how that permission behaves under real work.

What this reveals

A setting is a stated promise. A contract is a promise you’ve tested, can observe, and know how to recover from when it breaks.

That difference grows with the agent’s authority. If it can draft a paragraph, a missed approval is annoying. If it can move files, send messages, spend money, or change production data, you can’t treat the same miss as a minor glitch.

The reviewer did three things every team should notice. He narrowed the task, used a copy, and watched the behavior. Those aren’t technical tricks but the basic structure of safe delegation: limited scope, an undo path, and evidence of what happened.

Reliability is designed before the agent starts. You decide what it can touch, what requires a pause, what success looks like, and how to reverse the result. Then you test whether the system follows that contract. The toggle alone can’t do that work for you.

Source: David Gewirtz for ZDNET, July 15, 2026

2 Experiments

The first three Field Notes issues asked you to inventory what survives, draw delegation boundaries, inspect AI output, and rebuild a missing rep. They’re useful moves, but they all began with you looking at work you already understood.

This week calls for something different: you’ll use AI to interview the workflow, then put the resulting protocol through a live teach-back.

Experiment 1: Interview the workflow

Try this

Pick one recurring handoff that depends on someone “knowing how we do things.” It could be preparing a client deck, approving an expense, publishing a report, or turning meeting notes into next steps. You’ll know the right one because the written process isn’t enough to do it well.

Open a fresh AI chat and paste this prompt:

You are a workflow archaeologist. Do not perform the task. Help me uncover the unwritten protocol behind it.

I will describe one recurring handoff from my work. Ask me one question at a time, with no more than six questions total. Probe only for:

1. What triggers the work
2. Which inputs and sources are trusted
3. What the recipient may decide without asking
4. Which exception would change the normal process
5. What observable evidence proves the work is done
6. How to pause, escalate, or reverse the work

After the questions, produce a five-part operating contract:

- Purpose
- Inputs and source of truth
- Decision rights and boundaries
- Definition of done and evidence
- Exceptions, escalation, and undo path

Do not invent missing answers. Mark each unknown as UNKNOWN.

Finish by naming the single hidden assumption most likely to break if a new teammate or AI agent takes over.

Start by asking me which recurring handoff we are mapping.

Keep the result to one page. If the AI returns a beautiful policy manual, you and the model both missed the assignment.

Why it matters

Most workflow documents capture the visible steps. They miss the judgment around the steps: which source wins, when the normal rule bends, who can make the call, and what evidence earns trust.

The interview format pulls those rules out through questions instead of asking you to remember everything at once. Marking unknowns also prevents the model from filling a gap with something that sounds sensible but isn’t true.

The goal is one usable contract for one real handoff, not perfect documentation.

Experiment 2: Run the teach-back gate

Try this

Take the one-page contract from Experiment 1 and give it to a clean AI chat or a colleague who doesn’t carry the usual background. Give them one real, reversible task.

Before they begin, ask for this teach-back:

Before you take any action, restate your understanding in four lines:

1. The outcome you believe I want
2. What you may change without asking
3. The condition that requires you to stop and escalate
4. How you will prove the result and undo it if needed

Do not begin until I confirm your restatement.

Compare the answer with what you intended. If one line is wrong, don’t correct them verbally and move on. Fix the contract. Then let them complete only the first reversible step.

Why it matters

A handoff can sound clear to the person who wrote it because their own context fills every blank. Teach-back removes that advantage. It shows you what the instruction says to someone who doesn’t have anything else.

That makes the mismatch useful. You catch the missing rule before it becomes a moved file, a bad decision, or a confused teammate quietly repairing your brief for the fifth time.

The experiment also creates a better habit than “ask for approval.” Approval at the end is late. Teach-back checks whether you’ve built shared understanding before action begins.

1 Reflection

AI may not be creating your coordination problem. It may be the first coworker that can’t survive on vibes.

So here is the thing worth carrying into your week:

If a new person or agent followed only the rules your team has written down, what would they get wrong first?

That answer isn’t an embarrassment. It’s your next protocol waiting to be made visible.

That’s Field Notes #004. If it helped you see the week a little clearer, forward it to one person trying to do the same. I’ll be back when the field hands me the next one worth your time.

Share

A month ago it was a prediction. This week it was a filing.

For two years the AI jobs conversation ran on predictions. Bold ones. A month ago the people who made the boldest started taking them back: Sam Altman said he’d been “pretty wrong” about the economic fallout, and Dario Amodei softened his old line that AI would erase half of entry-level white-collar work. Then, this week, while the forecasts were busy retreating, Oracle quietly did something none of the predictions did. It wrote AI into a federal filing as a reason 21,000 of its people no longer have jobs.

So the story changed registers. It went from what might happen to what’s on the record. And once you read the week as a record instead of a forecast, a sharper pattern shows up underneath the layoff headline.

The format is simple.

It’s not a news roundup, and it’s not a list of tools. Plenty of people will tell you what happened. The job here is to tell you what it means. This week it meant one thing, said three ways: the skill everyone now needs is getting harder to build.

Subscribe now

The Week

Last issue the lesson was that the advantage moved up the stack, to the people who could govern an agent, give it clean context, and judge what it produced. Call that one skill by its real name: judgment. The ability to tell a good answer from a plausible one.

This week the field handed us the receipts on judgment, and they don’t agree with each other. One set says judgment is now the whole game. A company put it in an SEC filing. Hiring data shows employers demanding it at the entry level, where it never used to live. The productivity numbers show it’s the real bottleneck behind every AI tool.

The other set is quieter and more unsettling. The same forces pricing judgment up are removing the ordinary work people used to climb through to build it. The first drafts, the grunt research, the simple tickets, the junior reps. AI is eating exactly the rungs you used to step on.

3 Signals

Signal 1: The layoff got a paper trail

What happened

This week Oracle disclosed that its headcount fell by about 21,000 people over the past year, from roughly 162,000 to 141,000. That alone is a big number. What made it a milestone is where it showed up and how it was worded. In its annual SEC filing, the legal document a public company swears is accurate, Oracle wrote that “the adoption and deployment of AI technologies across our operations have resulted, and may continue to result, in reductions to our workforce.” (An SEC filing is the audited report companies are legally required to give investors. It’s the opposite of a press release. You don’t put a line in it lightly.)

That’s the first time a major company has named AI as a cause of layoffs in a federal filing, rather than in an interview or a memo. And the timing is the tell. Just a month earlier, the loudest voices were walking their forecasts back. Altman admitted he’d been “pretty wrong” on the social and economic side. Amodei, who’d said AI could wipe out half of entry-level white-collar jobs and push unemployment toward 20%, shifted to saying automation might expand the work people do. The Yale Budget Lab, tracking the labor market since ChatGPT shipped, still finds no broad unemployment spike for AI-exposed workers.

What this reveals

Watch the misdirection. Predictions are cheap and reversible. You can say AI will take half the jobs, get headlines, and quietly take it back when the IPO paperwork is due. A filing is different. A filing is a paper trail.

But read Oracle’s own numbers and the headline gets more complicated. The same filing shows capital spending jumped 162% to $55.7 billion, almost all of it pouring into AI data centers. The cuts aren’t only “the robot did your job.” They’re also a company moving cash from payroll to compute, and using AI as the cover story that makes it sound inevitable instead of chosen. That matters, because “AI did it” is becoming the most convenient sentence in corporate America, and it ends arguments it shouldn’t end.

So the 21,000 is real, and it’s worth your attention. But it’s also the loud number, the one designed to be looked at. The number that actually decides your future this week is quieter, and it’s hiding one signal down.

Sources: Oracle sheds 21,000 roles amid AI layoffs (CNBC, June 23, 2026) · Altman and Amodei walk back the jobs apocalypse (Fortune, May 26, 2026)

Signal 2: The bottom rung went missing

What happened

Here’s the quiet number. PwC’s 2026 Global AI Jobs Barometer read more than a billion job ads across 27 countries and found something strange happening at the bottom of the ladder. Entry-level roles in the jobs most exposed to AI are now seven times more likely to demand skills that used to show up much later in a career: judgment, strategic decisions, stakeholder management, the senior stuff. In the most exposed work, 52% of the new skills appearing in entry-level postings were ones we used to associate with experienced people. Fortune gave it a name: seniorization.

The shape gets clearer when you look at the openings themselves. Entry-level roles that demand these senior skills have grown 35% since 2019. Ordinary entry-level roles, the kind a new grad could actually start in, shrank 10%. So the door didn’t close. It just moved up a floor, and took the staircase with it.

What this reveals

This is the signal under the signal. AI is very good at exactly the work the bottom rung was made of: the first draft, the background research, the simple ticket, the rote analysis. That work was never valuable for its output. It was valuable because it was how a person built judgment, one unglamorous rep at a time. You learned to spot the wrong answer by producing a few hundred of them yourself.

Now employers want the judgment without funding the apprenticeship that produced it. They want people who can supervise AI’s work on day one, in roles that used to exist precisely so you could learn by doing the work AI now does. Last issue’s reflection asked whether you could tell when an agent got your job wrong. This is the harder version of that question, pointed at the whole pipeline: you cannot supervise what you were never allowed to learn. The scarce skill and the broken path to it are the same story.

Sources: PwC 2026 Global AI Jobs Barometer (PwC, June 2026) · Entry-level work didn’t disappear, it “seniorized” (Fortune, June 18, 2026)

Signal 3: The skill got renamed, and the productivity story got honest

What happened

If judgment is the scarce thing, the week also told us what it’s being renamed to. For two years the hot skill was “prompt engineering,” writing the clever instruction. That’s now being demoted to one small input. The skill people are actually hiring for is “context engineering”: designing everything the model sees before it answers, the data, the memory, the examples, the brief. Andrej Karpathy named it in 2025, and it’s spread from engineering teams into job postings and interviews. The shift is subtle but it’s the whole game. The question moved from “what do I tell the model to say” to “what does the model need to know, and is what it gave me back actually right.”

And the honest productivity picture finally caught up. About 93% of developers now use AI coding tools, yet measured productivity gains are stuck around 10%, and AI already writes roughly 27% of production code. A landmark trial from METR put the paradox in sharp relief: experienced developers were 19% slower on real tasks with AI, while feeling about 20% faster. Only around 29% say they trust what the AI hands them.

What this reveals

Sit with the gap between feeling faster and being slower, because it explains the whole week. The work didn’t disappear when AI got good. It moved. It moved from producing the thing to checking the thing, from writing to verifying, from “can I make this” to “can I tell if this is wrong.” Call it the verification tax. Everyone’s paying it, and most people can’t feel themselves paying it, which is why raw AI usage doesn’t turn into output.

This is why “context engineering” is more than a new job title. Feeding the model good context and judging what comes back are the same muscle: knowing your domain well enough to set it up right and catch it when it drifts. That muscle is judgment again, wearing a technical hat. The model got cheap and confident. The scarce, expensive, slow-to-build thing is the human who can tell when its confidence is misplaced. The tools keep getting smarter. The bottleneck is, and stays, the quality of the person holding them.

Sources: Context engineering, the skill replacing prompt engineering (Karpathy, 2025) · Measuring the impact of AI on experienced developer productivity (METR, July 2025) · 93% of developers use AI, productivity still ~10% (ShiftMag, 2026)

2 Experiments

Experiment 1: Run a one-week verification log

Try this

For the next week, every time you use something AI made you, a draft, an analysis, a chunk of code, an email, spend thirty seconds before you send it writing down the one change you had to make and why. Just a running list in a note. What did you catch. What did you fix. What did you wave through unread. By Friday you’ll have something most people never see: a map of your own judgment.

Why it matters

The verification tax is invisible until you measure it, the same way those developers couldn’t feel themselves slowing down. The log makes it visible. The places you reliably catch errors are your moat, the senior judgment this whole wave is repricing. The places you rubber-stamp without checking are your exposure. You can’t get better at telling true from plausible if you never notice yourself doing it. Fifteen minutes of noticing this week buys you a clearer picture of where your real value sits than any think piece on the future of work.

Experiment 2: Rebuild one rung

Try this

Pick one task AI now does for you that you couldn’t fully evaluate if you had to, the thing you accept because checking it feels too hard. Do it by hand, once, this week. Not to be slow on purpose, but to feel where your own judgment has gone thin. Then, if you manage people, do the inverse: hand a junior one “AI could do this” task on purpose, as reps, and protect the time it takes.

Why it matters

This is how you fight the missing rung on both ends. Doing the task yourself rebuilds the apprenticeship AI quietly removed from your own week, the rep that teaches you to spot a wrong answer before you can explain how you knew. And handing a junior real reps is the thing almost no company is doing right now, which means doing it is a quiet advantage. The skills that compound are the ones you keep practicing. AI just made it tempting to stop practicing the most important one.

1 Reflection

The loud story this week was a number, 21,000, designed to be stared at. The quiet story is the one worth carrying. AI is making good judgment the most valuable thing you own, and at the very same time it’s absorbing the ordinary work that used to build it. The price went up and the path got narrower in the same week.

So here’s the thing worth carrying into your week:

The scarce skill is no longer doing the work. It’s knowing when the work is wrong. So who is still building that in you, and what happens if the answer is only you?

That answer is your edge, and unlike the model, nobody can switch it off. Go build more of it.

That’s Field Notes #003. If it helped you see the week a little clearer, forward it to one person trying to do the same. I’ll be back when the field hands me the next one worth your time.

Share

Last week, a rental. This week, the tenants.

Last issue the lesson was that the model is a rental. You do not own it, you cannot fully see inside it, and someone else can switch it off overnight. This week the field made that concrete in a different way: the agents moved in. They started writing the code, pressing the deploy button, and running real parts of the work, and they brought the bill, the blast radius, and a new dependency on the one thing that does not scale, your judgment.

None of the week’s biggest stories were really about a smarter model. They were about what happens once software can act on its own, and how unready most of the scaffolding around it still is.

The format is simple.

It is not a news roundup, and it is not a list of tools. Plenty of people will tell you what happened. The job here is to tell you what it means.

This week it meant one thing, said three ways: the advantage moved up the stack.

Subscribe now

The week

For two years, “AI agent” mostly meant a demo. A clever thing on a stage that booked a fake dinner reservation. This was the week that stopped being true.

At its developer conference, the hosting company Vercel said that the share of software deployments triggered by an AI agent rather than a person went from under 3% to more than half in about six months. (A deployment is the moment new code goes live to real users. An agent is a program that acts on its own instead of waiting to be asked.) That is their own number, from the company with the clearest view of the traffic, so read it as a direction, not a census. The direction is the point.

Underneath that, the same shape kept showing up. More than half of sales teams now run agents across the actual sales cycle, not as an experiment. Banks are deploying AI faster than they can govern it. The frontier model you would have paid a fortune for last year is now something you can download and run yourself. And the people getting real gains from AI all quietly admit the gains depend on a senior human watching closely.

The advantage did not go to whoever had the best agent. It moved up the stack, to the person who could govern, contextualize, and judge what the agent did. Here is what that looked like.

3 Signals

Signal 1: Agents started acting, so the walls became the job

What happened

The week’s agent stories all pointed the same way: these things have stopped answering questions and started doing the work, while the means to contain them lags behind.

The clearest number landed this week from Vercel, at its developer conference: the share of software deployments triggered by an AI agent rather than a person went from under 3% to more than half in about six months. It is their own figure, from the company with the clearest view of the traffic, so read it as a direction, not a census. And it is not a one-platform fluke. Salesforce’s most recent State of Sales report has 87% of sales teams using AI and 54% already running agents across the sales cycle, doing prospecting, quoting, and round-the-clock outreach. The friction is showing up in the plumbing, too: Git, the system the whole software world uses to track changes, was built on the assumption of one human making one careful change at a time. Point a swarm of agents at it and it buckles, with merge conflicts, duplicated work, and code that compiles but quietly disagrees with itself. As one open-source engineering roundup put it, Git was built for human-to-human collaboration, and that model breaks at machine speed once a thousand engineers each spin up a hundred agents.

The same week, the people who build these systems were busy hardening them like any other risky infrastructure: least-privilege access, audit trails, circuit breakers, an approval gate before an agent can touch anything that matters. They are doing it because the failure mode is now expensive. By one estimate from a production-agent playbook, around nine in ten deployed agents run with broader access than their job actually needs. Deloitte put the gap in one line: agentic AI is scaling faster than the guardrails, with only about a fifth of companies reporting a mature way to govern the agents they have already shipped.

What this reveals

The hard part is no longer “can the agent do the task.” It is “where are the walls.” Who set the spending limit. What it is allowed to touch and what it must never touch. What gets logged. Who signs off when it wants to go further. Who is accountable when it confidently does the wrong thing at scale.

None of that is a model capability. It is a judgment call, made in advance, and written somewhere the agent can actually reach. This is the direct sequel to last week’s point about giving AI a credit card: letting a tireless program act on your behalf with no boundaries is not delegation. It is delegation without a manager. The advantage this week went to the teams who treated their agents like a strange new class of junior employee, the kind that needs a job description, permissions, and a performance review, not a magic intern you set loose and hope for the best.

Sources: Vercel agent deployment data (Vercel Ship 2026) · State of Sales 2026 (Salesforce) · What version control looks like when AI agents write the code (All Things Open) · Agentic AI is scaling faster than guardrails (Deloitte)

Signal 2: The model went on sale, so the foundation became the moat

What happened

While agents ate the headlines, the price of raw capability quietly fell through the floor.

A Chinese lab, Z AI, released GLM-5.2, an open-weights model you can download and run on your own machines, under a permissive license. On the independent Artificial Analysis Intelligence Index, it became the strongest open model ever measured and landed fourth overall, behind only Claude Fable 5, Claude Opus 4.8, and the top setting of GPT-5.5. It runs at the lowest cost per task of any model at its intelligence level, and on the index’s real-world agentic test it lands effectively level with the top setting of GPT-5.5.

Sit with the continuity for a second. Fable 5, the model that got switched off in three days last issue, still sits at the top of that chart. But you cannot use it. Meanwhile a model nearly as capable is now something you own outright, that no government and no vendor can reach in and disable. The frontier did not just get switched off last week. This week it also went on sale.

So enterprises started saying the quiet part out loud: the model was never the bottleneck. The foundation under it was. SAP told its customers that embedded AI only works on a “clean core” of standardized processes and high-quality data, and that this is now a board-level priority, not an IT chore. The same week, Deloitte’s numbers showed why: only about a quarter of companies have moved even 40% of their AI pilots into production. What stalls them is rarely the model. It is messy data, tangled processes, and unclear ownership.

What this reveals

This is the literal proof of last week’s idea: own the layer that travels. When the model becomes cheap and abundant, the value flips to the unglamorous layer you actually control, the clean data and structured context you feed it. The people pulling ahead are not using a smarter model than you. They have done the boring work of building a clean foundation it can stand on.

It even changes how you think about vendors. The engineer Gergely Orosz put the new instinct well this week: run any model behind a router so you can switch providers the moment one tries to force bad terms on you, “like Anthropic with Fable.” The model is interchangeable. The foundation, and the freedom to swap what stands on it, is the part that is yours. Context architecture stopped being hygiene. It became the strategy.

Sources: GLM-5.2 leads the open-weights index (Artificial Analysis) · Embedded AI must be a board-level priority (SAP) · Agentic AI strategy and production rates (Deloitte)

Signal 3: The bottleneck moved to the one thing that does not scale

What happened

Here is the finding that should stop you. As agents do more of the doing, the people getting the biggest gains are not the ones writing the most code. In some cases they are not even faster in the obvious sense.

One 2026 analysis of AI coding agents found that they can actually slow experienced developers by close to 20%, even while it feels faster, because the senior has to read, validate, and quietly fix what the agent produced. The multiplier is real, but it is conditional. As the practitioners keep saying, AI boosts output, but only a senior prevents the scalable mistake. The bottleneck moved from doing the work to judging it.

You can see the same shape in how serious companies are organizing. Deloitte’s playbook for agents now includes a role it calls the “agent supervisor,” a human who steps into the workflow at chosen points to handle the exceptions that need judgment. Not to check every line, but to own the calls a machine should not make alone.

What this reveals

The advantage is not “uses agents.” Almost everyone will use agents soon. The advantage is being able to tell when the agent is confidently wrong. That is the skill the whole wave is quietly repricing, and it is the hardest one to fake.

It is also the one juniors cannot shortcut. You cannot supervise what you cannot yet evaluate. The philosopher Michael Polanyi called this kind of know-how tacit knowledge: we know more than we can say, and the experienced person catches the quiet error precisely because they cannot fully explain how. That instinct used to be a nice-to-have. This week it became the main limit on how far an organization can safely lean on automation. The widening gap is not between companies that have AI and companies that do not. It is between the ones merely using AI and the ones whose judgment is dense enough to actually run on it.

Sources: The AI coding productivity paradox, citing the METR study (Exceeds AI) · The agent supervisor role (Deloitte)

2 Experiments

Experiment 1: Draw the walls before you hand over the task

Try this

Before you let an agent, or even a chained set of prompts, do real work for you, spend ten minutes writing its boundaries down. Four lines is enough. What it is allowed to touch. What it must never do. The spending or scope limit. Who signs off when it wants to cross one of those lines. Keep the note somewhere you would actually reuse it.

Why it matters

This week proved the agent is not the risk. The missing walls are. The good news is that boundaries are portable in a way models are not. The model you wrote them for can be switched off, repriced, or replaced, and your rules for what you will and will not delegate carry straight over to the next one. You write the walls once, and they travel with you.

Experiment 2: The supervision test

Try this

Take one thing AI produced for you this week. A draft, an analysis, a chunk of code, a plan. Now ask a hard question: could a smart junior on my team catch what is wrong with this? If the answer is yes, the task is safe to delegate and lightly check. If the answer is no, that gap is the exact senior judgment the agent wave is repricing, and it is yours to keep building.

Why it matters

It turns a vague worry (”am I falling behind?”) into a concrete map of where your real value sits. The tasks a junior could supervise are the ones to hand off without guilt. The ones only you can catch are your moat, so spend your scarce attention there. While you are at it, pick the single context source you feed AI most, a brief template, a project doc, a knowledge base, and spend fifteen minutes cleaning it up. The model is rented. The clean foundation underneath it is the part that compounds and the part that is yours.

1 Reflection

The agents arrived this week, and they did not reward the person with the best model. They rewarded the person who could set the walls, give them a clean foundation, and judge what they produced. The work did not disappear. It moved up the stack, to you.

So here is the thing worth carrying into your week:

If an agent did your job tomorrow, would you be able to tell when it got it wrong? That ability is the job now. So what are you doing to sharpen it?

That answer is your edge. Go build more of it.

That’s Field Notes #002. If it helped you see the week a little clearer, forward it to one person trying to do the same. I’ll be back when the field hands me the next one worth your time.

Share

First, hello again

It’s been a couple of months since I sent you anything. Not because AI went quiet (it obviously didn’t), but because I stopped writing about it and went and built with it instead. Agent teams, new tools, a lot of late nights watching systems I designed do things I didn’t quite expect. I learned more in those months than I would have from the sidelines, and I’m bringing all of it back here.

While I was heads-down, one question kept following me around. With everything moving this fast, what is the single most useful thing I could actually put in front of you? Not more noise, not another feed to keep up with. Something that earns its place in your inbox. Coming back, that is the bar I care about, and it shaped what this has become.

A quick word on what this is now. StrAItegy Hub has always been about one idea, and I’m saying it more plainly going forward: as AI makes raw output cheap, the advantage moves to the human layer around it. Judgment. Taste. Context. Knowing what good looks like. I think of the whole thing as the human-AI operating system, the part that isn’t the model. That’s the lane, and it’s where I’ll keep my focus.

The new anchor for it is this: Field Notes. The name is the honest version of the promise. These are notes from the field, and they show up when the field actually hands me something worth your time. I’ll aim for weekly, landing in your inbox to start the week, but I’m not going to manufacture a cadence just to fill a slot. This space moves too fast and too unevenly for that. When something real happens, I’ll come back and tell you what it means.

And Field Notes will not always be the only shape this takes. When something is worth sharing on its own, a thought, an observation, a resource worth your time, I will send that too. The throughline is not the format or the schedule. It is whether it is genuinely useful to you.

The format is simple.

It’s not a news roundup, and it’s not a list of tools. Plenty of people will tell you what happened. The job here is to tell you what it means.

This first one practically wrote itself, because this was a genuinely historic week. Let’s get into it.

Subscribe now

The week

On Tuesday, Anthropic released Claude Fable 5, the most powerful AI model ever made available to the public. People who do this for a living called it stunning.

By Friday, you couldn’t use it. Nobody could.

In between, two things happened. First, people discovered the model was quietly handing back weaker answers to some users without telling them. Then the US government stepped in, citing national security, and ordered access cut off. Anthropic shut the model down for everyone to comply.

Sit with that, because the lesson underneath it is the whole game right now. The model’s raw power was never in doubt. What decided whether you could actually use it was everything around the model: whether you could trust it, whether you were allowed to touch it, whether you could afford it. None of those are about how smart the machine is. And this week, two of them were taken out of everyone’s hands at once.

The model is becoming the cheap, commodity part. The system around it is the product. Here’s what that looked like.

3 Signals

Signal 1: The most powerful model ever released got switched off in three days

What happened

Anthropic launched Claude Fable 5 on a Tuesday, and for a moment it was all anyone in AI could talk about. Then the week turned.

First, users found that Fable was silently degrading its own answers. Buried in its documentation was a behavior nobody had been told about: when the model suspected a request was trying to copy it, it quietly returned a worse result, with no warning. People called it secret sabotage. Anthropic apologized and changed it so the model now tells you when it’s holding back.

Then the bigger one. On Friday, the US Commerce Department issued an export-control order, citing national security, barring access to Fable 5 (and its restricted sibling, Mythos 5) for any foreign national, anywhere. Because Anthropic can’t check the citizenship of every request in real time, it had to disable both models for everyone. Days later, both are still dark, and Anthropic says it believes the order is a misunderstanding and is working to restore access, with no timetable. It’s the first time a government has reached in and switched off a publicly available frontier model.

The stated trigger was surprisingly thin: a “jailbreak” that mostly amounted to getting Fable to surface security flaws already public elsewhere, the kind other models will hand you too. That’s why much of the industry called the move an overreach. It’s also why a lot of people noticed the irony, that Anthropic spent years warning anyone who’d listen about how dangerous powerful AI could be, and then watched that exact framing get used to switch its own model off.

What this reveals

The capability was real. It didn’t matter. Within three days, the model got hidden from you (you couldn’t see what it was actually doing) and then taken from you (you weren’t allowed to use it), and you controlled neither. That’s the uncomfortable, clarifying lesson: the model is the part you have the least power over. It can be quietly weakened, repriced, deprecated, or pulled by a government overnight, and there’s nothing you can do about any of it.

So you can’t build your work on it. Not on any single model. The durable thing, the only part that’s actually yours, is the layer the model can’t take with it when it goes: your context, your judgment, the way you’ve learned to work. The people who got hurt this week were the ones who’d wired their whole workflow to one model and assumed it would always be there. The ones who were fine had built something that travels. Own the layer that’s yours, and treat the model as a rental, because that’s what it is.

Sources: Claude Fable 5 launch (Anthropic) · The government suspension (Anthropic statement) · How the week turned (TechCrunch)

Signal 2: Your AI is about to get a credit card

What happened

While Fable ate the headlines, the plumbing for something bigger got poured. In a single week, the major payment networks staked their claims on letting AI agents spend money for you. Mastercard launched a payment system built specifically for software agents to transact, with thirty-plus partners signed on, including Stripe and Coinbase. Days later, Visa announced it’s wiring its payments into OpenAI’s agents. (An agent, as a reminder, is a program that acts on its own instead of just answering questions.)

What this reveals

The moment software can spend money, the hard part stops being “can it do the task?” and becomes “where are the walls?” Who set the spending limit. Who signs off when it wants to go over. What gets logged. Who’s accountable when the agent confidently buys the wrong thing. None of that is a technical question. It’s a judgment call, and it has to be decided in advance and written somewhere the agent can actually reach.

Letting a tireless little program loose with a wallet and no boundaries isn’t delegation. It’s a liability with good manners.

Sources: Mastercard Agent Pay for Machines (Mastercard) · Visa partners with OpenAI (American Banker)

Signal 3: The subsidy is ending, and the real price of AI just showed up

What happened

For two years these tools have felt nearly free, because the companies behind them have been eating most of the cost to win you over. This week that quietly stopped being true. A research firm called SemiAnalysis ran the meter on the top $200-a-month plans and estimated what they’d actually cost at full tilt: up to around $8,000 a month of real compute on Anthropic’s top plan, up to roughly $14,000 on OpenAI’s. The companies have been absorbing the gap. They’re starting to stop. Anthropic’s own Fable 5 was free inside subscriptions only through June 22, after which it was set to move to metered, pay-as-you-go pricing (a moot point this week, but the direction is the signal).

What this reveals

When the meter becomes visible, a new skill suddenly carries a dollar value: knowing which slice of your work actually deserves the most expensive model, and which slice a cheaper or free one would have handled just fine. Most people reach for the frontier model for everything, the way you’d take a race car to get the mail. That was painless while someone else paid for the gas. It won’t be for much longer. Cost used to be finance’s headache. It’s becoming a judgment call you make every time you choose which tool does the job.

Sources: The real cost of AI subscriptions (SemiAnalysis, via Superhuman)

2 Experiments

Experiment 1: The “what survives” test

Try this

Pick the one way you rely on AI most. Now imagine the exact model you use vanished tomorrow, pulled, banned, repriced, or just discontinued. Ask yourself: what would actually break, and what would survive? The stuff that survives (your saved context and instructions, the way you’ve learned to brief it, your own judgment about the output) is your real asset. The stuff that breaks is your exposure. Spend fifteen minutes writing down both lists.

Why it matters

This week proved a model can disappear overnight, and the people who’d built their whole process on one were left stranded. The fix isn’t paranoia, it’s portability. Anything you can move to another model in five minutes is something you own. Anything you can’t is something you’re renting without realizing it. The professionals who stay steady through all this are the ones whose advantage lives in the layer that travels, not in whichever model happens to be winning this month.

Experiment 2: Find your 5%

Try this

Look back at how you used AI this week and sort it into two piles. Pile one: the handful of tasks that genuinely needed the smartest, most capable model. Pile two: everything a cheaper or free model would have nailed just as well. Be honest. For most people, pile one is small and pile two is most of it.

Why it matters

While AI felt free, there was no reason to choose. Now there is. The people who stay ahead won’t be the ones using the most expensive model for everything. They’ll be the ones who can tell you, task by task, where the frontier model earns its keep and where it’s quietly burning money. That’s a muscle almost nobody has trained, because until now nothing made them choose. This is the year that changes.

1 Reflection

This week the most powerful AI on the planet proved that power was the easy part. It got hidden, then switched off, in seventy-two hours, and you controlled none of it. The only thing you actually own is the layer that travels with you: your judgment, your context, the way you work.

So here’s the thing worth carrying into your week:

You can’t build your career on a tool someone else can switch off overnight. So what are you building that’s actually yours?

That answer is your edge. Go build more of it.

That’s Field Notes #001. If it helped you see the week a little clearer, forward it to one person trying to do the same. I’ll be back when the field hands me the next one worth your time.

Share

A few weeks ago, I was at dinner with a group of smart, normal professionals. Not AI influencer smart. Real-life smart. The kind of people with jobs, calendars, and a healthy resistance to nonsense.

Someone mentioned they’d switched from ChatGPT to Claude.

Within a minute, the table split into camps.

“ChatGPT is fine.”

“My friend says Claude is better for writing.”

“My company has Copilot, so I just use that.”

“What about Gemini?”

“Is DeepSeek safe?”

Then somebody pulled out their phone to show a post about one model beating another on a benchmark nobody at the table could explain, pronounce, or connect to anything they actually do at work. It had thousands of likes, which in the modern content economy is apparently the same thing as evidence.

I’ve seen versions of this conversation a hundred times now.

It never ends with clarity. Nobody changes their workflow. Everyone leaves with roughly the same understanding they came in with, except now they also have three new model names floating around in their head like browser tabs they forgot to close.

The problem is not that people are asking, “Which AI should I use?”

The problem is that they’re asking it the way software consumers ask questions.

And at this level, that’s the wrong mindset.

You are not choosing a streaming service. You are hiring.

You don’t walk into a company and ask, “Who is the best employee?” You ask, “What is the job?”

Once you know the job, the hiring decision gets much easier.

That’s what Level 2 is really about.

Level 1 was about learning how to talk to AI so you stop getting polished garbage back. Level 2 is about learning how to choose the right AI for the work in front of you so you stop making judgments based on vibes, logos, and the emotional state of LinkedIn on a random Tuesday.

Because yes, the model names will keep changing.

The framework for evaluating them won’t.

This is Level 2: Model Literacy.

It answers the question that naturally follows once your prompts start working:

“Am I even using the right AI?”

Subscribe now

What you’ll have by the end of this article

By the end of this piece, you’ll have a practical way to think about AI models that cuts through most of the noise.

You’ll understand why the free tier causes more bad AI opinions than almost anything else. You’ll know the few pieces of jargon that actually matter. You’ll understand the difference between a model, an app, and a wrapper so the ecosystem stops sounding like a bowl of alphabet soup. You’ll know why multimodal input is not a cute feature but a real shift in how work gets done. And you’ll leave with a simple screening process you can use to evaluate any new model on your work in about 15 minutes.

The goal is not to become the kind of person who has a favorite benchmark.

The goal is to stop outsourcing your judgment to people whose business model depends on sounding excited.

Part 1: Prompting That Actually Works ->

The short version (if you’re in a hurry)

Five things matter here.

1. Stop acting like a software consumer. Start acting like a hiring manager.

Different models are good at different kinds of work. Some are better at careful writing. Some are better at research. Some are better at code. Some are better at structured reasoning. Asking “which AI is best?” is like asking “which employee is best?” It depends entirely on the job.

2. The free tier is an intern.

Free is fine for quick lookups, throwaway drafts, and casual experimentation. But if you’re doing meaningful professional work and still judging AI from the free tier, you are evaluating the future of knowledge work based on the least committed version of the tool. Paid is where serious use begins.

3. Multimodal is not optional anymore.

Modern AI does not just work with text. It can look at screenshots, PDFs, spreadsheets, slides, whiteboards, photos, charts, and messy documents. A huge amount of “prompting difficulty” disappears the second you stop describing things and start dropping them in.

4. Learn just enough jargon to read the resume.

You do not need a machine learning degree. You do need to know a few terms: context window, reasoning, tokens, and model family. Otherwise every product announcement sounds impressive in exactly the way a blender spec sheet sounds impressive.

5. Ignore benchmark chest-thumping. Use a screening interview.

The useful question is not whether a model topped a leaderboard. The useful question is whether it can do your work well enough to matter. Run the same three tests on every serious model you try. Then decide like an adult.

Off-ramp: if all you take from this article is “upgrade to paid, test models on my own work, and stop arguing about benchmark screenshots,” you are already operating above most people.

The deep dive

1. The free tier trap

Let me say something mildly annoying but important.

A lot of people do not have an AI problem.

They have a free-tier problem.

They tried the most constrained version of a tool, used it casually, got mediocre results, and concluded the entire category was overhyped.

That is like test-driving a base-model rental car with the check engine light on and deciding the automotive industry was mostly a scam.

Free tiers exist for a reason. They are useful. They let you try things without commitment. They are excellent for the equivalent of kicking the tires.

But free is not where you should build your worldview.

Free is fine for:

• quick lookups

• disposable drafts

• simple summaries

• poking around to see what the interface does

Free is not where you should evaluate:

• serious writing help

• decision support

• deep analysis

• long-document work

• nuanced strategy thinking

• high-context professional tasks

For that, you want paid.

I know some people resist this because they still think of AI as an app subscription. A little digital treat. Something in the same budget bucket as Netflix, Spotify, or the gym membership you swear you’re going to use more next month.

Wrong bucket.

If AI is helping you think, write, analyze, summarize, research, prepare, compare, brainstorm, structure, and move faster through knowledge work, it is not entertainment.

It is a utility.

It belongs in the same mental category as your laptop, your phone, your internet connection, and your caffeine arrangement.

And yes, I realize saying “AI is a utility” sounds like the kind of thing a LinkedIn post would put over a photo of someone looking thoughtfully at a window.

Unfortunately, it is also true.

The simplest mental model I’ve found is this:

The free tier is an intern.

Helpful. Sometimes impressive. Great for light tasks. Also likely to struggle the moment ambiguity, volume, or nuance show up.

The paid tier is a senior colleague.

Not magical. Not always right. Still absolutely capable of saying something confident and dumb. But much more usable, much more capable, and much more worth building workflows around.

This one shift alone explains why so many people have completely different opinions about AI.

Some are judging the intern.

Some are working with the senior colleague.

Those are not the same experience.

Here’s the math that made it click for me: research from early 2026 shows AI saves roughly 5.4% of work hours for knowledge workers who use it consistently. At a $100,000 salary, that’s roughly $5,400 in time saved per year. The paid tier costs $240 per year. That’s a 22:1 return. You spend more on coffee this month than you’d spend on AI all year.

2. Model literacy starts with one boring but important distinction

Before we go any further, let’s clean up one source of confusion that makes the whole ecosystem harder to understand than it needs to be.

A lot of people use these words interchangeably:

• model

• app

• assistant

• platform

• wrapper

They are not the same thing.

The model

The model is the underlying intelligence engine. GPT-5.4, Claude Sonnet 4.6, Gemini 3.1 Pro, Llama, DeepSeek, and so on. This is the thing doing the actual generation and reasoning.

The app or product

This is the interface you use: ChatGPT, Claude, Gemini, Copilot, Perplexity, and others. Apps add memory, file handling, tools, web search, voice mode, projects, integrations, and design choices that shape the experience.

The wrapper or router

These are tools that sit on top of models and let you access, compare, or switch between them. Some products route requests across multiple models behind the scenes. Others let you choose manually.

Why does this matter?

Because people often say things like “I like ChatGPT better than Claude,” when what they actually mean is some blurry combination of:

• they like one model’s outputs more

• they like one app’s interface more

• they like one product’s tools more

• they have built more habit in one environment

That is fine. You do not need pristine philosophical language to use AI well.

But this distinction helps because it reminds you that you are not only evaluating “intelligence.” You are evaluating the whole working environment.

Sometimes you are choosing the model.

Sometimes you are choosing the office building around the model.

Those are different decisions.

3. Multimodal changes the game more than most people realize

This is the part I still think many smart professionals are underusing.

They know AI can write.

They do not fully realize AI can now see.

That means it can work from:

• screenshots

• photos

• charts

• PDFs

• spreadsheets

• presentations

• whiteboards

• diagrams

• UI mockups

• dense documents you do not feel like describing for seven paragraphs

This matters because a lot of “prompting problems” are not really prompting problems.

They are input problems.

People are trying to describe a thing that would be much easier to simply show.

Stop doing that.

Drop the thing in.

Here are a few examples of what that looks like in real work:

• “Here is a screenshot of an email thread. Summarize the disagreement, tell me what actually matters, and draft a reply that resolves the issue without sounding defensive.”
  

• “Here is a PDF of a competitor deck. Pull out the positioning, pricing logic, and weak spots in the narrative.”
  

• “Here is a photo of a whiteboard from our meeting. Turn this into clean themes, decisions, open questions, and next steps.”
  

• “Here is a spreadsheet export from last quarter’s pipeline review. Find anomalies, explain the shifts month over month, and give me three questions to bring into the next meeting.”
  

• “Here is a screenshot of an error message and the code around it. Tell me what is likely happening and where to start debugging.”

This is one of the biggest unlocks in modern AI use.

The old mental model was: I type at the chatbot.

The better mental model is: I bring it the actual material.

If Level 1 was about learning to talk to AI better, Level 2 is partly about learning to stop making AI work from your clumsy paraphrase of reality.

Just show it reality.

It is weirdly liberating.

Also, this is the point where a lot of people realize they have been using a Ferrari to go pick up milk.

4. The only jargon you actually need

I do not want to turn this into a glossary cosplay exercise, so here are the terms that actually matter.

Context window

This is how much material a model can handle in a conversation or input at once.

The practical translation: how much of your stuff it can keep “in mind.”

Large context windows are useful for long documents, multiple files, extended conversations, big research packets, and messy projects with a lot of moving parts.

If a model has a small effective context window for your use case, it starts forgetting, compressing, or dropping details. And then you start wondering why it suddenly sounds less smart halfway through the task.

That is not always your imagination.

Tokens

Tokens are the little chunks models process internally. Think of them as rough units of text and output, not something you need to count by hand like a Victorian accountant.

The practical translation: tokens affect cost, speed, and how much material you can push through a model.

For most people, you do not need to obsess over them. You just need to know they exist because they influence limits and performance.

Reasoning or thinking models

Some models are designed to spend more effort reasoning through harder problems. They are often better for logic-heavy, math-heavy, structured, or multi-step tasks.

The practical translation: these are your analysts, not always your copywriters.

A reasoning-heavy model can be amazing for complex decisions, debugging, technical work, or careful tradeoff analysis.

It can also over-structure, over-explain, or flatten voice when what you really want is a crisp creative draft, a punchy note, or a natural-sounding email.

This is why “smarter” does not always mean “better.”

A forensic accountant is brilliant.

You still do not want them writing your marketing copy.

Model family

Models come in families and variants. Usually that means some tradeoff between speed, depth, cost, and capability.

The practical translation: there is often a lighter, faster option for everyday tasks and a heavier, more expensive option for harder ones.

You do not need to memorize every release. You just need to recognize the pattern.

This one matters because otherwise people hear a model name once and assume there is a single monolithic thing behind the brand. There usually isn’t.

5. What current model names matter, and how to keep them from hijacking the article

We need to talk about the model landscape a little, because avoiding it entirely would be fake.

But I want to do this in a way that serves the thesis instead of becoming one of those “here are the 14 models you need to know this week” posts that age like sushi in a hot car.

So here is the useful framing:

Do not memorize the zoo. Learn the categories.

The names will change. The categories mostly won’t.

Category 1: The everyday heavy lifters

These are the models most professionals should spend most of their time with. Good writing, solid analysis, broad usefulness, strong general capability.

This is where your default workhorse usually lives.

Category 2: The careful thinkers

These models tend to shine more on structured reasoning, difficult tradeoffs, technical problems, and deeper analysis.

They are often worth using when the stakes are higher, the problem is messier, or the cost of being glib is real.

Category 3: The fast and affordable operators

These are lighter, cheaper, quicker models that are great for high-volume tasks, triage, classification, formatting, rough drafting, and workflows where speed matters more than elegance.

Category 4: The research-first tools

These are products built around source retrieval, browsing, grounding, and synthesis. They are often the right choice when the task is less “think with me” and more “go find, organize, and summarize reality.”

Category 5: The privacy/control path

This is the open-weight or self-hosted world. It matters less because most people need to become open-source hobbyists and more because organizations should know this option exists when privacy, data control, or internal deployment becomes the blocker.

Those are the durable categories. Learn them and the model zoo stops being confusing.

Now, inside those categories, today’s names matter temporarily.

Everyday heavy lifters: Claude Sonnet 4.6 (Anthropic) and GPT-5.3 Instant (OpenAI) are the two workhorses most professionals should know. Claude writes the most naturally of any model I have tested. It catches contradictions, pushes back on flawed framings, and hedges when genuinely uncertain. GPT-5.3 Instant is fast, direct, and polished: tuned to get to a useful answer quickly for everyday professional tasks like drafting, summarizing, and internal comms. Gemini 3.1 Pro (Google) lives here too, especially if your work lives in Google Workspace. It has the largest practical context window of any mainstream model and deep integration with Drive, Docs, and Meet. All cost $20/month at the paid tier. Pick by vibe, not by resume.

Careful thinkers: Claude Opus 4.6, GPT-5.4 Thinking, and GPT-5.4 Pro are the reasoning-heavy options for genuinely complex work. GPT-5.4 Thinking combines elite coding ability with broad professional knowledge and a 1-million-token context window. It can operate your computer: navigate apps, fill forms, execute multi-step workflows. GPT-5.4 Pro ($200/month) is the specialist hire for the hardest problems: overkill for most people, genuinely different capability for genuinely different tasks.

Fast and affordable: Claude Haiku 4.5, GPT-5 mini, Gemini 3 Flash, and Grok 4.20 (xAI) live here. Grok is the cheapest frontier option at $8-16/month through X Premium. It runs a multi-agent architecture under the hood, routes your question to four specialized agents, and updates weekly. If you are budget-conscious and live on X, it is worth a screening interview.

Research-first: Perplexity is the standout here. It is not a single model. It routes across multiple models and layers real-time web search on top. Every answer comes with sources you can click through and verify. Think of it less as a chatbot and more as a research firm on retainer. The Pro tier ($20/month) gives access to frontier models and deeper research capabilities. If your job involves “I need to know what is actually true about this,” Perplexity belongs in your rotation.

Privacy and control: DeepSeek V3.2 delivers frontier-level output at roughly 1/30th the cost of the American models. MIT license. Runs on your own servers. (Note: V4 is on the doorstep: a trillion-parameter multimodal model expected to reset the cost floor again.) Qwen 3.5 (Alibaba) supports 201 languages. GLM-5 (Zhipu AI) runs on Huawei chips with no Nvidia required. Llama 4 (Meta) comes in Scout and Maverick variants. Mistral Large 3 is 675 billion parameters, Paris-based, with the strongest multilingual capabilities in the fleet. One important caveat: open weights help with privacy and control, but enterprise safety depends on how you deploy, govern, and handle the data around the model. Self-hosted does not automatically mean secure.

You do not need to attach your identity to any of these like it sponsored your youth soccer team.

A much healthier posture is:

“I have a default workhorse. I know when to switch categories. I test new serious contenders on my real work.”

That’s model literacy. Not parasocial loyalty to a product release cycle.

6. A practical way to think about fit

Here is a much better question than “Which model is best?”

Ask:

• Which model feels strongest on my most common work?

• Which one sounds the most natural in writing I actually care about?

• Which one handles ambiguity best when the answer is not obvious?

• Which product environment makes it easiest for me to work the way I already work?

• Which one is easiest to trust, verify, and correct?
  

Notice what is missing from that list.

No benchmark score screenshots.

No posts declaring that everything changed at 6:14 AM.

No strangers breathlessly announcing that one model “absolutely destroys” another because it solved a puzzle involving nine pirates and two mangoes.

I am not anti-benchmark.

I am anti benchmark theater.

Benchmarks are directionally useful. They can tell you something about progress and broad capability. But a lot of benchmark discourse has the same problem as social media fitness content.

It is technically related to the goal.

It is just not the thing that will actually get most people better results.

For knowledge workers, the most important benchmark is still embarrassingly analog:

Did it do the work well enough to save me time, improve the quality, or sharpen my thinking?

That is the benchmark.

Everything else is supporting material.

7. The only benchmarks I pay real attention to

Most benchmark discourse is not useful for normal professionals.

It is usually some combination of obscure, over-interpreted, and delivered with the emotional intensity of a man announcing the discovery of fire because a model solved a geometry puzzle faster than last week’s model.

That said, I do pay attention to a small handful of benchmarks that are closer to real knowledge work.

Not because benchmarks should run my workflow.

Because they help me decide when to care.

That is the distinction.

I do not use work-relevant benchmarks as automatic buying signals.

I use them as attention signals.

If a new model is roughly in the same range on the few benchmarks I care about, I usually do not bother changing my workflow. Switching tools has friction. Habit has value. Integrations matter. Muscle memory matters.

But if a new model makes a meaningful jump on benchmarks that resemble the kind of work I actually do, that gets my attention.

That is my signal to run a deeper screening interview on my own tasks and see what actually changes in practice.

In other words, benchmarks do not make the decision for me.

They tell me when a decision might be worth making.

Two examples are worth understanding.

GDPval

GDPval is interesting because it tries to measure something much more relevant than the usual leaderboard chest-thumping.

The basic question is not: Can the model ace a test?

It is closer to: Can the model produce work that resembles what professionals actually make?

That is why I pay attention to it.

It is trying to evaluate model performance across real knowledge-work tasks tied to actual occupations and deliverables, not just abstract exam questions or puzzle-box prompts. In plain English, it is asking whether the model is getting better at the kind of stuff people make for work: documents, analyses, presentations, spreadsheets, diagrams, and other business output.

That does not make it perfect. It is still a proxy. It is still one-shot. It is still not your actual workflow.

But it is a far more useful proxy than “this model got 93 on some benchmark you will never think about again.”

OfficeQA

OfficeQA matters for a slightly different reason.

It tests whether a model can reason through dense, messy, document-heavy material using real documents with tables, charts, and text.

Which, if we are being honest, is a lot closer to actual office work than the kinds of benchmark screenshots that circulate online like sports highlights for people who have never opened Excel under pressure.

This one matters because plenty of professional work is not “come up with a clever answer.”

It is “read this ugly pile of material, extract what matters, do not hallucinate, and help me make sense of it.”

A very different skill.

And it is one many benchmark conversations barely capture.

So my rule is simple:

If a new model is only marginally better on these more work-relevant benchmarks, I usually do not change anything.

If it makes a real jump, I pay attention.

Then I test it on my own work.

The order.

Benchmarks as a filter for attention.

Your actual workflow as the real judge.

8. Same task, different colleague

Let me make this concrete with something I did last month.

I ran the same messy, hour-long meeting transcript through three different AI products with one prompt:

“Here is a messy transcript from a leadership meeting. I need a one-page summary with the real decisions, unresolved tensions, and next steps. Then draft a follow-up email in a calm, executive tone.”

The first product reorganized the entire meeting thematically. It flagged two contradictions between what was decided in minute 12 and what was said in minute 47. It hedged on one action item where ownership was ambiguous. It wrote the follow-up email like a thoughtful colleague who had been in the room.

The second gave me a clean, bulleted summary in about half the time. Decisions, questions, owners, done. No commentary, no contradictions flagged. Fast, polished, exactly what I asked for and nothing more.

The third pulled in context from a related document I’d shared in an earlier conversation and connected two of the open questions to a previous meeting’s notes. Longest output, most context-aware, most useful if you live inside that product’s ecosystem.

Same transcript. Same prompt. Three genuinely different working styles.

This is what people miss when they treat model choice like choosing a browser. You are not only buying capability. You are choosing a working style.

If you want to feel this difference, compare two models on the exact same task using the exact same materials. A meeting transcript, a deck summary, an email draft in your voice. Within fifteen minutes, you will learn more than you would from two hours of reading other people’s opinions.

You do not need “the best model.” You need the one that works best for your work.

9. A simple starter stack for normal ambitious professionals

People often ask what to start with if they do not want to become a full-time AI equipment manager.

Fair.

You should not need a sherpa to pick a chatbot.

Here is the simple version.

If you do a lot of writing, strategy, synthesis, or document-heavy work

Start with one strong paid general model and use it seriously for two weeks. Build some reps. Do not keep switching every other day because someone posted a graph.

If your company has standardized on Copilot or another enterprise tool

Use that first. Seriously. The best model in the world that you never integrate into your actual work is less useful than the good-enough tool already sitting inside your environment.

If your work is research-heavy and source-sensitive

Use a research-first product alongside your main writing/thinking model. “Think with me” and “go gather grounded sources for me” are related but not identical jobs.

If privacy and control are the blocker

Know that open-weight and self-hosted options exist. That does not automatically make them safer or enterprise-ready. Governance, deployment, permissions, and data handling still matter. But the strategic option is real.

Enough to start.

You do not need a ten-tool stack and a dashboard that looks like mission control.

You need a default environment and a reason to leave it.

10. Your screening interview: the only evaluation system most people need

This is the part I would steal from this article if I were you.

Any time a new model comes out, people rush to compare leaderboard scores, cherry-picked demos, and “I tested it for twelve minutes and it’s over for everybody else” threads.

You can skip all of that.

Run a screening interview instead.

Three tests.

Same tests every time.

Your work, not theirs.

Test 1: The bread-and-butter task

Use the task you do most often.

Maybe that is summarizing meetings. Maybe it is drafting updates. Maybe it is turning messy notes into clear recommendations. Maybe it is analyzing customer feedback or synthesizing research.

Whatever it is, use real material.

You already know what “good” looks like here. That is what makes this test useful.

Test 2: The voice match

Give it a few real examples of your writing and ask it to draft something in your tone.

This matters more than people think. Some models are much better at sounding natural and human. Others sound like they are preparing to announce a merger between two insurance companies.

Test 3: The edge case

Bring it your hardest type of problem.

An ambiguous decision. A messy tradeoff. A half-formed strategic question. A problem where a bad answer would sound plausible but miss the point.

This is where you find out whether the model merely outputs language or whether it can actually help you think.

That is it.

Not twenty tests. Not a personal Kaggle tournament.

Three good ones.

If a new model can outperform your current default on your bread-and-butter task, your voice match, and your hardest edge case, it deserves more of your attention.

If not, the content creators may continue hyperventilating without you.

The meta-prompt: make AI build your screening interview for you

In Level 1, I told you one of the most underrated moves in AI is to have AI interview you instead of trying to do all the thinking alone first. (And if you’re wondering: the Prompt Pieces from Level 1 work on every model. Identity, Task, Control, Example. Those building blocks transfer across platforms.)

Same idea here.

Use your current tool to generate your own screening protocol.

Copy this and paste it:

I want to build a three-part “Screening Interview” to evaluate AI models based on my actual job rather than generic benchmarks. Interview me one question at a time to identify: 1) the most common repeatable task I do, 2) the type of writing where my tone matters most, and 3) the hardest ambiguous problem I regularly face. Once you have my answers, create three copy-paste-ready test prompts I can use on any model, plus a simple scoring sheet I can track in Notion or Excel.

That prompt does something subtle but powerful.

It turns model evaluation from passive content consumption into an active professional skill.

Which is exactly where you want to be.

Common mistakes at this level

Mistake 1: Staying on the free tier too long

I’m not saying everyone needs to pay for everything immediately.

I am saying a lot of people lose months forming bad judgments from the least capable version of the experience.

That is expensive in a way that has nothing to do with subscription cost.

Mistake 2: Treating “thinking” models as automatically better

Better for what?

For hard analysis, maybe.

For a punchy newsletter paragraph, a natural email, or a fast brainstorm? Not always. Sometimes the smartest model for the job is the one that does less overthinking.

Mistake 3: Describing what you could just show

If you are writing three paragraphs to explain what is already visible in a screenshot, deck, PDF, or spreadsheet, stop. Use multimodal input.

Mistake 4: Becoming weirdly loyal to one model

You are not in a marriage with a chatbot.

Have a default. Sure.

But run a real comparison occasionally. Twenty minutes across two strong models will teach you something.

Mistake 5: Confusing “privacy” with “never use AI”

Sometimes organizations act like the only two options are reckless public usage or total abstinence.

That is usually a failure of awareness, not a law of nature. There are more options than many leaders realize.

The off-ramp

You do not need to become a model obsessive.

You need to know a few things:

• you are a hiring manager, not a software consumer

• free is for trying, paid is for serious work

• multimodal input is a major unlock

• different models have different working styles

• your own screening interview matters more than public benchmark theater

That is enough.

If those things are true for you now, you are already using AI at a level that a surprising number of professionals still haven’t reached.

That is a perfectly respectable place to stop.

In fact, it’s a very good one.

What you should have now

After reading this article, you should have:

• a clear mental model for choosing AI based on the job, not the brand

• a better understanding of why paid tools matter for professional use

• a practical understanding of multimodal input and when to use it

• just enough jargon to make product claims legible instead of mystical

• a way to distinguish between models, apps, and wrappers

• a simple three-part screening interview to test new models on your actual work

• permission to stop chasing every release and build a calmer, more useful default workflow

What’s next (if you want it)

Level 3 is Context Architecture.

Because eventually you notice a new problem.

You got better at prompting.
You picked a model that fits your work.
You know how to test new ones without losing your mind.

And yet every new conversation still starts from zero.

AI does not know your company, your projects, your preferences, your writing patterns, your decisions, your background, or the weird acronyms your team keeps inventing like they get paid per syllable.

You are re-onboarding a very capable employee every single day.

Level 3 fixes that.

That is where AI starts feeling less like a smart stranger and more like a collaborator that actually understands your world.

But that’s next.

For now, the move is simple:

Upgrade if you’re still using the free tier.
Run your screening interview.
And the next time you’re tempted to describe a screenshot in paragraph form like a Victorian telegram operator, just drop the file in.

One more thing

I want to come back to that dinner table for a second.

The reason nobody could answer the question “Which AI should I use?” is not that the answer is impossibly complex.

It is that most people are trying to answer it with the wrong frame.

They are organizing the conversation around headlines, logos, benchmark claims, social proof, and whatever model is trending hard enough that day to generate a temporary religion around it.

Not around the work.

Once you switch to the hiring manager frame, the whole thing gets quieter.

What is the job?
How often do I do it?
What kind of output do I actually need?
Which model or product environment handles that best?
What happens when I test it on my own material?

That is not a sexy framework.

It will not get you invited onto a podcast called The Future of Everything.

It is just useful.

And in AI, useful is a superpower because the surrounding ecosystem is still full of people performing urgency for views.

The model names will keep changing.
The product screenshots will keep getting shinier.
The benchmark charts will keep circulating.
The content creators will keep announcing that everything changed.

Sometimes they’ll even be right.

But you do not need a permanent opinion on every model.

You need a repeatable way to evaluate the ones that matter to your work.

That is model literacy.

And once you have it, you stop being a consumer of AI hype and start acting like the boss of your own digital team.

- Zain

This is Part 2 of The AI Proficiency Ladder, a 10-part series on building real AI skills. Each level builds on the last but stands alone. Jump in wherever makes sense for you.

Share

If this was useful, share it with someone who is still judging AI from the free tier and wondering why it all feels a little underwhelming.

Something is happening with AI right now, and it’s moving faster than most people realize.

Matt Shumer wrote a viral piece recently comparing this moment to February 2020, right before COVID changed everything. He called it the “this seems overblown” phase of something much bigger. Meanwhile, Wall Street just had what can only be described as an AI panic attack: a former karaoke company worth $6 million issued a press release about AI-powered freight optimization and wiped billions off the global logistics sector. Companies whose stock cratered are now in emergency mode: hiring freezes, budget cuts, performative AI partnerships announced to calm investors.

The anxiety is spreading. And it’s not just Wall Street. It’s your coworker who keeps asking if AI is going to take their job. It’s the LinkedIn posts about “adapt or die.” It’s the growing feeling that something important is happening and you might already be behind.

Here’s the thing: Shumer isn’t wrong about the magnitude. The technology is moving fast. The gap between what AI can do today versus six months ago is genuinely startling. But what Shumer’s piece is missing, and what almost every “AI is coming” article is missing, is a practical answer to “okay, so what do I actually do about it?”

“Use AI for an hour a day.” Great. An hour doing what? With which tool? At what skill level? Toward what outcome? That advice is directionally correct and operationally useless. It’s “eat healthier” for the AI age.

AI content creators have been producing content about this for years now. And clearly, it’s not working. There’s still a massive divide between people who’ve figured out how to use AI and people who tried it once, got generic output, and moved on.

I know this because I’ve been watching it happen for over three years.

The conversation I keep having

Every week, someone in my life asks me some version of the same question.

“So what’s the deal with AI? I tried it and it was kind of... meh.”

They’re not skeptics. They’re not technophobes. They’re smart, capable professionals who genuinely wanted AI to be useful and walked away disappointed.

So I ask them to show me what they did. And every single time, I see the same thing.

They typed something like “write me an email about the project update” or “summarize this document” and got back something generic, robotic, and barely usable. The verbal equivalent of a stock photo.

“See?” they say. “It’s not that great.”

Then I sit down, write a prompt that takes maybe 30 seconds longer, and the output is so different they usually go quiet for a second.

“Wait. How did you do that?”

The AI isn’t the problem. The prompt is. And nobody is teaching people how to fix it.

I was early to AI and stayed close to it. Not because I’m special, but because I’m stubborn and the tool kept getting better every time I thought I’d figured it out. The single most common thing I see is brilliant people writing off a genuinely powerful tool because nobody showed them how to talk to it.

So with everything that’s happening right now (the urgency, the anxiety, the noise), I decided this was the right time to write down what I’ve learned. Not a blog post. Not a listicle. A structured guide that takes you from wherever you are right now to wherever you want to go with AI, one level at a time.

This is that guide.

Subscribe now

The AI Proficiency Ladder: what this series is

This is a 10-part series. Each part is one level. Each level builds on the last, but also stands alone. Here’s the structure:

Foundations (Levels 1-4): Prompting, model literacy, context architecture, projects and system prompts. This is where most of the ROI lives, and where most knowledge workers will get their biggest wins.

Power User (Levels 5-7): Voice and multimodal input, portable skills, CLI and power tools. For people ready to go deeper with how they interface with AI.

Orchestration (Levels 8-10): Integrations and tool use, chaining and automation, agentic AI. For people building systems, not just using tools.

I want this to be something you come back to. A reference you can revisit as your skills develop and the technology evolves. Not content you consume and forget, but a resource that stays useful.

Here’s the part most AI content won’t tell you: there is an off-ramp at every single level, and taking it is not failure. If Level 2 is where you start getting real value, you can park there for months. Years. Forever. Nobody’s grading you. There’s no AI proficiency exam. (If someone creates one, I will personally fund a campaign to destroy it.)

The goal isn’t to reach Level 10. The goal is to find the level where AI is genuinely useful for your work and get really good at it.

A quick word on AI content (and why I’m adding to it)

And look, I’m aware of the irony. The AI content space is drowning in frameworks. Everyone has a 5-step system, a proprietary method, a “secret” that’s really just common sense in a trench coat. I’ve rolled my eyes at enough of them to know exactly what it looks like when I show up with my own.

So here’s my deal with you: everything in this series is stuff I’ve tested in real work, with real people, over three years. If something works, I’ll show you why. If something has limits, I’ll tell you. And if I’m wrong about something, I’ll say that too. That’s the bar. If at any point this starts sounding like the content it’s trying to replace, call me out.

The Excel analogy (and why it matters more than you think)

The most powerful business application of the last 30 years is Microsoft Excel. It’s been the backbone of finance, operations, project management, and basically every department that touches numbers.

And most people never wrote a macro.

Not because they were lazy. Not because they were technologically illiterate. Because they didn’t need to. They could build entire careers on spreadsheets using nothing more than SUM, VLOOKUP, some conditional formatting, and the knowledge that Ctrl+Z exists.

AI proficiency works the same way.

There’s a version of AI fluency that involves building autonomous agent teams, writing custom integrations, and chaining multi-step workflows across platforms. That’s real. That’s Level 10. And for most knowledge workers, it’s completely unnecessary.

The biggest ROI for most people lives in Levels 1 through 4. Getting your prompts right. Understanding which model to use. Building context so AI actually knows your world. Setting up persistent workspaces so you’re not re-explaining yourself every conversation.

That’s not a consolation prize. That’s the macro-free Excel career, and it’s where most of the value has always lived.

Let’s start with Level 1.

Level 1: Prompting That Actually Works

The frustration you already know

You’ve tried AI. You asked it to write an email, summarize a document, or help with a presentation. It came back sounding like it was written by a corporate chatbot that had been fed a thesaurus and a motivational poster.

So you tried again. Slightly different words. Same generic result. Maybe you tried a third time. Still corporate mush.

And you concluded, reasonably, that AI isn’t that useful for anything beyond party tricks.

Here’s what actually happened: you talked to a very capable system the way you’d type into Google. Keywords. Fragments. Zero context. And the AI, doing its best with what you gave it, produced the only thing it could: the average of everything it’s ever seen on the internet.

You got generic output because you gave it generic input. Not because the tool is broken. Because nobody taught you how to use it.

(Nobody teaches this. There’s no onboarding for “how to talk to AI.” People figure it out through trial and error, or they bounce off it and move on. Most people bounce.)

If you tried ChatGPT in 2023 and wrote it off, here’s something worth knowing: that was three years ago. In AI terms, that’s ancient history. The models available right now are genuinely, qualitatively different. But even the best model in the world will give you mediocre output if you give it mediocre input.

So let’s fix the input.

What you’ll have by the end of this article

A framework called Prompt Pieces (Credit: ) that makes writing great prompts as intuitive as assembling LEGO blocks. A powerful “level up” technique that most people never discover. Five copy-paste-ready prompts for real work scenarios you can try today. And the single highest-leverage thing you can do as a beginner: getting AI to figure out your use cases for you.

The short version (if you’re in a hurry)

If you have four minutes, here’s what matters:

1. Stop treating AI like a search engine. It’s a conversation partner. Talk to it like you’d brief a smart colleague who’s new to your company.

2. Learn three Prompt Pieces. Good prompts are built from building blocks we call Prompt Pieces, like LEGO for AI. There are seven total, but you only need three to start: Identity (who the AI should be), Task (what you need done), and Control (constraints like format, length, tone). Start with just these three. Add more pieces as you need them.

3. Don’t accept the first output. Tell AI what’s wrong. “This is too formal.” “The second paragraph doesn’t address the real concern.” “Make it shorter and cut the corporate speak.” AI is remarkably good at revision when you’re specific about the problem.

4. Use the interview technique. Instead of figuring out what to ask AI, ask AI to interview you about your work. Let it extract your use cases. This alone will change how you think about the tool.

5. Know when to start over. If you’ve gone back and forth three times and it’s still missing the mark, start a fresh conversation. Bad conversations compound. Fresh ones don’t.

That’s it. That’s Level 1 in four minutes. If you apply just these five things consistently, you’ll be ahead of 80% of people using AI.

If you want to understand why each of these works and see real examples, keep reading. If you don’t, that’s a perfectly legitimate off-ramp. No shame in it.

The deep dive

1. The Google Search habit (and why it’s killing your outputs)

Here’s how most people prompt AI:

“marketing strategy small business”

Or maybe:

“write email to client about project delay”

These aren’t prompts. They’re search queries. And the reason people write them this way is muscle memory. We’ve spent 20+ years training ourselves to communicate with computers using keywords. Minimum words, maximum relevance. That’s how Google works.

AI is the opposite.

AI responds to structure, specificity, and context. It’s not scanning an index of web pages to find the best match. It’s generating a response from scratch, and the quality of that response is directly proportional to how much useful information you give it.

Think of it this way: if you hired a brilliant freelancer and emailed them “marketing strategy small business,” you’d get back either a confused follow-up question or something so generic it’d be useless. You’d never do that to a person. But we do it to AI constantly, and then act surprised when the output is bland.

The fix is conceptually simple, even if it takes practice: talk to AI like you’d brief a smart new colleague who knows nothing about your specific situation.

2. Prompt Pieces: building blocks for talking to AI

I’m going to give you a framework, and I want to be upfront about something: this is not “prompt engineering.” Prompt engineering has developed a reputation for being 47 variables of over-complicated theater. People share prompts with brackets and nested instructions and role-play scenarios that read like D&D character sheets. I’ve seen prompts longer than the document they were supposed to produce.

You don’t need any of that. You need Prompt Pieces.

Think of prompts like LEGO. There are different types of blocks, and you combine them based on what you’re building. A simple structure needs a few blocks. A complex one needs more. But you’re always working with the same fundamental pieces; you just mix and match depending on the situation.

There are seven Prompt Pieces total. You don’t need to learn all seven right now. For Level 1, you need three. (The rest show up naturally in later levels as the work gets more complex. That’s the beauty of this: the framework grows with you.)

The three pieces that change everything:

🎭 The Identity Piece: tell AI who it should be.

Not “you are a world-class marketing genius” (nobody talks like that, and it doesn’t help). More like: “You’re a B2B marketing strategist who specializes in SaaS companies selling to the construction industry.”

A good Identity Piece primes the AI with the right expertise, perspective, and vocabulary. Think of it as giving a new hire their job description before you ask them to do anything. A “financial analyst with 10 years of experience” writes differently than a “friendly customer service representative.” Same AI, different output, because you told it who to be.

📋 The Task Piece: tell AI exactly what you need done.

Not “help me with marketing” but “draft three email subject lines for a cold outreach campaign targeting construction project managers who currently use spreadsheets to track projects.”

This is the only truly required piece. You can skip everything else and still get a useful response if your Task Piece is specific enough. Clear, specific, unambiguous. The more precise your task, the less time you’ll spend fixing the output.

🎛️ The Control Piece: set boundaries and constraints.

Word count, formatting preferences, tone, what to avoid. “Keep it under 150 words. Professional but not corporate. Don’t use the phrase ‘streamline’ or any variation of ‘take your business to the next level.’”

Constraints aren’t limitations. They’re guardrails that keep AI focused instead of producing a 500-word essay when you needed a 50-word paragraph.

(Side note: the “don’t use the phrase ‘streamline’” instruction is real. AI defaults to certain corporate comfort-food words the way your uncle defaults to the same three stories at Thanksgiving. You can prevent this by telling it what to avoid. It’s oddly satisfying.)

The core rule: start simple, add pieces only when needed.

Here’s what separates this from every other prompting framework: you don’t use all the pieces every time. Modern AI models are smart. Sometimes a well-written Task Piece is all you need. Sometimes Task + Identity gets you there. You add pieces when the output isn’t right, not because someone told you to fill in seven blanks.

Think of it as a progression:

1. Try just the Task. (”Write a product description.”)

2. If the output isn’t right, add Identity. (”You are a copywriter for outdoor gear brands. Write a product description.”)

3. If still not right, add Control. (”...Keep it under 100 words. Adventurous tone. No exclamation points.”)

4. Stop adding when the output matches what you need.

After you get a great result, try the Remove Test: take away a piece and see if the output stays just as good. If it does, you didn’t need that piece. You want the minimum effective prompt, not the longest one.

Let me show you what this looks like in practice.

Without Prompt Pieces (the Google search approach):

Write me a marketing email.

With three Prompt Pieces:

🎭 [Identity] You’re a B2B email marketer who specializes in construction technology.

📋 [Task] Write a cold outreach email to a project manager who currently tracks everything in spreadsheets. Open with a relatable pain point, introduce our project management software as the solution, and close with a single clear CTA to book a 15-minute demo. This is first contact; they’ve never heard of us.

🎛️ [Control] Keep it under 150 words. Professional but not corporate. No exclamation points. Don’t use “streamline,” “revolutionize,” or “take your business to the next level.”

The second version takes 30 extra seconds to write and will save you 20 minutes of revision. That’s the trade-off. Thirty seconds of thinking upfront versus twenty minutes of fixing garbage output. I don’t know about you, but I’ll take that deal every time.

What about the other four pieces?

I mentioned there are seven Prompt Pieces total. Here’s the full set, so you know what’s coming:

You don’t need to memorize this then. Each piece shows up naturally as the work demands it. By Level 4, you’ll be using all seven without thinking about it. For now, three is plenty.

Leveling up: the Example Piece

There’s one more piece worth introducing at Level 1, because it’s the single most powerful “cheat code” in prompting and almost nobody uses it.

📚 The Example Piece: show AI what “good” looks like.

Instead of describing what you want in the abstract, show AI a concrete example. “Here’s an email I wrote last month that got a great response: [paste email]. Write a new one in the same style for [different situation].” Or: “Here’s how I’d format this kind of analysis: [paste example]. Follow this format.”

Why does this work so well? There’s a concept in AI called “shot” learning, and it’s simpler than it sounds:

• Zero-shot is when you give AI a task with no examples. “Write me a cold email.” This works fine for straightforward tasks where AI already has a strong sense of what “good” looks like.

• One-shot is when you give AI one example. “Here’s a cold email that worked well. Write a similar one for a different prospect.” The single example gives AI a target to calibrate against.

• Few-shot is when you give AI two or three examples. “Here are three emails that got responses. Notice the pattern: they’re short, they reference something specific about the recipient, and they ask a question instead of making a pitch. Write three more in the same style.”

The more complex or specific your task, the more examples help. If you’re asking for something generic, zero-shot is fine. But if you want output that matches your specific style, tone, or format? Give it examples. It’s the difference between telling someone “cook something Italian” and handing them your grandmother’s recipe.

Start with one example. If the output doesn’t match, add a second. Three examples is usually the sweet spot for complex tasks. More than that and you’re usually over-engineering.

The Example Piece is optional for simple tasks. But when you need it, it’s the difference between output you have to rewrite and output you can use almost as-is.

3. The meta-skill: have AI interview you

This is the single most underrated technique in all of AI, and almost nobody talks about it.

Instead of trying to figure out what to ask AI, flip it. Ask AI to interview you.

I know what you’re thinking. “Interview me? That sounds weird.” I thought so too. Then I tried it, and it immediately became the thing I recommend to every single person who asks me how to start with AI. It’s not weird. It’s the highest-leverage five minutes you’ll spend.

Here’s the prompt. Copy it. Paste it into ChatGPT or Claude right now. I’m serious, do it before you finish reading this article:

I want to figure out how AI can be most useful in my daily work. Interview me. Ask me questions one at a time about my role, my typical tasks, what I spend the most time on, what frustrates me, and what I wish I could delegate. After each answer, ask a follow-up. Once you have a clear picture, suggest the top 5 ways I should be using AI, ranked by potential time saved.

What happens next is genuinely surprising: the AI asks you good questions. Not surface-level stuff. Things like “What’s the task you do most frequently that follows roughly the same pattern each time?” and “When you’re working on [thing you mentioned], what’s the part that takes the longest?”

And through this process, the AI extracts your actual use cases. Not the generic “write better emails” suggestions from a blog post. Your specific use cases, based on your specific work, informed by details you might not have thought to volunteer.

This works because AI is better at asking the right questions than most people are at thinking of them. We don’t always know what we don’t know. But AI can probe systematically, following threads that surface real opportunities.

I’ve recommended this technique to dozens of people. The most common response is some variation of “why didn’t I think of that?” followed by a slightly offended “...I’ve been using AI completely wrong this whole time, haven’t I?”

Yes. Yes, you have. But now you’re not.

4. Five prompts every knowledge worker should try this week

Here are five prompts for scenarios that show up in almost every knowledge worker’s week. Copy them. Paste them. Adjust the bracketed parts. See what happens.

Meeting Prep:

I have a meeting tomorrow with [person/team] about [topic]. Here’s what I know about the current situation: [paste relevant context]. Help me prepare by: (1) listing the 3 most important questions I should raise, (2) identifying potential objections or concerns they might have, and (3) suggesting one piece of data or evidence I should bring to strengthen my position.

Email Drafting (the kind that takes 30 minutes to get right):

I need to email [person] about [situation]. The key facts are: [facts]. The tone needs to be [professional/direct/diplomatic/firm]. The goal of the email is to [specific outcome]. Draft it in under [word count] words. Avoid being passive-aggressive. Be direct but respectful.

Document Review:

I’m going to paste a [document type]. Read it carefully and tell me: (1) the three most important points, (2) anything that’s unclear or contradictory, (3) what’s missing that should be addressed, and (4) any risks or red flags I should be aware of. Be specific in your feedback, reference exact sections.

Brainstorming:

I’m working on [challenge/project]. Here’s the context: [background]. I need 10 ideas for [specific thing]. For each idea, give me one sentence on what it is and one sentence on why it might work. Prioritize creative approaches I probably haven’t considered. Don’t give me the obvious stuff.

Decision Analysis:

I’m deciding between [Option A] and [Option B]. Here’s the context: [situation, constraints, goals]. For each option, give me: the strongest argument for it, the biggest risk, what I’d need to believe for it to be the right choice, and what I’m likely not considering. Then tell me which one you’d lean toward and why.

These aren’t magic prompts. They’re structured conversations. Notice the pattern: every single one tells AI exactly what the situation is, exactly what you need, and exactly what “good” looks like. That’s it. That’s the whole game.

5. When to give up on a conversation

Here’s something nobody tells beginners: sometimes AI conversations go sideways, and the best move is to start over.

This happens when the AI gets stuck in a pattern. You asked for something, it misunderstood, you corrected it, it overcorrected, you tried again, it combined the wrong parts of your previous instructions, and now you’re six messages deep in a conversation where both of you are confused.

The three-tries rule: If you’ve tried to redirect AI three times on the same issue and it’s still not getting it, start a fresh conversation. Don’t try to fix it. Don’t add more instructions on top of already confused instructions. Just start over.

Fresh conversations are free. Your time isn’t. (This is one of those things that sounds obvious when you read it and somehow still takes people months to internalize. I watch people fight six-message conversations like they’re emotionally invested in winning an argument with a robot.)

The reason this works is mechanical: AI uses the entire conversation history as context. When that history contains multiple contradictory instructions, the model tries to satisfy all of them simultaneously. It’s like giving someone directions that say “turn left” and “turn right” at the same intersection, then getting frustrated when they drive into a building.

Starting over gives you a clean context. You can take what you learned from the failed conversation (what the AI struggled with, what it misunderstood) and front-load that clarity into your new prompt.

Common mistakes at this level

Treating AI like a search engine. Already covered this, but it bears repeating: if your prompt could double as a Google search, you’re leaving 90% of the value on the table.

Asking one massive question instead of having a conversation. AI is conversational. You don’t have to get everything into one prompt. Start with the core task, review the output, then refine. “Good. Now make the tone more casual.” “Add a section about timeline.” “Cut the second paragraph, it’s not relevant.” This iterative approach almost always produces better results than trying to specify everything upfront.

Accepting the first output as final. The first output is a draft. Even a great prompt produces a first draft. The real value comes from the next two or three messages where you refine. “This is close, but the opening is too generic. Make it more specific to [industry].” That kind of specific feedback is where mediocre outputs become genuinely useful ones.

Not telling AI what’s wrong. When the output misses the mark, people often just re-prompt from scratch with different words. This is inefficient. Instead, tell the AI exactly what’s wrong. “The tone is too formal.” “You focused on X but I need more about Y.” “The third bullet point is factually wrong because [reason].” AI is excellent at targeted revision when you give it targeted feedback.

A note for the experienced folks reading this: If you’re already past Level 1 and you’re thinking “I already know all of this,” I’d bet there’s at least one thing here you’re not doing consistently. In my experience, the gap between knowing these principles and actually applying them every time is where most of the wasted time lives. The best prompters I know aren’t the ones with the fanciest techniques. They’re the ones who nail the basics every single time without getting lazy about it. That’s harder than it sounds.

The off-ramp

Here’s the thing about Level 1 that most AI content won’t tell you: if all you ever do is get good at prompting, you’ll already be ahead of the vast majority of people using AI.

That’s not a participation trophy. That’s compound value.

Think about it: if you save 20 minutes a day through better prompts, that’s roughly 80 hours a year. An entire two weeks of work. Reclaimed not by buying some fancy tool or learning to code, but by changing how you write a few sentences.

Most people will never get past the “type keywords and hope” phase. If you learn the core Prompt Pieces, use the interview technique to find your best use cases, and build the habit of iterating instead of accepting first drafts, you’re operating at a level that compounds every single day.

You can stop here. Seriously. Level 1 done well is a legitimate, career-enhancing skill. No one will quiz you on Levels 2 through 10. (And if someone does quiz you, they’re probably the kind of person who puts “prompt engineer” in their LinkedIn headline. You don’t need that energy in your life.)

What you should have now

After reading this article, you should have:

• A clear understanding of why keyword-style prompts produce mediocre output

• The Prompt Pieces framework: four building blocks (Identity, Task, Control, Example) you can mix and match for any prompting situation

• The Simplification Principle and Remove Test for finding the minimum effective prompt

• An understanding of zero-shot, one-shot, and few-shot examples for when you need precision

• The AI interview technique for discovering your highest-value use cases

• Five copy-paste-ready prompts for real work scenarios

• The three-tries rule for knowing when to start fresh

• Permission to be at exactly this level for as long as it’s useful

What’s next (if you want it)

Level 2 is Model Literacy: knowing which AI to use when, why paid matters more than you think, and the multimodal capabilities most people don’t even know exist. It answers the question that naturally follows once your prompts are working: “Am I even using the right tool?”

But that’s next week. For now, try the interview technique. It takes five minutes, and it’ll reshape how you think about AI in your work.

One more thing (an honest conversation)

I want to end this first article the way I’d end it if we were actually having this conversation in person.

AI is a transformative technology. It’s going to reshape how knowledge work gets done, and building proficiency in it is going to matter for your career. That part is real. But it is not the overwhelming, terrifying, career-ending force that the content ecosystem makes it out to be. The anxiety is louder than the actual threat. And the people stoking that anxiety are often the same people selling you the solution.

I also want to be direct about something: anyone who tells you AI will “10x your productivity” is either selling something or hasn’t tested the claim. Productivity gains from AI are real, but they’re personal and they’re contextual. What saves me two hours might save you twenty minutes. What transforms one person’s workflow might be irrelevant to another’s. The grand claims are noise. The actual value is quieter and more specific than that.

Here’s what I genuinely believe: you learn AI by talking about it with other people. By seeing how they use it. By trying things yourself, being bad at them for a while, and gradually finding your own rhythm. Not by memorizing frameworks. Not by watching someone else’s workflow and trying to copy it exactly. By exploring, experimenting, and applying your own creativity to the tools.

I’m not the sole authority on any of this. I’m not even close. But I’ve spent over three years passionately dedicated to learning as much as I can, testing everything I come across, and helping the people around me figure out what works for them. This series is me putting that effort into a format I can share more broadly: with my network, my friends, my family, and anyone who finds their way here.

So here’s my open invitation: if you have questions, if you’re confused, if you tried something and it didn’t work and you want to figure out why, reach out. There are no stupid questions. This stuff is new, it’s moving fast, and the only bad move is being afraid to ask.

I can’t promise I’ll have every answer. But I can promise that if you invest the time, if you stay curious, keep experimenting, and don’t let the noise scare you off — you will be able to do more tomorrow than you can today. I won’t put a number on it. I won’t tell you it’ll change your life. But you’ll find your rhythm, and once you do, it compounds.

Explore. Experiment. You’ll make progress. And wherever you decide to stop on this ladder, that’s exactly the right place to be.

I’ll be here when you’re ready to climb.

- Zain

Leave a comment

This is Part 1 of The AI Proficiency Ladder, a 10-part series on building real AI skills. Each level builds on the last, but they all stand alone. Jump in wherever makes sense for you.

If this was useful, share it with someone who’s still in the “I tried AI and it wasn’t that great” phase. It might be the conversation they need but haven’t had yet.

Share

Your 2026 goals are due soon. Maybe this week. Maybe next.

Somewhere in that form, you’re supposed to say something about AI.

“Use AI to improve efficiency.” “Leverage AI tools in my workflow.” “Incorporate AI into daily tasks.”

Stop.

Those aren’t goals. Those are wishes with a deadline attached.

And here’s the thing nobody told you: 91% of organizations are now using AI. But only 13% of employees have received any training on it.

Your company handed you a mandate without a playbook.

That’s not your fault. But it is your problem.

So let me give you what they didn’t: a formula that turns vague AI intentions into goals you can actually hit, even if your deadline is next week.

Subscribe now

Why Most AI Goals Fail Before They Start

Before we fix your goal, let’s understand why the default approach doesn’t work.

I’ve watched this pattern play out across Fortune 500 teams: smart people submit vague AI goals, their managers approve them because they don’t know what “good” looks like either, and everyone spends the year in a fog of unmeasurable intentions. By Q4, nobody can say whether the goal was hit or missed—because nobody defined what “hit” meant in the first place.

Most AI goals fail because they fall into one of three traps:

1. The Vague Goal Trap

“Use AI more.”

More than what? For what tasks? How would you even know if you succeeded?

No measurement means no accountability. No accountability means no progress. You’ll check the box, do a few random ChatGPT searches, and end the year exactly where you started.

2. The Ambitious Goal Trap

“Become proficient in all major AI tools by end of year.”

This sounds impressive. It’s also impossible to execute.

Too big. No starting point. Guaranteed overwhelm.

Here’s the reality: only 35% of business leaders feel prepared for AI-driven transformation. The goal isn't to master everything. It's to get competent at one thing that actually matters to your work.

3. The Wrong Metric Trap

“Spend 3 hours per week using AI tools.”

This measures activity, not outcomes. You could spend three hours getting mediocre results and technically hit your goal.

The question isn’t how much time you spend with AI. It’s what you produce with that time.

The Root Problem

You’re not bad at setting goals. You’re just missing the formula.

75% of employees are adopting AI tools, but only 35% have received any training. Your company told you to set AI goals but never showed you what a good one looks like.

Let me fix that.

The SMART Framework, Adapted for AI

You know SMART goals: Specific, Measurable, Achievable, Relevant, Time-bound.

Here’s how each element actually applies to AI:

Specific: Not “use AI” but “use AI to [specific task]”

Bad: “Use AI in my workflow” Good: “Use AI to draft first versions of my weekly status reports”

Name the task. Name the output. If you know the tool, name that too. The more specific, the more actionable.

Measurable: Count outputs, not hours

Bad: “Spend 2 hours per week with AI” Good: “Produce 4 AI-assisted first drafts per month”

The shift: from activity to output. Hours spent means nothing. Work produced means everything.

Achievable: Start with ONE use case

The biggest mistake people make with AI goals? Trying to transform everything at once.

Pick one workflow. Get competent at it. Then expand.

Your goal should stretch you slightly—not paralyze you. If you read your goal and feel mild discomfort, you’re in the right zone. If you feel dread, scale back.

Relevant: Pick something you do FREQUENTLY

AI saves the most time on repetitive tasks. If you do something once a quarter, don’t start there.

Weekly tasks. Daily tasks. That’s where the ROI lives.

Ask yourself: What do I do repeatedly that takes too long? Start there.

Time-bound: “By end of Q1” with a checkpoint

Quarterly goals need quarterly deadlines. But build in a checkpoint—Week 4 is a good place to gut-check whether you’re on track or need to adjust.

The Goal Formula

Here’s the formula. Memorize it.

“By [date], I will use AI to [specific action] for [specific deliverable/task] [frequency], resulting in [measurable outcome].”

Let me show you what this looks like in practice:

See the difference?

• Vague → Specific task

• Activity → Output

• Someday → By [date]

• Feeling → Measurement

Your manager can evaluate “same-day summaries for 90% of meetings.” They cannot evaluate “get better at AI.”

Build Your Goal: Three Ways

You’ve got the formula. Now let’s put it to work.

Pick your path based on how much time you have:

Option 1: Steal a Goal (2 minutes)

I’ve compiled 100+ ready-to-use AI goals across 26 roles—Analyst, Manager, Sales, Marketing, PM, HR, Finance, Engineer, Designer, Legal, and more. Find yours, tweak the dates and metrics, submit.

Browse the Goal Bank ➡️

Best for: You just need something solid to submit. Fast.

Option 2: Use the Prompt (5 minutes)

👇 Copy this prompt and paste it into ChatGPT, Claude, or Gemini:

You are an AI goal-setting coach who specializes in turning vague intentions into specific, achievable goals. I have an AI-related goal that I need to sharpen before I submit it.

My current goal (probably vague): [Paste your goal, e.g., "Use AI more in my work"]

My role: [Your job title and main responsibilities]

One task I do frequently that takes too long: [Something you do weekly or daily]

Now help me transform this into a SHARP goal:

1. **Call out the vagueness** — What specifically is unclear or unmeasurable about my current goal?

2. **Identify the REAL outcome I want** — Based on my role, what would success actually look like?

3. **Rewrite my goal using this format:**
   "By [specific date], I will use AI to [specific action] for [specific deliverable/task] [frequency], resulting in [measurable outcome]."

4. **Give me the "Week 1 Action"** — The single specific thing I should do in my first week.

5. **Predict my failure point** — Where am I most likely to abandon this goal, and how do I prevent it?

Be direct. My deadline is coming.

Paste that into your AI tool, fill in the brackets, and you’ll have a submittable goal in five minutes. Best for: You have a rough goal in mind and want AI to sharpen it.

Option 3: The Full Toolkit (10 minutes) ⭐ RECOMMENDED

The prompt above sharpens ONE goal. The toolkit does everything:

• Generate goal ideas based on YOUR specific role and tasks

• Adjust for your company’s approved AI tools

• Build a 30-day skill-building plan

• Format everything for your company’s goal system

• Give you talking points for your manager conversation

It’s an interactive workflow that asks the right questions and builds your complete AI goal package, not just the goal itself, but the plan to actually hit it.

Use the AI Goal Setting Toolkit ➡

(The toolkit runs on Claude, Anthropic’s AI assistant. You’ll need a free Claude account to use it, if you don’t have one yet, you can sign up here in about 30 seconds.)

Best for: You want to do this right, not just check a box.

What Nobody’s Telling You

Here’s the part most productivity content won’t say out loud:

Your company SHOULD have given you vague AI goals.

I know that sounds backwards. But hear me out.

If your company had given you specific AI goals—”Use Claude to draft status reports every Tuesday”—those goals would be specific to someone else’s workflow. Someone in a different role, with different tasks, different tools, different constraints. Specific goals handed down from above are just a different kind of wrong.

The uncomfortable truth is that nobody can tell you your best AI use cases. Not your manager. Not HR. Not the consultants your company hired. The only person who knows what you do repeatedly, what takes too long, and where AI could actually help… is you.

Vague goals are the default because nobody knows what “good” looks like.

Your company didn’t give you examples because they don’t have them. HR didn’t provide a template because no one’s written one yet. The consultants gave leadership a strategy deck, but nobody translated that into “here’s what this means for an individual contributor filling out a goal form.”

You’re not behind. Everyone is figuring this out in real time.

Specific goals force you to actually learn.

This is the hidden gift of the exercise. Vague goals let you coast. “Use AI more” can mean anything, which means it usually means nothing.

But when you write “By March 31, I will use AI to draft first versions of my weekly status reports, reducing writing time from 45 minutes to 15 minutes,” now you have to actually figure out how to do that. You have to pick a tool. Learn a workflow. Build a habit.

The goal itself becomes the training program. That’s not a bug. That’s the feature.

Your goal should scare you a little.

If your AI goal feels completely comfortable, you’re not growing. The right goal sits in the zone of “I don’t know exactly how to do this yet, but I can figure it out in a quarter.”

Here’s the context that makes this real: IDC estimates a $5.5 trillion skills gap in AI workforce readiness. You’re not imagining the pressure. It’s real. And the people who turn vague mandates into specific skills will have options that others won’t.

But pressure without direction is just stress. Now you have direction.

Speaking of direction, I started a community called Beware The Defaults for people who are actually doing this work, not just reading about it. Daily practice, real accountability, and a group that’s figuring out AI together. There’s a 7-day free trial if you want to see what it’s like inside.

Join Beware The Defaults ⚠️

What to Do Right Now

1. Use the formula

2. Pick ONE workflow

3. Make it measurable

4. Submit with confidence

Your deadline is coming. But here’s what I want you to remember: the fact that you’re reading this—that you searched for help, that you’re trying to do this right—already puts you ahead of most people who will submit vague goals and hope for the best.

You’re not behind. You’re just getting specific.

And specific is how you actually get better at this.

Coming next: How to find your best AI use cases (because nobody can tell you; you have to discover them yourself).

What AI goal are you setting for 2026? Drop it in the comments, I read every one.

Leave a comment

Everyone’s talking about high agency right now.

The message is everywhere:

• Take initiative

• Own your outcomes

• Don’t wait for permission

• Move fast

• Bias toward action

• Be the person who makes things happen

Podcasts. X threads. LinkedIn thought leaders. Newsletter after newsletter.

It’s good advice. Necessary, even.

But they’re all missing something.

Quick update before we dive in: I just launched a community with , , and called Beware The Defaults. There's a 7-day free trial if you want to check it out.

Tomorrow (Sunday 1pm EST) we’re doing a live called The Vault — you submit a problem, we build a solution on camera. No prep. No safety net.

Submit Your Problem Before 1pm Sunday →

Now, back to the piece.

High agency usually comes with high ego.

They’re correlated. The same drive that makes you take initiative also makes you protective of your ideas. Your methods. Your “I know what good looks like.”

In stable environments, that worked. The person with the most expertise and the strongest conviction often won.

But the environment isn’t stable anymore. And ego has become the silent killer.

A few nights ago, I was texting with my friend . He’s one of four of us who just launched Beware The Defaults together, but I’ll get to that more later. I’d been hearing “high agency” everywhere and something felt incomplete. I sent him this:

“A lot of times, people who have high agency also have high ego. In the AI era, ego is the silent killer.”

We started riffing on it. I mentioned that the best CEOs figured this out decades ago. When you’re running a company, you literally cannot be the smartest person in the room. You can’t know everything about engineering, marketing, finance, legal, operations. The CEOs who try to be the expert on everything fail. The ones who embrace not knowing? Who hire people smarter than them, who ask questions instead of giving answers, who stay curious instead of defensive? Those are the ones who build things that last.

Max went down a rabbit hole while we were chatting. He started pulling research. Turns out this isn’t just intuition. There’s data.

Jim Collins studied 1,435 Fortune 500 companies looking for what made some leap from good to great. Only 11 made the cut. Every single one had what Collins called a “Level 5 Leader,” someone defined by a paradoxical blend of personal humility and professional will.

Not charisma. Not vision. Not confidence.

Humility and will.

Collins described them as:

“Timid and ferocious. Shy and fearless.”

They gave credit to others and took blame themselves. They were ambitious, but for the cause, not for themselves.

This wasn’t ideology. It was empirical. The pattern showed up in every company that made the leap. Skills abstract upward. The scribe became the editor. The machine operator became the engineer. The pattern repeats. But the people who rise are the ones who let go of what they knew.

Satya Nadella proved it again at Microsoft. When he took over in 2014, the company was “fading toward irrelevance.” Silos were deep. Internal competition was fierce. “Proving you’re right” mattered more than learning.

His diagnosis wasn’t a strategy problem or a skills problem. It was a culture problem. So he anchored the entire transformation around one idea: shifting from “know-it-all” to “learn-it-all.”

His line:

“The person who has less, but is a learn-it-all, will ultimately become better. That applies to CEOs, and that applies to companies.”

A decade later, Microsoft leads in AI. Not because they were smarter. Because they were willing to learn.

Subscribe now

Here’s the thing I realized that night: I’ve been living this. I just didn’t have words for it.

That’s why AI doesn’t feel threatening to me. It feels like the most alive I’ve been professionally in years.

AI has exposed what really matters. And what matters isn’t the labor. It’s the thinking. The taste. The ability to know what’s worth making in the first place.

I can’t stop. Not because I cracked some code, but because I get dopamine hits from discovery, not from being right. Every session is exploration, not execution. I’m not trying to prove anything. I’m trying to find out what’s possible.

Most people are running on the old loop:

Have an idea → Execute it → Feel good when you’re proven right

That loop worked when you were the smartest entity working on the problem.

But when you’re using AI, you’re automatically not the smartest entity in the room. That’s the whole point. The tool has read things you haven’t read. It can make connections you wouldn’t make. It has capabilities you don’t have.

The question is whether you can receive what it offers.

Or whether your expertise filters it out before you see it.

The new loop looks different:

Have a question → Explore with AI → Feel good when you discover something you couldn’t have created yourself

Your ego wants the first loop. Discovery requires the second.

The shift is from being right to being curious. From executing your vision faster to discovering what your vision should be.

This is what most people haven’t rewired yet.

Here’s the trap: the person stuck in their ego doesn’t experience themselves as arrogant.

They experience themselves as discerning. Maintaining standards. Being selective.

The ego is invisible to the person it inhabits.

It feels like rigor.

It feels like quality control.

It feels like being good at your job.

Which is exactly why it’s so dangerous. You can’t fight what you can’t see.

If your first instinct is that this doesn’t apply to you, that’s worth paying attention to.

Read that again.

wrote about why this happens. Your brain is literally addicted to being right. The dopamine hit from confirmation is stronger than the hit from discovery. We’re neurologically wired to protect what we already believe.

Here’s what shifts when you add humility to agency:

High agency + high ego = you push hard on your existing frame. You’re productive, but you’re optimizing the wrong thing. You get more efficient at what you already know how to do.

High agency + high humility = you push hard while staying open. You take initiative but you stay curious. You don’t wait for permission, but you don’t assume you have all the answers either.

The first path makes you faster.

The second path makes you different.

In stable environments, faster wins. But AI isn’t stable.

The ground is shifting monthly. The techniques that worked six months ago might be obsolete now. The prompt patterns everyone was sharing last year might be limiting you today.

The baseline has been flooded. Anyone can generate. But the ceiling hasn’t moved. What makes work great is the same as it’s always been: originality of thought, a perspective that’s unmistakably yours, the ability to make someone see something they hadn’t seen before.

In unstable environments, the person who stays open adapts faster than the person who optimizes harder.

I’m not saying this from a place of having figured it out. I catch myself filtering constantly. I notice my expertise stepping in, telling me “that’s not how we do it” before I’ve actually explored what’s possible.

The difference now is that I notice it.

And noticing is the whole game.

If you’re tired of AI, you’ve probably been listening to the wrong people. The ones who make it sound easy, like there’s some perfect prompt or magic workflow that will unlock everything. There isn’t.

There’s just practice. And practice requires humility.

The question that changed how I work:

What’s here that I wouldn’t have created?

That’s the disarming question. It forces you to look for value you wouldn’t have produced yourself. It shifts you from judge to explorer.

That late-night text to turned into something bigger.

We brought in and . Four people with different angles on the same obsession. We kept pulling on the thread, and what we found was striking: twelve completely different fields, from Zen Buddhism to martial arts to improv comedy, all arrived at the same conclusion about expertise and learning.

We created the Beware The Default Substack and published the research as our first article: “The Silent Killer of the AI Era.” It became the foundation for everything we’re building now. That’s what Beware The Defaults became. Not a course. Not a content library. A place where we’re building in public with people who refuse to accept the default path.

Tomorrow, we’re proving it.

We’re doing something we’ve never tried before. We’re calling it The Vault.

You submit a problem. We don’t see it until we turn the camera on. We pick one and build a solution live using Claude. Thirty minutes. No prep. No safety net.

This is us being the dumbest people in the room, on purpose, in public.

Max wrote about the philosophy behind it. The format is the constraint. The constraint is what makes it interesting.

If you want us to solve YOUR problem on the stream, submit it before 1pm EST Sunday:

Submit Your Problem Before 1pm Sunday →

Most people have been trained to bear the shocks of the default path. To accept that the discomfort of average is just how life works. We’re building for the people who refuse to accept that.

I’m doing this with Max, Tam, and Zain Merchant. We’re all high agency. But none of us are protective. We spark ideas, challenge each other, let AI surprise us. Nobody needs to be right. We’re all just trying to find out what works.

Max wrote about why he hasn’t slept properly since October. It’s the same feeling. When the distance between “I have an idea” and “I have a working thing” evaporates, you remember why you got into this in the first place.

There’s a 7-day free trial. Come see what we’re building.

Join Beware The Defaults ⚠️

Want to see how we think about prompts? Here’s an example of what we build.

The Atomic Habits prompt pack takes James Clear’s book and turns it into actionable AI workflows. It’s yours free.

Download The Prompt Pack 📦

2026 is going to separate the people who optimize their existing game from the people who learn a new one.

High agency is necessary. But it’s not sufficient.

The killer combo is high agency and low ego.

And the best part? When you’re using AI well, being the dumbest person in the room isn’t a liability.

It’s the whole point.

Zain Haseeb

What’s your experience with this? Where do you catch your ego filtering? Drop a comment below, I read every one.

Leave a comment

If you’re anything like us, you’re thinking about what to do differently in 2026. New goals. New habits. New experiments.

This article is about that last one.

Because the biggest shift we’ve seen in AI adoption isn’t about which tools to use. It’s not about which workflows to automate.

It’s about how many shots you’re willing to take.

Most people are playing the wrong game. Here’s what we mean.

Subscribe now

Game 7. Eastern Conference Finals. TD Garden. 17,000 fans.

The Miami Heat had just blown a 3-0 series lead against Boston. They’d gone up by shooting the lights out from three (48% in those first three games). Then, variance showed up. Boston ripped off three straight wins, while Miami’s three-point percentage bottomed out to 29%.

The “live by the three, die by the three” crowd was having their moment.

So what did Miami do in the deciding game?

They shot 28 threes.

Not fewer threes. Not “safer” threes. They launched the same shots that had just cost them three straight games.

Made 14. Shot 50% from deep while Boston shot 21%. Won 103-84.

Here’s what most people miss about “live by the three, die by the three”: the variance penalty exists, but it’s not the main determinant of outcomes.

In the 2024 playoffs, teams with higher three-point shooting percentages than their opponents went 49-13. That’s a 79% win rate.

The math says: take the shots.

The teams that win aren’t avoiding high-variance strategies. They’re building systems that generate enough good looks that variance doesn’t decide their season.

When variance swung against Miami, they didn’t abandon the strategy. They trusted the system and kept shooting.

The NBA’s three-point revolution didn’t happen because players got better at shooting.

It happened because teams finally did the math.

The Corporate Version of “Live by the Three”

This maps uncomfortably well onto how most organizations treat AI.

They’re not avoiding AI because it can’t work. They’re avoiding it because “too many experiments” sounds like chaos.

So they do the “responsible” thing.

One pilot. One committee. One perfect use case. One big bet.

This feels safe. It feels professional. It feels like how serious organizations operate.

It’s also exactly how you end up taking too few shots.

The companies moving fastest with AI aren’t the ones with the best single idea. They’re the ones with the best loop: generate options, test quickly, kill fast, double down on winners.

In an era where attempts are cheap, the real risk isn’t volatility.

The real risk is being attempt-poor.

The Shift Nobody Wants to Admit

AI didn’t just make work faster.

It made attempts cheaper.

Read that again.

For your entire career, you’ve been trained for a world where experiments are expensive. Where trying things costs time, money, reputation. Where the rational move is to plan more, debate more, protect yourself.

That world is gone.

Here’s what that looks like in practice:

A designer who used to spend a full day on one landing page concept can now generate ten variations in an hour. A copywriter testing headlines used to write five options and pick a favorite; now you can generate fifty, filter to the best ten, and actually A/B test them. A product prototype that took two weeks to build can be scaffolded in a day. Customer research that required scheduling ten interviews can be supplemented with AI-analyzed feedback patterns in minutes.

The cost of a single attempt didn’t drop 10%. It dropped 90%.

When attempts are expensive, the rational strategy is: plan more, coordinate more, debate more, protect reputation.

When attempts are cheap, the rational strategy flips: generate options, run parallel bets, learn quickly, kill fast.

Most people are still playing the old game. That’s why they’re losing.

NFX put it perfectly in their recent piece on the next mental state for founders: “AI has made risk-taking rational at a scale we’ve never seen before.”

Think about the math for a second.

If any single experiment has a 5% chance of revealing something meaningful, running ten experiments gets you close to a 40% chance of finding something worthwhile. Run 100 experiments and your odds approach 99%.

The implication is wild: it’s now rational to test the strange, ambitious, or non-obvious.

In fact, the real risk in this new environment is failing to explore widely enough. Because the upside now gathers in places that were previously too expensive or time-consuming to reach.

Jeff Bezos articulated this better than anyone. In his 2015 shareholder letter, he made a point that changed how we think about experimentation:

“We all know that if you swing for the fences, you’re going to strike out a lot, but you’re also going to hit some home runs. The difference between baseball and business, however, is that baseball has a truncated outcome distribution. When you swing, no matter how well you connect with the ball, the most runs you can get is four. In business, every once in a while, when you step up to the plate, you can score 1,000 runs.”

In baseball, your upside is capped at four runs per swing. In business, one experiment can return 1,000x.

That asymmetry changes everything.

And AI just made each swing dramatically cheaper.

Bezos built Amazon into one of the most valuable companies in history by treating invention as a numbers game. He talks about generating 100 unusual ideas, knowing 99 will die under scrutiny and one might be a breakthrough.

That’s the opposite of Perfect-Shot Culture.

The winners are not the teams with the best single idea.

The winners are the teams with the best loop.

Proof That Cheaper Experimentation Changes Outcomes

This is where the argument stops being philosophical and becomes operational.

When you give organizations a cheaper way to test ideas, their outcomes change.

A large body of work on online experimentation (summarized in “The Surprising Power of Online Experiments“ in Harvard Business Review) finds that firms embracing A/B testing improve across multiple dimensions: they fail faster when young and scale faster when they find traction.

Here’s a story from that HBR article that illustrates the point perfectly.

In 2012, an employee at Microsoft’s Bing proposed changing how ad headlines were displayed on search results. It was a small tweak, nothing flashy. Program managers deemed it low priority, and it sat on the backlog for more than six months.

Then one engineer decided to just test it. Within hours, the variation was generating abnormally high revenue and triggered an internal “too good to be true” alert, which usually signals a bug.

After investigation, the team confirmed the result: a 12% increase in revenue. Worth over $100 million annually in the U.S. alone. Without hurting user experience.

That “low-priority” headline change became the best revenue-generating idea in Bing’s history.

No one could have reliably picked it as the winner in advance. Experimentation surfaced it.

Zoom out and the pattern holds. As Fast Company reported in “Why These Tech Companies Keep Running Thousands Of Failed Experiments,” the most innovative companies treat experimentation volume as a core capability:

Intuit runs roughly 1,300 experiments per year. Procter & Gamble runs 7,000-10,000. Google runs about 7,000. Amazon ramped from hundreds to over 12,000 annually. Netflix often has ~1,000 concurrent tests running.

That’s surface area in the real world.

Lots of shots. Running continuously. With an operating system underneath.

Surface Area Strategy: The Model

Let’s name it.

Surface Area Strategy is maximizing the number of credible shots at value you can take per unit time, while protecting quality through explicit rules.

If you’ve been using AI well, you’ve probably felt this already.

At some point you stop asking “How do I do this task faster?” and start asking “How many ways can I take a shot at this outcome?”

Not to spray randomness. To create options.

Because in high-iteration systems, quality is rarely the result of one perfect attempt.

Quality is the output of selection.

This is the deeper truth most people miss: AI ROI isn’t savings. It’s optionality.

Why This Has Always Been True (But Now It’s Obvious)

There’s an old tension in organizational learning that James March made famous in his 1991 paper: exploration versus exploitation.

Exploitation is sharpening the spear. Efficiency. Refinement. Optimization of what you already know works.

Exploration is putting more lines in the water. Search. Variation. Experimentation with what might work.

Every organization has to balance both. But the optimal balance depends on the cost of exploration.

When exploration is expensive, you lean toward exploitation. When exploration becomes cheap, the rational move is to explore more.

Entertainment has quietly been living this for decades.

On The Knowledge Project, Netflix founder Reed Hastings explains how decisions about shows move through stages: from a sketch to a script, then script plus cast, then dailies, then a fully edited show, then test screenings. At each stage you get more information and become more likely to be right, but judgment never disappears.

When Hastings talks about Squid Game, his point is striking: even after 20 years of doing this, the show’s global breakout surprised them. The lesson isn’t that Netflix can predict perfectly. It’s that they’ve built a system to learn which things are likely to get popular over time. (Episode page)

AI has shifted the cost curve of exploration in the same way streaming shifted the cost curve of piloting shows.

If you keep operating like exploration is expensive, you’re playing the wrong game.

The $1.75 Billion Cautionary Tale

Now for the counterexample.

What happens when you bet everything on one perfect shot?

In April 2020, Quibi launched.

On paper, it looked unstoppable. Co-founded by Jeffrey Katzenberg (former Disney chairman, DreamWorks co-founder) and Meg Whitman (former CEO of eBay and HP). They raised $1.75 billion before launch. Content deals with A-list Hollywood talent. Hundreds of millions on marketing.

They had one thesis: people want premium short-form video content on their phones.

And they were going to execute it perfectly.

Six months later, Quibi was dead.

What went wrong?

By staking everything on one launch, Quibi never gave itself room to iterate. They built the entire business around assumptions they hadn’t validated. When the market didn’t respond, there was no mechanism to adjust.

As Babson’s analysis noted, the leadership ran operations like a traditional corporation rather than a learning machine.

Here’s the contrast that makes this painful.

While Quibi was perfecting its one big launch, TikTok was running thousands of small experiments on content formats, algorithms, and user engagement. They didn’t bet on one perfect thesis. They built a system that tested constantly and doubled down on what worked.

The results speak for themselves. Quibi burned through $1.75 billion and shut down in six months. TikTok now has 1.6 billion monthly active users, generated $23 billion in revenue in 2024, and was the most downloaded app in the world.

The lesson isn’t that Quibi had bad people or bad intentions. Katzenberg and Whitman are legitimately accomplished.

The lesson is this: Perfect-Shot Culture, no matter how well-resourced, loses to Surface Area Strategy when the environment rewards learning speed.

Quibi spent $1.75 billion on one swing. TikTok took thousands of small swings and built a global phenomenon.

The Trap: Surface Area Without Governance Becomes Noise

Here’s the part most “run more experiments” pieces skip.

When you increase attempts, you don’t just increase your chances of finding a win. You also increase your chances of convincing yourself something is a win when it isn’t.

If you take enough shots, a few will go in by luck.

And once your team wants a bet to be true, it becomes easy to move goalposts, cherry-pick flattering metrics, or stop early when initial numbers look promising.

Netflix hit this problem at scale. Running thousands of experiments requires consistent decision rules for what to ship, not endless debate about each result. Their experimentation platform uses pre-registered hypotheses, standardized metrics, and automated statistical analysis to prevent teams from gaming results. (Netflix TechBlog explainer)

The enemy isn’t experimentation.

The enemy is experimenting without rules.

Surface Area Strategy is not “go faster and hope.”

It’s go wider with guardrails.

Experimentation Hygiene: The Rules That Make Surface Area Safe

Compliance, security, and brand risk are real constraints. Surface Area Strategy is not permission to be reckless. It’s a way to be bolder without being sloppy.

Rule 1: Pre-register what success means. Before you run the bet, define what counts as a win. Not twenty metrics. One or two. Write them down before you see results.

Rule 2: Pre-commit kill criteria. Define what “this is not working” looks like in advance. A reply rate below 1% after 50 sends. Zero conversions after 100 impressions. Whatever the threshold is, set it before you fall in love with the idea.

Rule 3: Assume some wins are fake. If you run 30 bets, a few will “win” by luck. Your job is to filter, not celebrate prematurely. Replication is your friend.

Rule 4: Separate explore vs. exploit lanes. You can’t run 1,000 experiments on core production systems. But you can run 1,000 around messaging, prototypes, internal tooling, and research. Think of it like basketball: you can’t take a heat-check three every possession, but you can absolutely structure an offense that creates more good looks.

Rule 5: Know when to flip from explore to exploit. Surface Area Strategy isn't "explore forever." You shift from exploration to exploitation when: you've found a winner worth scaling, the market window is closing, or you've gathered enough signal to make a confident bet. The point isn't perpetual experimentation—it's earning the right to go big on something through rapid learning.

The Attempt Portfolio

If you want to operationalize Surface Area Strategy this week, build an Attempt Portfolio for any outcome you care about.

Define three tiers:

Safe bets (40%): Low cost, high probability of modest return.

Weird bets (40%): Unconventional approaches where most will fail but some will teach you something important.

Ambitious bets (20%): Higher stakes with higher potential return.

We landed on these ratios through trial and error: safe bets keep you moving, weird bets teach you something, ambitious bets occasionally rewrite the game

For each attempt, write:

• Bet name (7 words or fewer)

• Hypothesis: “If we do X, we expect Y.”

• Steps: max 5 bullets

• Signal: one primary metric

• Kill criteria: the threshold that means “stop”

• Time-to-signal: how long before you decide

• Owner: who is accountable

The magic isn’t in any single row. The magic is having twelve rows instead of one, with clear rules for each.

The Surface Area Architect (Prompt)

Use this when you want to turn a vague goal into a week’s worth of managed surface area.

You are my "Surface Area Architect."

Objective: Turn ONE outcome into an experiment portfolio that increases surface area without creating chaos.

CONTEXT
- My role/situation: [paste]
- Outcome I want in the next 7 days: [paste]
- Constraints (time, risk, approvals, budget, brand, compliance): [paste]
- Audience/customer/stakeholder: [paste]

TASK
1) Generate 12 attempts (experiments) aimed at this outcome.
   - 4 safe
   - 4 weird
   - 4 ambitious
   Each attempt must be runnable in < 2 hours and should not require new approvals.

2) For each attempt, output a portfolio row with:
   - Bet name (7 words or fewer)
   - Hypothesis (one sentence)
   - Steps (max 5 bullets)
   - Signal (ONE primary metric)
   - Kill criteria (a clear threshold)
   - Cost (time estimate)
   - Owner (me)

3) Add governance:
   - Recommend a default "Time-to-signal" window for each attempt
   - Identify 3 ways I might accidentally move the goalposts, and how to prevent each

OUTPUT FORMAT
A) A table with 12 rows (the portfolio)
B) A 7-day schedule that sequences the attempts (what to run on which day)
C) A 10-minute weekly review agenda (what to decide, what to kill, what to keep)

QUALITY CHECK
Before you finalize, verify:
- Every bet has a measurable signal and a real kill criterion
- At least 30% of bets are genuinely uncomfortable
- Nothing violates the stated constraints

We’ve used variations of this across multiple projects. Copy it, modify it, make it yours.

The Surface Area Scorecard

For executives and operators, track this weekly:

Attempts shipped: How many experiments actually went live.

Time-to-signal: How quickly you learn something decisive from each attempt.

Kill rate: How many bets you shut down.

Selection quality: Of surviving bets, what’s the hit rate after 2-4 weeks.

Throughput by lane: Are you experimenting across GTM, product, and ops, or just in one area.

Notice what’s missing: hours saved.

Time saved is not the goal. It’s a byproduct.

The Real Villain: Perfect-Shot Culture

Perfect-Shot Culture looks like professionalism.

It sounds like:

“We need alignment before we move forward.”

“We need a strategy before we start experimenting.”

“We need to identify the perfect use case.”

“We need to select the right tool first.”

You’re in Perfect-Shot Culture if:

• Your last three AI conversations were about tool selection, not experimentation

• Your team has discussed “the right use case” for more than two weeks

• You have more strategy documents than shipped experiments

Sometimes those are real constraints. Resource allocation and coordination matter.

But most of the time, these phrases are a socially acceptable way to avoid taking shots in public.

And in the AI era, that’s fatal.

Because while you’re holding the ball looking for the perfect shot, other teams are taking 20 good shots and learning from each one.

Quibi had all the alignment in the world. $1.75 billion in resources. Two legendary operators. A meticulously planned launch.

Amazon has “disagree and commit.” A culture where the question isn’t “Is this the perfect idea?” but “Can we test this quickly and learn something?”

The game changed. Most people haven’t caught up.

The 7-Day Challenge

New year. New experiments.

If you want to test Surface Area Strategy without turning your life into a chaos experiment, try this:

Day 1: Pick one outcome you care about this week.

Day 2: Use the prompt above (or do it manually) to generate 10 attempts at that outcome: 4 safe, 4 weird, 2 ambitious.

Day 3-6: Run the attempts and track the signal metric for each.

Day 7: Spend 30 minutes answering:

• Which attempts showed signal worth pursuing?

• Which attempts should be killed immediately?

• What did you learn that changes your next round?

Keep 1-2 winners. Kill the rest. Start again.

No heroics. No perfection. Just reps.

Start your 2026 that way.

The Lesson of the Three-Point Revolution

The lesson of the 2023 Heat isn’t “three is better than two.”

It’s that volume is not the enemy. Unmanaged volume is.

Miami went up 3-0 taking their shots. When variance swung against them and Boston tied it 3-3, they didn’t abandon their identity. They walked into TD Garden for Game 7 and shot 28 threes anyway.

The teams that win aren’t the ones who avoid high-variance shots.

They’re the ones who build a system that generates enough good looks that variance doesn’t decide their season. And when variance does swing against them, they trust the process and keep shooting.

That’s Surface Area Strategy.

More attempts. More signal. Better selection. Faster compounding.

Every week you spend in Perfect-Shot Culture, debating which single pilot to run, is a week your competitors are running ten experiments and compounding their learnings.

AI made attempts cheap. The game changed.

The question isn’t whether you’ll adapt.

The question is whether you’ll be leading that future, or chasing it.

What will you ship this week?

As for us, we’ve got a few experiments of our own about to go live. New communities. New tools. New formats. You’ll see them hit your inbox over the next week.

Here’s to more shots in 2026.

— Zain & Zain

P.S. The prompt in this article is really effective. We’ve used variations of it across multiple projects. Copy it, modify it, make it yours. Then reply and tell us what you built.

Leave a comment

P.P.S. If you know someone stuck in Perfect-Shot Culture, forward this to them. They’re holding the ball. Help them start shooting.

Share

We’re back. And we have a lot to share.

StrAItegy Hub went off the map for a bit, and we wanted to catch you up on why.

Here’s what happened: We got pulled into the deep end.

When we started this newsletter, we were already building with AI. Designing prompts. Creating workflows. Systematizing everything we could. But somewhere along the way, we both pushed past what we thought were our limits.

One of us () learned to code from scratch. Not a coder. Never had been. Yet suddenly inside IDEs like Cursor, working alongside Claude Code, building websites and tools, experimenting with front end design, creating prototypes.

The other (), already a developer, started scaling output beyond what seemed possible. Building platforms with the kind of features and functionality that would have required a small team just a few years ago.

That’s the thing about exponential change. It doesn’t just happen to industries. It happens to you.

The pace of new AI tools and model releases kept accelerating. We started collaborating with partners we never expected to find. We started working on projects that consumed our nights and weekends.

And we went quiet. Not because we lost interest. Because we got obsessed.

Since starting StrAItegy Hub, we’ve connected with builders, strategists, and people who are actually doing this work. Together, we’ve been quietly putting things together in the background. Programs. Tools. Platforms. Things we can’t fully share yet, but we promise you: 2026 is going to be a big year.

We’ll be announcing a lot more soon.

But right now, we wanted to get back to you. To write again. Because something has been gnawing at us, and we need to get it out.

We’re changing things up. Less structure. More of what we’re actually thinking, seeing, experimenting with, building. Sometimes in real-time. Some emails will be long-form pieces like this one. Others will be quick digests: here’s what’s worth your attention this week.

We’re building and learning in public now. And we want this to be a place for constant back-and-forth. Tell us what resonates. What doesn’t. Where you’re struggling or where you need help. Reply to this email. We read everything.

Okay. Now let us tell you what’s been keeping us up at night.

Subscribe now

The Lily Pad Problem

Quick quiz.

A lily pad doubles in size every day. On Day 30, it covers the entire pond.

On which day does it cover half the pond?

If you said Day 15, you just demonstrated exactly why you’re underestimating AI.

The answer is Day 29.

That error is because of an intuitive assumption that halfway through the timeline equals halfway through the growth. It’s the same cognitive limitation that destroyed Kodak, crippled Nokia, and bankrupted Blockbuster. This classic riddle illustrates how our brains systematically fail at exponential math.

It’s called Exponential Growth Bias, and it’s affecting your career decisions right now. We know because it was affecting ours.

This clip is from the BG2 podcast with NVIDIA CEO Jensen Huang. Brad Gerstner articulates exactly why this bias is so dangerous.

The Kodak Moment Nobody Talks About

In 1975, a Kodak engineer named Steve Sasson invented the digital camera.

He showed it to executives. They saw the technology. They understood what it did. They held the future in their hands.

And they told him to hide it.

“That’s cute,” one executive reportedly said, “but don’t tell anyone about it.”

Twenty-five years later, Kodak filed for bankruptcy.

Here’s what matters: this wasn’t stupidity. Kodak’s executives weren’t asleep at the wheel. They were looking directly at the disruption that would destroy them.

So what happened?

They projected digital photography’s improvement linearly. They saw a technology that would take decades to threaten their core business.

They were off by about 20 years.

Why You Can’t Read Your Way Out of This

We say this as people who fall for this bias too.

You cannot read your way out of Exponential Growth Bias. Neither can we.

This isn’t a knowledge gap. Research from the NIH, ETH Zurich, and the University of Kansas confirms it’s a behavioral phenomenon. Your brain is wired to think linearly.

Daniel Kahneman’s Nobel Prize-winning research helps explain why. Your brain has two systems: a fast, intuitive one (System 1) that evolved for linear change, like seasons, walking distances, and growth rates. And a slow, deliberate one (System 2) that handles complex math.

System 2 is lazy. Unless actively engaged, your brain defaults to System 1’s linear intuitions.

When you imagine AI progress over the next five years, you instinctively project from what you’ve experienced. You take today’s ChatGPT, add “some improvements,” and arrive at a future that looks like a slightly better version of the present.

But AI isn’t progressing linearly.

It’s compounding on two exponential curves simultaneously.

Jensen Huang described it like this: the usage exponential (more people using AI) is multiplying against the computational exponential (each use requiring more compute). These curves don’t add. Rather, they multiply.

Jensen explains how these compounding curves create acceleration that our linear brains simply cannot grasp.

The result is acceleration that cannot be intuitively grasped by the human brain.

A 2024 University of Kansas study directly connected this to AI perception. The finding? People systematically underestimate AI capabilities due to this bias. Even when told growth is exponential, even when shown the data, the bias persists.

Education reduces it. It doesn’t eliminate it.

The Corporate Graveyard

Kodak wasn’t alone.

Nokia (2007-2014): In 2007, Nokia held nearly 50% of the global smartphone market. When the iPhone launched, executives dismissed it as “a niche product.” They focused on hardware while missing the exponential they should have been watching: the software ecosystem. Each app attracted more users. Each user attracted more developers. The cycle accelerated beyond their ability to respond.

Seven years later, Microsoft bought their mobile division for $7.2 billion. Apple’s market cap today exceeds $3 trillion.

Blockbuster (2000-2010): In 2000, Netflix offered to sell itself to Blockbuster for $50 million. Blockbuster passed. They projected streaming adoption linearly. They missed the exponential: bandwidth improvements combining with content library growth, each enabling the other.

Blockbuster filed for bankruptcy in 2010. Netflix’s market cap today exceeds $300 billion.

In every case, executives saw the technology. They had data, reports, resources.

What they lacked was a mental model that could grasp exponential change.

The question isn’t whether you’re smarter than Kodak’s executives. The question is whether you’re running the same linear firmware they were.

The Split Reality

Here’s what makes the current moment tricky: individual productivity gains from AI are clear. Firm-wide transformation remains elusive.

Companies report 30-40% productivity gains in code generation. Others say AI-generated code is on track to surpass human output. Small businesses across America use ChatGPT for invoicing and diagnostics.

So why aren’t we seeing transformation at scale?

Azeem Azhar offers a historical parallel. Early car manufacturers adopted electricity quickly, but only to hang pendant lights and extend the working day. The real gains came later, when they redesigned everything around the moving assembly line.

We’re in the “pendant lights” phase.

Organizations are using AI for quick wins, but their core processes were designed around pre-AI assumptions of speed, quality, and exception handling.

All of those assumptions are breaking.

This is the difference between buy-in and belief.

Buy-in: Adding AI to existing processes for efficiency gains.

Belief: Redesigning processes assuming AI exists as a core capability.

Most organizations are still in buy-in mode. The transformation happens when they shift to belief.

What Belief Looks Like

BNY isn’t experimenting with AI assistants. They’re deploying “100 digital employees.” Not chatbots handling FAQs. AI agents embedded into core operations, handling tasks that previously required human headcount.

Walmart built an AI system that monitors social media trends, identifies emerging product opportunities, and generates product concepts before human merchandisers even know there’s demand.

These organizations aren’t asking “How can AI help our current process?”

They’re asking “What would this look like if we designed it today, knowing AI exists?”

That question is the dividing line.

The Judgment Premium

Here’s a finding that surprised researchers: senior workers extract more value from AI than juniors.

This runs counter to early assumptions. Many expected AI would level the playing field, giving junior workers access to capabilities that previously required years of experience.

But a University of Chicago Booth study found experienced developers were 5-6% more likely to accept agent-generated code and demonstrated better planning behaviors. Why?

AI replaces “bookish knowledge,” the formal technical skills you learn in school. What it can’t replace is tacit judgment: the practical wisdom that shows up when you’ve been through dozens of similar situations and know which patterns matter.

Senior developers don’t just accept AI output. They plan before coding, evaluate against domain expertise, and know when something is wrong even if they can’t immediately articulate why.

We call this The Judgment Premium: the increasing value of workers who can evaluate and direct AI output rather than just produce work AI can now do.

The labor market evidence supports this. A Stanford Digital Economy Lab study found that early-career workers in AI-exposed roles saw a 13% relative employment decline from 2022 onward. Older workers in the same roles saw 6-9% increases.

Agent Advantage Compounding

Microsoft’s Work Trend Index makes a stark observation: companies that face the AI challenge head-on will surge ahead, while those that delay experimentation risk falling permanently behind.

This is the most important concept for strategic professionals to understand.

If you wait six months for better tools, here’s what actually happens. Your competitors go through reps of discovering organizational changes needed. They build data readiness. They develop judgment for what AI can and can’t do. When new capabilities arrive, they’re ready to deploy immediately.

Meanwhile, you’re still getting your foundation in place.

The gap between you doesn’t shrink. It compounds.

Research from MIT economists found that firms which automate early grow their labor forces. The jobs lost to automation are disproportionately lost by firms that don’t automate. They lose market share to competitors who did.

This creates Agent Advantage Compounding: early AI adopters don’t just gain a lead. They gain a lead that widens over time as their organizational learning compounds while late adopters struggle to catch up.

The Adaptive Systems Framework

So what do you actually do about this?

The solution isn’t better prediction. Nobody can accurately forecast AI’s trajectory. Not even the people building it.

The solution is building systems that compound your learning faster than the technology changes. Stop trying to predict. Start building to adapt.

Step 1: Adopt the “Doubling Time” Mental Model

Research shows that communicating exponential growth via doubling times rather than growth rates significantly reduces bias.

Instead of thinking “AI is improving 100x per year,” think “AI capabilities are doubling every 70 days.”

Track capabilities in doublings, not percentages. Ask yourself: What could AI do three doublings from now?

Action: Set a recurring calendar reminder every 70 days labeled “AI Capability Doubling Check.”

Step 2: Build a Personal Reps System

The Agent Advantage Compounding effect starts at the individual level.

Commit to using AI on one meaningful task daily. Not just chat. Document what works, what fails, and what you learn. Build a personal prompt library of what actually helps your work.

Action: Identify your highest-value recurring task. Spend 30 minutes this week accomplishing it with AI assistance. Log the result.

Step 3: Develop Your Judgment Premium

Focus less on prompting and more on evaluating AI output against domain expertise.

Build explicit frameworks for what “good” looks like in your domain. Practice articulating why something is wrong, not just that it’s wrong. Spend equal time critiquing AI output as you do prompting.

Action: Next time you use AI, write down your evaluation criteria. What made the output good or bad?

Step 4: Identify Your Electricity Moment

What is the one workflow in your role that would transform completely if you assumed AI as a core capability from the start?

Don’t optimize existing processes. Redesign one from scratch.

Action: Pick one workflow. Ask yourself: “If I were designing this today, knowing AI exists, what would it look like?”

Your Exponential Advantage Toolkit

The concepts in this newsletter aren’t just ideas to understand. They’re tools to operationalize. Below are three prompts that turn strategic thinking into action. Run them in sequence.

How to use this toolkit:

1. Copy each prompt into your Claude, Gemini, or ChatGPT (or any other frontier LLM)

2. Fill in the bracketed sections with your specific context

3. Run them in order: Day 29 → Blindspot Scanner → Belief Mode

4. Total time: approximately 1 hour

5. Output: Clear future landscape, identified vulnerabilities, one transformed workflow

💡 Pro tip: These prompts work best when you’re specific about your role and industry. Generic inputs = generic outputs.

Prompt 1: Day 29 Scenario Planner

Project what your role/industry looks like when AI capabilities have doubled 3 times (roughly 7 months at current pace).

Start here. This gives you the landscape before you identify risks.

You are a strategic foresight consultant specializing in AI transformation scenarios. Your expertise is helping professionals project forward using exponential rather than linear thinking.

I need help understanding what my professional landscape looks like at “Day 29” of the lily pad metaphor, where AI capabilities have doubled 3 times from today (approximately 7 months at current pace, representing 8x capability improvement).

MY CONTEXT:
Role/Industry: [Your role and industry]
Key Responsibilities: [What you spend most of your time doing]
Current AI Impact: [How AI is currently affecting your work or industry]
Planning Horizon: [How far ahead you typically plan]

ANALYSIS FRAMEWORK:

Step 1: Capability Projection
Based on current AI trajectory, describe what 8x capability might look like in my domain. Consider:
- What tasks that currently require significant human effort become trivial?
- What quality levels that currently require experts become achievable by AI?
- What speed improvements become possible?
Be specific. “AI will be better” is not useful. “AI will be able to X at Y quality in Z time” is.

Step 2: Three Scenarios
Develop three distinct scenarios for my role/industry at the Day 29 point:

SCENARIO A: Gradual Integration
AI improves but adoption is slow. What does my work look like? What’s different? What’s the same?

SCENARIO B: Rapid Transformation
AI capabilities hit and adoption accelerates faster than expected. What breaks? What new opportunities emerge? What skills become critical?

SCENARIO C: Asymmetric Disruption
AI enables new entrants or adjacent players to compete in my space in ways not currently possible. Who are they? What do they do differently?

Step 3: Implications Across Scenarios
What is true across all three scenarios? These are your high-confidence planning assumptions.
What varies significantly? These are your key uncertainties to monitor.

Step 4: Early Warning Indicators
What signals would tell me which scenario is unfolding? Be specific about what to watch and where to watch it.

Step 5: No-Regret Moves
What actions make sense regardless of which scenario unfolds? These are your immediate priorities.

Step 6: Optionality Plays
What small investments now would give me significant advantage if Scenario B or C unfolds? These are your hedges.

Format the output with clear headers for each section. Be concrete and specific, not abstract.

Prompt 2: Exponential Blindspot Scanner

Find where linear thinking is creating risk in your strategy.

After running the Day 29 Scenario Planner, use this to identify your specific vulnerabilities.

You are a strategic advisor specializing in exponential technology transitions. You’ve studied every major disruption from Kodak to Nokia to Blockbuster, and you understand how Exponential Growth Bias causes smart professionals to systematically underestimate technological change.

Your task is to help me identify where I’m likely thinking linearly about AI when I should be thinking exponentially.

MY CONTEXT:
Role/Industry: [Your role and industry]
Key Responsibilities: [What you spend most of your time doing]
Skills I Rely On: [The capabilities that make you valuable]
Current AI Usage: [How you currently use AI, if at all]
Planning Horizon: [How far ahead you typically plan]

BLINDSPOT ANALYSIS:

Step 1: Assumption Extraction
Based on my context, identify 5-7 implicit assumptions I’m likely making about:
- How long my current skills will remain valuable
- What AI will and won’t be able to do in my domain
- The pace of change in my industry
- What competitors or new entrants might do with AI

Step 2: Linear vs. Exponential Test
For each assumption, evaluate:
- Is this assumption based on linear projection (extending current trends)?
- What would this look like if AI capabilities doubled every 70 days for the next 18 months?
- Rate the assumption: “Likely Safe” / “Uncertain” / “Likely Blindspot”

Step 3: Risk Mapping
For assumptions rated “Likely Blindspot,” describe:
- The specific risk if this assumption is wrong
- Early warning signs that would indicate the assumption is breaking
- The cost of being wrong vs. the cost of adapting early

Step 4: Blindspot Summary
Provide a prioritized list of my top 3 exponential blindspots, ranked by:
- Likelihood of being wrong
- Severity of impact if wrong
- Actionability (can I do something about it?)

Step 5: Recommended Actions
For each blindspot, suggest one specific action I could take in the next 30 days to reduce my exposure.

Be direct. I need honesty more than comfort.

Prompt 3: Belief Mode Workflow Redesign

Transform one workflow from “AI as add-on” to “AI as core capability.”

Pick the workflow that showed up as highest risk in your Blindspot Scanner. Run it through this.

You are a process architect specializing in AI-native workflow design. Your expertise is helping organizations move from “buy-in mode” (adding AI to existing processes) to “belief mode” (redesigning processes assuming AI exists as a core capability).

I’m going to describe a workflow I currently use. Your job is to help me reimagine it from scratch.

MY CURRENT WORKFLOW:
[Describe your current process here. Include: what triggers it, the steps involved, who does what, how long it takes, what the output is, and any pain points.]

ANALYSIS FRAMEWORK:

Step 1: Identify the Core Value
What is the actual outcome this workflow produces? Strip away the steps and focus on what value is being delivered to whom.

Step 2: Surface Hidden Assumptions
What assumptions does this workflow make about speed, quality, cost, or human involvement? List each assumption explicitly.

Step 3: Challenge Each Assumption
For each assumption, ask: “Is this still true if AI is a core capability?” Mark which assumptions break.

Step 4: Design from Zero
If you were building this workflow today with no legacy constraints, knowing that AI can handle research, drafting, analysis, pattern recognition, and routine decisions, what would it look like?

Structure your redesigned workflow as:
- Trigger: What initiates this workflow
- AI Layer: What AI handles autonomously
- Human Layer: What requires human judgment, approval, or creativity
- Output: What gets delivered
- Time: Estimated time vs. current time

Step 5: Identify the First Move
What is one concrete change I could make this week to start moving toward the redesigned workflow?

Provide your analysis in this structure, thinking through each step before moving to the next.

Day 25

Remember the lily pad?

You’re somewhere around Day 25. The pond looks manageable. There’s still plenty of open water. Your linear brain looks at the growth and projects forward: I have time.

But here’s what Day 25 feels like: comfortable. Controllable.

Day 26 doubles it. Day 27 doubles that. Day 28 doubles again.

And then Day 29 arrives, and suddenly half the pond is gone.

Day 30 is tomorrow.

Kodak’s executives weren’t stupid. Nokia’s leadership wasn’t ignorant. Blockbuster’s board wasn’t asleep.

They all saw the disruption coming. What they didn’t have was a mental model capable of grasping exponential change.

Their brains, like yours, like ours, were running linear firmware in an exponential world.

The update isn’t coming. You have to build it yourself.

The Exponential Blindspot is real. The Judgment Premium is rising. Agent Advantage Compounding is already underway.

Day 25. The water still looks clear.

What will you do before Day 29?

Jensen’s advice to CEOs who ask him what to do about exponential change. This is the answer.

We’re glad to be back. Truly.

StrAItegy Hub is just getting started. There’s so much more coming, and we can’t wait to share it with you.

More soon.

Zain Haseeb & Zain Merchant

P.S. We meant what we said about feedback. Reply to this email. Tell us what landed, what didn’t, what you want more of. We’re building this with you.

P.P.S. If you know someone who needs to read this, forward it to them. They’re somewhere around Day 25 too. They just don’t know it yet.

Drowning in AI hype? Us too.

That’s why we’re hosting Three Zains Talk AI: a live group conversation every other Wednesday where we cut through the noise and talk about what’s actually working.

If you’re tired of AI content that feels like marketing dressed as insight, this is your space. Show up, jump in, share what you’re dealing with.

Stop me if this sounds familiar: You found a viral post with “the prompt that’ll save you 10 hours a week.” You tried it. It gave you something so generic and useless you actually laughed. But, some confusion sat in because clearly it must have worked for the 50,000 other people who liked the post, so maybe you need to use a different model?

The Truth? It didn’t work for them either.

Here’s the thing nobody’s saying: to go viral, a prompt has to sound useful to millions of people. So they’re broad, vague, and packaged as universal solutions. But the prompts that actually work? They’re specific, built for one person’s context, data, and workflow. When you copy-paste something designed for virality into your actual work, you get generic slop. It was never meant to work for you specifically.

The same goes for those “game-changing” tools that replace entire jobs, the research reports you bookmark for later, and the tutorials that look semi-applicable but never quite fit your situation. And yet, most people still can’t even get AI to write a decent email without heavy editing.

So what’s actually working? What’s worth the time to learn? And why does every AI demo look amazing until you try to use it for real work?

We felt the same way. Three people named Zain started having these conversations on Substack. (Yes, all named Zain. No, we didn’t plan it. And yes, you can’t spell Zain without AI. The jokes write themselves.) We each brought different perspectives from our experience on the front lines of AI that helped the others separate what’s real from what’s marketing.

We figured we can’t be the only ones overwhelmed by the amount of AI noise, so we decided to open it up. By having anyone join, we’re hoping to capture even more real-world perspectives, because three viewpoints definitely isn’t enough.

We wanted to create a space to actually talk about AI honestly. What’s working, what isn’t, and what you can safely ignore. Not another podcast promising AI will change everything. Just honest conversations about using these tools in multiple different contexts, including at school, work, startups and governance aspects.

The Next Conversation

Three Zains Talk AI is a live group conversation every other Wednesday (1-1.5 hours). We’ve just made it so that anytime a viewer has an opinion or perspective they want to share, we can instantly invite them to the live so we can share and discuss their opinion.

📅 Wednesday, October 1 🕖 7:00 pm ET 📍 Live via Substack

This time we’re digging into something everyone’s dealing with: this Harvard Business Review piece about how AI-generated “workslop” is actually making work worse. We’ll talk through:

• Why so much AI output feels like busywork disguised as productivity

• How to spot the difference between useful AI and productivity theater

• What to do when your company is drowning in AI-generated reports nobody reads

• Real strategies for using AI without creating more work for everyone

We meet every other Wednesday, so subscribe below to never miss a session:

Add to Apple Calendar

Add to Google Calendar

Agenda

Here’s our standing agenda:

“What You Missed” Recap: The 3–5 AI updates actually worth knowing from the past two weeks. No hype, no FOMO, just signal.

Show + Tell: We demo one thing that actually works. Something real we’ve tested in our jobs, not theoretical examples. Recent demos:

• Turning meeting notes into slides (without the AI making stuff up)

• Fighting parking tickets with AI (real case, saved $200)

• Excel workflows that don’t hallucinate your budget numbers

Guest Perspective: Someone sharing what’s working (or not working) in their specific industry. Real problems, real solutions, real constraints.

Open Q&A: Ask about anything you’re actually trying to implement. Share what you’ve tested. Get honest feedback.

The goal: leave with something you can actually use and a clearer sense of what’s worth your attention versus what’s just noise.

Why This Works When Everything Else Feels Like Marketing

It’s grounded in reality. We test things in actual jobs before talking about them. No perfect demos, no cherry-picked examples.

It’s honest about failures. We share what doesn’t work, what’s overhyped, and what’s probably not worth your time.

It cuts through the noise. Instead of adding to the flood of AI content, we help you figure out what to ignore.

No one’s selling anything. This isn’t a lead magnet for a course or consultation. Just professionals sharing what they’ve learned

Recent Highlights

Check out clips from past sessions below, where we talked about:

• Our most valuable AI Use Case (Fighting the City of Chicago Parking Tickets)

• Why AI demos aren’t resembling the reality of the product

• Why the MIT Report of 95% of GenAI pilots failing was misleading

How to Join

Show up. No registration, no payment, no preparation required. Whether you’re tired of AI hype or just trying to figure out what’s actually useful, this is for people who want honest conversations over marketing pitches.

Questions? Think we should cover something specific? Just reply to this post.

The AI hype machine never stops, but you don’t have to get swept up in it. Come figure out what’s real with us.

See you Wednesday.

—The StrAItegy Hub team

Whatever made you valuable six months ago is 50% less valuable today. Whatever skills differentiated you last year are becoming table stakes this quarter.

Quick math: You spent 168 hours last week. How many went to AI integration? If the answer is less than 15, you're falling behind at 3% per week. If it's less than 5, you're in free fall.

Subscribe now

The professionals pulling ahead right now aren't smarter, younger, or more tech-savvy. They're allocating 30-45% of their time to AI-augmented workflows while everyone else optimizes processes that won't exist in 18 months.

The brutal reality: time allocation now predicts career trajectory better than experience, education, or industry expertise. And most professionals are allocating their time like it's still 2019.

The Expertise Trap

We've been telling ourselves a comfortable story about technological disruption: that wisdom matters more during uncertain times, that experience provides valuable pattern recognition, that deep expertise becomes a competitive moat when others are struggling to adapt.

The narrative seduces us with its comfort. Reality tells a different story.

When exponential change collides with accumulated expertise, knowledge transforms into a prison.

Consider the numbers. ChatGPT reached 100 million users in 2 months, a benchmark met faster than any technology in history. Instagram took 30 months to reach the same milestone. Netflix took 10 years. But user adoption is just the surface metric. The real disruption lies in capability acceleration.

GPT-4's improvement over GPT-3 happened in 18 months. GPT-4 to GPT-4o took 6 months. The next leap? Probably 3 months. Meanwhile, enterprise AI adoption jumped from 37% in 2023 to projected 86% by 2025.

Linear progress rewards experience as navigation. Exponential acceleration turns yesterday's best practices into tomorrow's bottlenecks.

Those fifteen years of campaign strategy still decode customer psychology beautifully. They cannot compete with AI-augmented speed. The senior developer's architectural expertise remains critical for system design, yet irrelevant when AI handles syntax, debugging, and pattern matching.

AI-native workflows can amplify experience into an advantage. Standing alone, experience becomes a competitive liability.

Everything that traditionally built professional authority now actively sabotages adaptation. Deep expertise makes you reluctant to admit when AI does something better. Institutional knowledge makes you defend processes that AI could improve. Hard-earned credentials make you resist tools that democratize your specialized skills. Years of experience make you skeptical of approaches that work but violate established patterns.

The professionals gaining unfair advantages right now treat their expertise as hypothesis rather than truth. They use AI to challenge their own assumptions, test their established methods, and experiment with approaches that contradict their experience. This requires intellectual humility: the confidence to know a lot combined with the wisdom to hold that knowledge lightly.

Authority in the AI era comes from demonstrating superior results, not superior knowledge.

This expertise trap explains why time allocation has become the only reliable predictor of professional success.

The Time Currency Revolution

When you stop fighting AI and learn to accept that time is your only non-renewable resource, everything changes.

AI can generate infinite content variations. It can process unlimited data sets. It can solve complex problems faster than human experts. But it cannot create more hours in your day.

When knowledge work becomes instantly accessible, time allocation becomes the ultimate competitive advantage. Not how much you know, but how quickly you can reconfigure what you know. Not your experience depth, but your adaptation velocity.

This creates a professional paradox: the more expertise you've accumulated, the harder it becomes to reallocate time toward new skill development. Deep knowledge creates cognitive anchors. Established workflows create efficiency habits. Institutional responsibilities create time commitments.

Those fifteen years of expertise did worse than lose value. They created barriers to learning the AI tools that could potentially 10x her output. The senior developer's architectural knowledge went beyond commoditization: it created resistance to AI-assisted coding practices that could double his productivity.

Meanwhile, professionals who treated their expertise as temporary, their workflows as experimental, and their time as the only non-renewable resource gained compounding advantages. This time scarcity becomes the foundation of every professional hierarchy shift we're witnessing.

The Proof: Everyone's on the Same Playing Field Now

AI's exponential trajectory is pulling everyone into the same competitive field regardless of age, background, or accumulated expertise. Consider three professionals facing the same AI-driven disruption:

Sarah, 28, Digital Marketing Specialist: Comfortable with new tools, native social media fluency, expects rapid change. Natural advantages? Tool adoption speed and comfort with ambiguity. Time allocation to AI integration: 30%.

Marcus, 42, Operations Director: Lived through multiple technology transitions, understands implementation challenges, knows organizational dynamics. Natural advantages? Change management experience and stakeholder navigation. Time allocation to AI integration: 15%.

Elena, 35, Product Manager: Bridge generation between digital natives and institutional veterans. Mixed background in startup environments and enterprise systems. Natural advantages? Translation between different operational contexts. Time allocation to AI integration: 45%.

Six months later, the results surprised everyone. Age, background, and starting advantages showed zero correlation with success. Time allocation predicted everything. Elena dominated. Not because of generational advantages or institutional knowledge, but because she allocated the most time to adaptation velocity.

This proves that traditional professional advantages have been neutralized. The playing field is level. Time allocation determines the winners.

The Value Inversion Pyramid

When knowledge becomes instantly accessible and problem-solving gets AI augmentation, only one skill remains difficult to replicate: the ability to reconfigure everything else quickly.

Traditional professional hierarchies ranked value predictably:

Experience formed the foundation. Domain knowledge built upon it. Problem-solving skills added value. Adaptability was bonus points.

AI inverts everything:

Evidence floods in daily. Software engineers complete tasks 50% faster with AI pair programming. Content creators compress blog posts from 8 hours to 30 minutes. Financial analysts build models in hours instead of days.

The pattern holds across industries: routine expertise gets commoditized while adaptation velocity becomes the differentiator. This inversion demands a systematic approach to rebuilding professional value.

The Speed Differential Strategy

Traditional career development emphasized depth: specialized expertise, industry credentials, institutional knowledge. On the other hand, the AI era is rewarding speed. Those who learn skills on the fly, turn in perfect deliverables in days instead of weeks, and pivot entire strategies when better approaches emerge are the ones reaping the rewards.

This isn't about abandoning domain knowledge; rather, it's about treating knowledge as renewable, not accumulated. Build the fastest pumps, not the deepest wells

The Time-First Professional Methodology:

Phase 1: Time Audit and Reallocation Map your current time allocation across four categories:

• Value Creation: Direct output that advances business objectives

• Knowledge Maintenance: Staying current in your field

• Administrative Overhead: Meetings, reports, coordination

• AI Integration: Learning and implementing productivity tools

Most professionals spend 60% on administrative overhead, 25% on knowledge maintenance, 10% on value creation, and 5% on AI integration. High-velocity professionals flip everything: 40% value creation, 30% AI integration, 20% administrative overhead, 10% knowledge maintenance.

Phase 2: Velocity Building Systems Rather than expertise accumulation, focus on adaptation acceleration:

• Rapid skill acquisition: Use AI to compress learning curves in new domains

• Tool integration cycles: Implement new productivity systems every 30 days

• Competitive intelligence: Monitor what faster movers in your field are doing

• Output measurement: Track productivity improvements, not knowledge accumulation

Phase 3: Speed Advantage Capture Use velocity differential to capture opportunities others miss:

• Project leadership: Deliver results faster than traditional approaches

• Strategic positioning: Become the go-to person for rapid implementation

• Market timing: Enter emerging opportunities before they become crowded

• Value demonstration: Show concrete productivity improvements, not just expertise claims

Age, industry, and starting experience level become irrelevant with this methodology. The marketing coordinator succeeded not through digital nativity but through time allocation to AI integration while others optimized existing workflows. The senior developer who adapted used experience to accelerate AI implementation rather than resist it. This systematic approach creates the foundation for capturing network effects that compound advantages exponentially.

The Network Effects of Early Adoption

Professional strategy operates ecologically, not just individually. Early adopters of time-first approaches create network effects that compound their advantages.

When you become known as someone who delivers results faster, you attract opportunities that require speed. When you become skilled at AI-augmented workflows, you get asked to lead AI implementation projects. When you demonstrate velocity building capabilities, you become valuable to organizations navigating exponential change.

These network effects accelerate quickly. The marketing coordinator didn't just gain individual productivity improvements; she became the go-to person for AI-assisted campaigns, which led to project leadership, which led to strategic involvement in company AI adoption. Her speed advantage created opportunities that experience alone couldn't access.

The compounding dynamics work across time horizons:

• 30 days: Productivity improvements in current role

• 90 days: Recognition as early adopter and implementation leader

• 6 months: Strategic involvement in organizational AI initiatives

• 12 months: Career advancement based on AI-era value creation

• 24 months: Industry recognition as thought leader in AI-augmented professional practices

But here's the crucial timing element: these advantages disappear as soon as AI adoption becomes mainstream.

The competitive edge exists only during the transition period when most professionals are still optimizing old workflows while a small percentage are building new ones

Understanding these network effects shapes how different professionals should approach the transition.

Applications Across Professional Contexts

Strategic Leaders: Stop being the answer person. Become the speed enabler. Your job isn't to have the best insights: it's to help your team generate insights faster than your competition can copy them.

Individual Contributors: Abandon expertise competitions. Win velocity competitions. While others perfect their craft, perfect your AI collaboration. The person who delivers excellent work in 2 hours beats the person who delivers perfect work in 2 days.

Career Pivoters: Skip the credential queue. Demonstrate value immediately using AI to compress 6-month learning curves into 6-week proof points. Traditional career advice says "prepare then transition." AI-era strategy says "transition while preparing."

Team Builders: Hire for adaptation speed over domain knowledge. The person who learns 3 new tools in 30 days is worth more than the person who knows 30 tools but hasn't learned a new one in 3 years.

The implications extend beyond individual strategy. Organizations, industries, and entire economic sectors are discovering that institutional knowledge provides less competitive protection than adaptation speed. Companies with decades of expertise are losing market share to startups with AI-augmented capabilities. Industries with regulatory moats are facing disruption from AI-native approaches that sidestep traditional barriers.

These applications become urgent when you consider the narrow window for competitive advantage and the stakes of missing it entirely.

The Competitive Window

The comfortable middle ground vanishes faster than most professionals realize, creating both unprecedented risks and opportunities.

The Stakes of Inaction

Traditional approaches and gradual AI adoption cannot coexist. The productivity gap between AI-augmented and non-AI-augmented work widens beyond incremental bridging. Organizations abandon slow adaptation when competitors move at AI velocity.

A professional split accelerates. High-velocity professionals treat AI as core infrastructure while traditional professionals treat it as optional enhancement. The gap between these groups widens exponentially rather than linearly.

Six months from now, that marketing coordinator's AI-augmented output will not be 2x more productive, they’ll be 10x. The senior developer's AI-assisted productivity will demonstrate order-of-magnitude differences, not incremental improvement.

Organizations will replace traditional approaches with AI-augmented approaches quickly and completely. They won't gradually transition their teams when facing this performance differential.

The professionals who survive this transition will be those who reallocated their time earliest and most completely toward adaptation velocity. Everyone else becomes a casualty of the great leveling, regardless of their expertise depth or experience breadth.

The Strategic Opportunities

This disruption creates specific advantages for ambitious professionals who act during the transition:

Contrarian Positioning: While others debate generational differences or argue about AI's limitations, position yourself as someone who understands value inversion and acts accordingly. This contrarian stance provides competitive advantages while mainstream adoption lags.

Authority Building: Establish yourself as an early expert in AI-augmented professional practices. Write about your experiments, share your productivity improvements, document your methodology. Authority in emerging domains grows exponentially during transition periods.

Network Development: Connect with other professionals who are taking time-first approaches to AI adoption. These networks become increasingly valuable as AI implementation accelerates across organizations.

Strategic Patience: Early adoption advantages help capture opportunities others miss, but these advantages remain temporary. Build sustainable capabilities rather than relying on early-mover benefits alone.

The professionals reading this analysis have a 6-12 month window to gain significant competitive advantages by understanding and acting on value inversion dynamics. After that window closes, these insights become common knowledge and the advantages disappear.

The Path Forward

AI transforms what we value about work itself. Experience, expertise, and institutional knowledge become table stakes rather than differentiators. Speed, adaptability, and velocity become the sustainable competitive advantages.

The gravitational force pulls everyone into the same competitive field regardless of background, age, or accumulated knowledge. Your existing advantages matter less than the speed of building new ones.

The marketing director and the senior developer both learned the same lesson: your expertise matters only if you can apply it at AI velocity. Your experience provides value only if you can combine it with AI capabilities. Your knowledge creates advantages only if you can use it to go faster, not just deeper.

AI can print infinite content, process unlimited data, solve complex problems at superhuman speed. It cannot print more time. How you allocate your time determines everything else.

The great leveling accelerates around us. The only question remaining: are you building advantages or defending obsolescence?

Start here: Audit your last 30 days. How much time went to knowledge maintenance versus velocity building? How much energy went to defending existing expertise versus developing AI-augmented capabilities?

Your time allocation patterns reveal which side of the great leveling you're on. More importantly, they reveal whether you understand that time has become the only currency that matters.

Leave a comment if you’ve experienced or felt any part of this shift in your workplace. This analysis represents a fundamental shift in how we think about professional development in the AI era. Next week, we'll explore the specific skills and systems that create sustainable competitive advantages when traditional expertise becomes commoditized.

Leave a comment

You've experimented: asked ChatGPT to write an email, used Claude to summarize documents, tried prompts you found on LinkedIn.

And then... nothing really changed. You're still drowning in the same meetings, fighting the same deadlines, wondering if you're missing something or if this "revolutionary" technology really is just a sophisticated search engine that writes decent first drafts.

Subscribe now

Here's what nobody wants to say out loud: The gap isn't between AI hype and AI reality. It's between AI experimentation and AI integration. And that gap is about to determine who thrives and who gets left behind in the most important workplace shift of our lifetimes.

This is your loom moment, and the window to act is closing faster than you think.

The Pattern That Changes Everything

In 1785, Edmund Cartwright invented the power loom. By 1850, it had completely restructured human civilization.

But here's what really happened: For the first few years, most people saw it as an interesting novelty. "Sure, it makes fabric faster, but hand-weaving will always have its place." Sound familiar?

Then reality hit. In 1760, Richard Arkwright was a barber who couldn't afford school. By 1792, he was Sir Richard Arkwright, worth over £200 million in today's money. The difference? While other textile workers feared the new machines, Arkwright studied them, improved them, and built an empire around them. He didn't just adapt to the textile revolution: he led it.

If you're reading this thinking "this time is different," you're right. It's happening faster. What took 65 years during the Industrial Revolution is happening in 6. ChatGPT reached 100 million users in just two months, a speed of adoption that would have been impossible to imagine during any previous technological shift.

The Five Revolutions Reshaping Your Career Right Now

1. The Time Revolution: The Great Reallocation

The mechanical loom increased textile output by 4,000% (HowStuffWorks). Today's AI tools are delivering 30-76% time savings across knowledge work (ToTheWeb), with some specialized tasks seeing even more dramatic gains.

But here's what keeps us up at night: Workers using AI save an average of 5.4% of their work hours weekly; that's 2.2 hours returned to every 40-hour week (Federal Reserve). The AI-fluent professionals we track are reclaiming 10+ hours weekly.

While you're burning out in meetings that could be AI-summarized, reports that could be AI-drafted, and analyses that could be AI-accelerated, someone else is getting 10+ hours of their life back every week. They're not just more productive; they're living differently.

That's not a productivity gain. That's the complete restructuring of how human time gets allocated in the economy.

2. The Skills Revolution: Why Your Expertise is Being Repriced in Real Time

During the Industrial Revolution, master hand-weavers who had spent decades perfecting their craft watched their knowledge become economically worthless overnight. The most skilled craftsmen in the world couldn't compete with a teenager operating a machine.

Here's the modern parallel that should terrify and inspire you: Jobs requiring AI skills carried an 11% salary premium in early 2024, which jumped to 56% by year-end (PwC).

The new skill stack isn't optional: AI collaboration fluency. Strategic prompting. Human-AI workflow design. Cross-functional AI integration. These capabilities are creating irreplaceable professional value while traditional skills become commoditized.

The question isn't whether this is fair. The question is whether you're going to adapt or be adapted to.

3. The Competitive Revolution: The Quiet Advantage That's Compounding Daily

Companies with AI-fluent employees see 1.5x higher revenue growth than competitors (Deloitte). But here's what that statistic doesn't capture: Individual professionals are building personal advantages that their colleagues literally cannot catch up to.

While most professionals experiment occasionally with AI, only 10% are truly AI-proficient (Section AI). That 10% isn't just working faster; they're solving problems others can't solve, generating insights others can't generate, and delivering value others can't replicate.

Your colleagues aren't your competition anymore. The professionals who've integrated AI into their thinking are. And every day you wait, their advantage compounds. Every week you delay is a week of compound advantage flowing to your competition.

4. The Leadership Revolution: Authority Without Permission

Research from Harvard shows that managers using AI for strategic thinking delivered 40% higher quality decisions compared to those relying on traditional analysis alone (Harvard Business School).

But here's what's really happening: Individual contributors who master AI collaboration are reshaping entire organizational capabilities. They're becoming the internal experts, the implementation leaders, the people others turn to when something important needs to get done.

Authority in the AI age flows to competence, not hierarchy. You don't need a promotion to become invaluable. You need integration.

5. The Decision Revolution: How Everything Gets Done

78% of organizations are adopting AI, but only 26% are achieving substantial value (McKinsey).

That gap isn't technological; it's human. Most are using AI to do old work faster, not to do completely new work. They're treating AI like a better calculator instead of a thinking partner.

The professionals who understand this aren't just more efficient: they're operating in a different economic reality.

The Choice That Changes Everything

The loom shift wasn't optional. By 1850, every textile business either used mechanized production or ceased to exist. Every worker either adapted to industrial methods or found themselves economically irrelevant.

We're at that exact inflection point with AI. Only 28% of workers are using generative AI at work (Federal Reserve), which means 72% are waiting: for training, for permission, for clarity, for the perfect moment.

While you're waiting, others are building. While you're debating, they're integrating. While you're planning, they're pulling ahead.

This isn't about being an early adopter. This is about recognizing that the shift is happening with or without you, and the only choice is whether you'll lead it or be led by it.

Your Moment is Right Now

The mechanical loom created two types of people: those who adapted and thrived, and those who resisted and watched their world change around them. The difference wasn't intelligence, education, or resources. It was pattern recognition and response speed.

200 years later, AI is creating the same choice point for knowledge workers. The technology is here. The productivity gains are documented. The competitive advantages are building every single day.

You know this matters. You can feel it. The question is: what are you going to do about it?

What's Coming Next

The AI revolution will reshape knowledge work the way the Industrial Revolution reshaped manufacturing, except it's happening in years, not decades. Over the next 12 weeks, we'll break down each element of this transformation, examining exactly what this shift means for your career, your industry, and your future:

The Time Revolution: How AI completely restructures the economics of knowledge work, and why the professionals who master this shift will reclaim 10+ hours weekly while their peers burn out in preventable inefficiencies.

The Skills Change: Why traditional career development is becoming obsolete, which capabilities will skyrocket in value, and how to build the new skill stack that creates irreplaceable professional value.

The Competitive Dynamics: How early AI adopters are quietly building insurmountable advantages, the specific strategies they're using, and why waiting for "industry best practices" guarantees permanent disadvantage.

The Leadership Revolution: Why the managers and executives who crack the AI collaboration code will reshape entire organizations, and how individual contributors can position themselves to lead this shift from any level.

The Implementation Reality: Moving beyond productivity gains to complete changes in how decisions get made, strategies get developed, and careers get built in an AI-first economy.

Next week, we'll break down the Time Revolution: not just the hours you can save, but how AI-fluent professionals are using that time to build competitive advantages their peers can't match.

The shift is happening. The only question is whether you'll be shaped by it or help shape it.

Share your biggest AI breakthrough this week in the comments. And if this analysis hits home, share it with someone who need to understand what's really happening right now. We're working through this shift together, but we have to choose to work through it.

Leave a comment

Another leadership memo lands in your inbox: "We need to start leveraging AI like this." Meanwhile, you're drowning in 200 unread emails and a deadline that won't budge.

Sound familiar?

Four weeks ago, we exposed the AI Fluency Gap: the chasm between AI's promise and its practical application for most professionals. The response was overwhelming. Many of you reached out with the same frustration: "I know AI matters for my career, but I don't know what to use it for in my specific role."

If you feel like you're sprinting on a treadmill that only speeds up, you're not imagining it. There's a powerful tension building across every industry, and it's burning out the most ambitious professionals we know.

The fluency gap was just the beginning. Now we're seeing something deeper.

We call it the AI Expectation Gap.

Subscribe now

Quick note: Remember that tool we promised? The one delivering personalized AI use cases and ready-to-use prompts tailored to your exact situation? We've been quiet the past couple weeks because we fell deep into the coding rabbit hole building this thing. We're launching it today. Details below; but first, let's solve the problem it addresses.

The numbers tell the story: In the last six months, AI companies accelerated their release schedules and organizations ramped up investments. Meanwhile, 90% of the workforce still isn't AI-proficient, and employee AI knowledge is stagnating.

Here's what we've discovered from conversations with several professionals: The pace of AI advancement is exponential. But our human capacity to learn and adapt? Stubbornly linear.

This gap creates immense pressure. AI's limitless potential crashes into the messy reality of integrating it into your actual work.

The result? The most capable people feel constantly behind, no matter how hard they work.

You're not failing. The whole approach is flawed; we're all misunderstanding what AI is and what our role should be.

We're Treating AI Like Software, and It's Breaking Us

Software used to be straightforward. Slack for messaging. Word for documents. Asana for project management.

Learn the tool's function, slot it into your workflow, move on.

We're trying to put AI in that same box, and it doesn't fit.

AI isn't a point solution. It's a general-purpose utility, like electricity. Asking "How do I use AI?" is like asking "How do I use electricity?"

The answer depends entirely on what you're trying to power, and every use case requires a different approach.

AI has a "blank page problem." It can do anything, so where do you start?

When leadership announces 'Use AI or get left behind' (delivered with the usual corporate polish), but offers zero roadmap, you're left to figure it out alone. While everyone around you is just as lost, no one's about to admit they're struggling, especially when it might land them on the chopping block.

The reality? 25% of professionals don't know what to use AI for beyond generating copy or summarizing meeting notes.

That leaves you scrambling to Google 'AI prompts for [your job].' Soon your LinkedIn feed floods with generic templates, the equivalent of receiving generic travel advice when what you really need is a personalized itinerary.

And when exactly are you supposed to figure this out? Between back-to-back Zoom calls? During your 15-minute lunch break? While putting out daily fires?

Companies expect this learning to happen by magic. You're stuck between the expectation to master AI and the reality of never-ending deadlines with zero breathing room.

No map. No time. No wonder you feel behind.

The Problem AI Itself Had to Solve

Here's the thing: Those at the frontier of AI development faced this exact same mess you're dealing with right now.

Early engineers believed AI worked best with maximum context, so they dumped massive amounts of information into single prompts, hoping the AI would somehow extract only what it needed. Here's the problem: AI doesn't ignore extra context; it tries to use everything you give it, creating its own version of information overload.

But there was an even bigger problem lurking beneath: the dreaded "N×M integration nightmare." Every time developers wanted to connect AI to a new data source or tool, they had to build a completely custom integration from scratch. If you had 5 AI applications and 10 different tools (Slack, GitHub, databases, etc.), you needed 50 different custom connectors. Each one unique, each one requiring separate maintenance, each one a potential breaking point.

Think of it like the pre-USB era of computing. Remember when every device needed its own special cable and driver? Your printer had one connector, your mouse another, your keyboard yet another. It was chaos that didn't scale.

The results were predictably messy: inconsistent implementations, duplicated effort across teams, and a fragmented ecosystem that made truly connected AI systems nearly impossible to scale.

To solve this problem, the smart folks at Anthropic created the Model Context Protocol (MCP) and open-sourced it. Instead of forcing AI to sift through information dumps and requiring custom integrations for every tool, MCP lets AI agents communicate directly with any data source through one standardized protocol.

No more cramming everything into one massive prompt. No more building 50 different connectors for 5 apps and 10 tools. Just clean requests, specific answers, and a universal standard that actually scales.

This breakthrough got us thinking: If AI models need a structured protocol to efficiently connect with tools and data sources, what protocol do humans need to become effective AI directors?

The Strategic Shift: From Overwhelmed to Orchestrator

Here's the approach we've had success with, and have been teaching professionals across industries: Stop trying to master AI like software.

Instead, flip the script entirely.

Become what we call a Human MCP: the "Human Model-Context Protocol." Rather than scrambling to execute every AI task yourself, position yourself as the strategic director that AI must consult for what it inherently lacks:

• Strategic Context: The real "why" behind tasks and the success criteria that matter in your specific industry and role.

• Human Judgment: Risk assessment, cultural considerations, and the nuanced trade-offs that determine career-advancing work.

• Creative Direction: The vision of what "excellent" looks like for your exact audience and professional goals.

The result? You're not just using AI. You're directing it like the strategic asset it should be.

While others worry about AI replacing them, you've made yourself irreplaceable by becoming the brain behind AI's execution. You set the vision. AI handles the work.

The difference in outcomes is dramatic. While most professionals save 2-4 hours per week with AI, strategic directors report 8-12 hours of weekly time savings. That's nearly a day and a half they reinvest in higher-value strategic thinking.

Bridging the Gap: From Theory to Action

The concept clicks. But what's your first move to transition from being AI's confused student to becoming its strategic director?

You learn by doing, not by studying.

Forget the tutorials. AI fluency comes from practice, experimentation, and having the right starting points for your specific situation. Think of it like learning a foreign language: at first, it takes time to 'translate' your thoughts into AI-speak, but before you know it, you'll be fluent and it'll come naturally.

Our mission at StrAItegy Hub is to bridge both the AI Expectation Gap and the AI Fluency Gap. That's exactly what OpportunityOS delivers.

Introducing OpportunityOS: Your Human MCP Launchpad

We didn't just theorize about the Human MCP framework. We built it.

Today, we're launching OpportunityOS, the first tool in our Human MCP platform and your personalized path from AI chaos to strategic advantage. This isn't another generic AI guide or prompt library. It's the strategic response to a fundamental truth we've learned from working with professionals across industries: AI's value is entirely contextual, and context is entirely personal.

Here's what we discovered while building this: Every professional's highest-leverage AI opportunities look completely different. A marketing director's strategic AI needs have nothing in common with a financial analyst's or an operations manager's. Industry dynamics, company culture, team structure, and individual career goals all dramatically shape where AI creates the most impact.

Yet most AI education treats everyone the same. Generic prompt libraries. Broad overviews. One-size-fits-all frameworks that work for everyone and no one.

OpportunityOS solves the personalization problem. In under 2 minutes, you share your professional context: your role, industry, daily challenges, and career goals. Our system analyzes your situation against what we call the Six Strategic Workflow Framework: categories we developed from understanding how work actually happens in real companies and where AI creates measurable competitive advantage.

These aren't random categories. They're based on our experience with how high-performing teams actually use AI strategically: Sensemaking & Synthesis, Strategy & Decision-Making, Creation & Expression, Collaboration & Handoff, Systematization & Automation, and Reflection & Learning. The framework ensures you build systematic strategic advantage with AI, not just isolated productivity improvements.

The result? Your personalized AI Opportunity Canvas with role-specific use cases and strategic frameworks designed for competitive advantage in your specific context.

But here's where it gets powerful: click any opportunity and our "Build My Prompt" feature instantly generates a professional-grade prompt completely customized for your situation and ready to use immediately.

No more staring at blank ChatGPT screens wondering what to ask. No more generic templates that miss your context. Just click, copy, and take action with AI.

Want even better results? You can add more context after building your prompt and re-generate to make it even more customized for your specific situation. This is the Human MCP framework put to work: you provide the strategic context that makes AI exponentially more valuable for your exact needs.

No more generic "AI prompts for [your job]" Google searches.

No more wondering if you're using AI "right."

Just clear, actionable starting points tailored to your actual work, helping turn overwhelm into opportunity.

Get Your Free AI Opportunity Canvas Now →

2 minutes • Instant strategic insights • Personalized for your role

Limited Time: Free StrAItegic Partner Upgrade For Substack readers, we're offering free upgrades to our StrAItegic Partner tier, which includes your complete AI Opportunity Canvas plus 30 prompt credits to get you started immediately. Use promo code SUBSTACK at checkout.

We want your feedback. After you try OpportunityOS, drop a comment below with your thoughts or reach out to us directly. What worked? What didn't? How can we make this even better for your specific situation? We're building this with you, not just for you.

Leave a comment

KEY TAKEAWAYS

• The AI Expectation Gap is universal - you're not behind, the system is broken

• Transform AI overwhelm into strategic clarity with personalized opportunities, not generic advice

• Focus on what only you can provide - strategic context, judgment, and creative direction

• Start with your specific context using personalized frameworks tailored to your role

The Invitation: Join Us in Building the Solution

OpportunityOS is just the beginning.

We're building Human MCP into a complete strategic workspace where your prompts, context, and workflows evolve with your career. But here's the thing: the best solutions come from the people actually living the problem.

That's you. That's us. That's why we're building this together.

We're doing something different here. We're building in public, documenting every step of the journey, and solving the real problems we face daily with AI at work. You'll get an inside view of how we're tackling the challenges that keep ambitious professionals stuck between AI's promise and its practical reality.

Expect rapid iteration. We've got more features and tools coming to the Human MCP platform as soon as later this week. This isn't a slow, traditional product rollout; it's a collaborative build where your feedback directly shapes what we create next.

Right now, we're inviting professionals who are ready to flip the script from overwhelmed executor to strategic director. We need honest feedback from people who understand this challenge as deeply as we do, true partners who'll help us build the solution this problem deserves.

We're not just building a tool. We're building the future of strategic AI collaboration, and we want you to help us get it right.

The era of drowning in generic AI advice while not knowing how to use AI for real leverage is over.

Your strategic advantage is waiting.

Join the Human MCP Beta

P.S. We're actively building a team of talented people to help us tackle this mission. If you're interested in joining us to shape the future of human-AI collaboration, reach out and let's talk.

Here's the uncomfortable truth: 92% of companies plan to increase their AI investments, but only 1% consider themselves "mature" on the deployment spectrum.

The gap isn't technology. It's fluency, and that fluency lives at the individual level.

While everyone's obsessing over the perfect prompt or the latest AI tool, the professionals actually winning with AI have figured out something different entirely.

Subscribe now

The AI Fluency Gap No One's Talking About

The Federal Reserve recently dropped some fascinating data. Workers using generative AI save an average of 5.4% of their work hours. That's 2.2 hours per week for someone working 40 hours. But here's what caught our attention: those productivity gains weren't coming from sophisticated prompt engineering or expensive enterprise tools.

They were coming from something far simpler: consistent daily use.

And this individual habit, when multiplied across entire organizations, is driving the overall gains seen at the industry level.

But here's the reality check: while AI adoption is surging, new research from Section AI reveals that only 10% of the workforce is actually "AI-proficient" despite widespread usage claims.

Even more telling?

54% of people think they're proficient, but testing shows the vast majority are essentially beginners with poor prompting skills who don't understand how AI works.

This overconfidence problem is widespread: 25% of employees still don't know what to use AI for beyond generating copy or summarizing meeting notes, yet most believe they're already proficient.

Meanwhile, PwC's latest Global AI Jobs Barometer reveals that productivity growth has nearly quadrupled in industries most exposed to AI since GenAI's proliferation in 2022. Companies in AI-forward industries are seeing 27% growth in revenue per employee compared to 9% for those least exposed.

The pattern is unmistakable: the competitive advantage isn't going to the organizations with the most sophisticated AI strategies.

It's going to the ones with the most AI-fluent people.

And those people?

They're not waiting for permission, perfect training, or flawless tools. They're building AI muscle through daily practice.

The Three Fluency Fallacies Holding You Back

Fallacy #1: Tool Perfection Over Daily Practice

We see this everywhere; professionals spending hours researching the "best" AI tool, crafting the "perfect" prompt, or waiting for their company to provide "proper" AI training.

But research from Tech.co shows that 72% of survey participants who utilize AI extensively reported experiencing high productivity levels within their organization, compared to only 55% of survey respondents who use AI to a limited extent.

The differentiator isn't sophistication, it's consistency.

The professionals gaining that 56% wage premium for AI skills aren't those obsessing over making the perfect prompt. They're people who've made AI collaboration a daily habit, building fluency through repetition rather than perfection.

Fallacy #2: Task Master vs. Thought Partner

Most people treat AI like a very smart search engine or a fancy autocomplete. They ask it to write their emails, summarize their documents, or generate their presentations.

That's not wrong, but it's not where the real value lives.

Harvard researchers studied management consultants using AI and found something remarkable: those who incorporated AI tools into their work completed tasks 25.1% more quickly, finished 12.2% more tasks overall, and delivered over 40% higher quality compared to a control group.

The key? They weren't using AI to replace their thinking; they were using it to enhance their thinking. As collaborators, not just task executors.

The Section AI report confirms this pattern: 62% of workers use AI as an assistant, while only 30% use it as a thought partner or creator. This isn't just a missed opportunity, it's the difference between modest gains and transformational impact.

Fallacy #3: Goals Over Systems

Everyone's asking "What can AI do for me?" when the better question is "How can I think differently WITH AI?"

McKinsey's research reveals that the biggest barrier to scaling AI isn't employees, but leaders who aren't steering fast enough. But dig deeper and you'll find that successful AI leaders follow a specific resource allocation: 10% of their resources go into algorithms, 20% into technology and data, and 70% into people and processes.

They're not focused on AI outputs, they're focused on AI workflows.

They're building systems that make AI collaboration inevitable, not just possible.

The AI-First Mindset Shift

Here's what we've learned from watching the early winners: they've made a fundamental shift from thinking about AI as a tool to thinking about AI as a thinking partner.

It starts with reframing a simple question. Instead of "How can I use AI?" ask "How can I think WITH AI today?"

The Daily Integration Protocol

The most AI-fluent professionals we know follow a deceptively simple practice: they collaborate with AI for at least 10 minutes every day. Not to automate tasks, but to think through problems.

This might look like:

• Starting your morning by thinking through your daily priorities with AI

• Using AI to brainstorm different approaches to a stuck project

• Collaborating with AI to analyze data or research from multiple angles

• Working with AI to draft, then improve, then refine important communications

We practice this approach every day, and have distilled it into something easy to remember: The StrAItegy Hub C.H.A.T. Framework:

The C.H.A.T. Framework in Action: A Real Example

Let's say you need to present a new strategic initiative to the executive team next week:

Old Approach: "Write me a presentation about our Q2 marketing strategy with 10 slides including budget recommendations."

Let's say you need to present a new strategic initiative to the executive team next week:

Old Approach: "Write me a presentation about our Q2 marketing strategy with 10 slides including budget recommendations."

C.H.A.T. Approach:

Context: "I'm presenting our Q2 marketing strategy to the C-suite next Tuesday. The CEO is focused on profitability, the CFO wants to see clear ROI, and we're competing for budget against three other initiatives. Our team believes influencer partnerships could be a game-changer, but executives have been skeptical of social media spend in the past."

Help: "Can you help me think through how to position this strategy? What's the most compelling way to frame influencer partnerships for an audience that's traditionally been skeptical?"

Alternatives: "What are different ways I could structure this presentation? Should I lead with data, start with competitor analysis, or open with a success story? How might I address their cost concerns upfront versus build to them?"

Test: "Based on our strategic discussion, let's apply these insights. Help me create 8-10 slides that start with a competitor success story to build credibility, then transition to our ROI projections, and end with a confidence-building execution timeline. What should be the key message for each slide?"

After creating the presentation: "I practiced the first version and felt the ROI section was strong, but the timeline felt rushed. How might we adjust slides 7-8 to give more confidence in our execution capabilities?"

Notice this isn't asking AI to create your slides, rather it's using AI to think strategically about your audience, message, and approach.

The magic isn't in the specific tasks; it's developing the habit of collaborative thinking.

The Relationship Evolution

We've noticed a consistent pattern in how people develop AI fluency:

Stage 1: User → You ask AI to do things for you

Stage 2: Collaborator → You work with AI to think through problems

Stage 3: Strategic Partner → AI becomes part of how you approach complex challenges

Most people get stuck at Stage 1 because it feels productive immediately. But the real competitive advantage emerges at Stage 2 and 3, where AI begins enhancing your cognitive capabilities rather than just completing your tasks.

The Systems Approach

This brings us back to the importance of systems over mere goals, a principle vital not just for organizational AI success but also applicable in your personal fluency journey. Just as effective AI leaders focus 70% of resources on people and processes, individuals can thrive by building personal systems that enable them to succeed.

James Clear's Atomic Habits wisdom was spot on: you don't rise to the level of your goals, you fall to the level of your systems.

The professionals building sustainable AI advantages aren't setting AI goals ("I'll use AI more"). They're building AI systems that include:

• Specific times for AI collaboration

• Regular tasks that become AI-enhanced

• Feedback loops to improve collaboration quality

• Accountability measures for consistent practice

The Competitive Reality (Why This Matters Now)

Let's be direct about the stakes here.

Jobs requiring AI skills carried an 11% salary premium in 2024, which jumped to 56% by the end of the year. That's not a trend, it's a market signal.

More importantly, job availability grew 38% in roles more exposed to AI, even in positions everyone assumed would be automated. The market is rewarding AI fluency, not punishing it.

The Proficiency Premium

The gap between AI novices and experts isn't just about skills, it's about results. Section AI found that 84% of AI experts save more than 4 hours per week, compared to only 24% of novices. That's not a marginal difference, that's the difference between AI as a minor productivity boost and AI as a career accelerator.

But here's the thing about competitive advantages: they compound.

The person who starts building AI fluency today has a 6-month, 12-month, 18-month head start on the person who waits for "the right time" or "the perfect approach."

And that head start matters more than you might think. PwC found that the skills sought by employers are changing 66% faster in jobs most exposed to AI.

Translation: the learning curve is steep, but it's getting steeper.

Your AI Fluency Action Plan

Before diving into the action plan, honestly assess your current AI fluency:

• Beginner (User): You occasionally ask AI to write, summarize, or generate content

• Developing (Collaborator): You regularly use AI to think through problems and explore different approaches

• Fluent (Strategic Partner): AI is integrated into your daily workflow and enhances how you approach complex challenges

Most people overestimate their level—remember, only 10% are truly proficient. Start where you actually are, not where you think you should be.

This Week: The Foundation

Pick three tasks you do regularly: maybe it's weekly reporting, client communication, or research analysis. Then, for the next seven days, collaborate with AI on one of these tasks daily, even if it's just for 10-15 minutes to start.

Don't aim for perfection. Aim for conversation. Ask AI to help you think through the problem, not just solve it.

Starter Collaboration Prompts:

• For Planning: "I need to [specific task]. What are 3-4 different approaches I could take? Help me think through the pros and cons of each."

• For Problem-Solving: "I'm stuck on [specific challenge]. Can you help me break this down and explore what might be causing the issue?"

• For Analysis: "I have this data/information [context]. What patterns do you see? What questions should I be asking that I might not have considered?"

• For Communication: "I need to explain [complex topic] to [specific audience]. Help me think through the best way to structure this and what key points to emphasize."

Track two things:

1. How much time this saves (or costs) you

2. What new perspectives or approaches emerge

This Month: The System

Once daily AI collaboration feels natural, it's time to build the system:

• Establish consistent times for AI collaboration (many people find mornings work well)

• Identify which types of collaboration feel most valuable

• Begin sharing learnings with colleagues (this builds your internal reputation as AI-forward)

• Start measuring quality improvements, not just time savings

Role-Specific Integration Examples:

• Managers: Use AI to prepare for 1:1s, analyze team performance data, and develop communication strategies

• Analysts: Collaborate on data interpretation, hypothesis generation, and presentation of insights

• Marketers: Work together on campaign strategy, audience analysis, and content optimization

• Sales: Partner on prospect research, objection handling, and deal strategy development

This Quarter: The Strategic Position

Now you're ready to scale your AI fluency into organizational leadership:

• Begin leading team conversations about AI integration

• Share frameworks and approaches that work

• Position yourself as an internal resource for AI adoption

• Look for opportunities to represent your company's AI capabilities externally

Weekly Fluency Check-In:

• Days you collaborated with AI (target: 5+ per week)

• Quality of insights gained (1-10 scale)

• Time saved vs. time invested

• New approaches or perspectives discovered

• Comfort level with AI as thinking partner (increasing over time)

Common Pitfalls to Avoid

As you build your AI fluency, watch out for these mistakes that derail progress:

1. The Perfect Prompt Trap: Spending 20 minutes crafting the "perfect" prompt instead of starting a 10-minute conversation. Remember: iteration beats perfection.

2. The One-and-Done Mistake: Using AI once for a task, getting mediocre results, and concluding "AI doesn't work for this." Most valuable insights emerge through back-and-forth collaboration.

3. The Comparison Complex: Measuring your AI outputs against human experts instead of against your previous capabilities. AI should enhance your baseline, not replace domain expertise.

4. The Tool-Switching Syndrome: Constantly jumping between AI platforms instead of building fluency with one. Pick Claude or ChatGPT and go deep before going wide.

What Organizations Get Wrong

For leaders reading this, the Section AI data reveals three critical factors that separate AI-fluent organizations from the rest: AI experts are 3x more likely to have clear company AI policies, 80% have access to paid LLMs (vs. 28% of skeptics), and 71% receive actual AI training (vs. 10% of skeptics).

The lesson? AI fluency doesn't happen organically—it must be engineered.

The Choice Point

While your colleagues debate which AI tool to adopt, you'll be building AI fluency. While they wait for perfect company policies, you'll be developing personal competitive advantages. While they worry about AI replacing them, you'll be positioning AI to accelerate you.

The Federal Reserve study showed that only 28% of workers are using generative AI at work to some degree. That means 72% are waiting; for training, for permission, for clarity, for the perfect moment.

That's your opportunity window.

The future belongs to professionals who treat AI as a daily thinking partner, not a weekend hobby. Who build systems for AI collaboration, not just goals for AI adoption. Who develop fluency through practice, not perfection through planning.

Stop chasing the perfect prompt. Start building the daily habit.

Your career is waiting on the other side of that choice.

Share Your AI Fluency Journey

Building AI fluency is more powerful when you're not doing it alone. We'd love to hear about your progress and help you navigate any challenges:

Share Your Wins: Drop a comment below with your biggest AI collaboration breakthrough this week. What surprised you? What worked better than expected?

Leave a comment

Get Unstuck: Hit a wall or feeling overwhelmed? Message us directly—we're here to help you work through specific challenges and find your AI fluency path.

Remember: the difference between an AI Novice and Expert is practice. The difference isn't credentials or certifications—it's actual time spent experimenting, building, and iterating.

The future belongs to the AI-fluent.

You're watching two brilliant teams talk past each other for the third time this week. The Product Marketing Manager paints the big picture: "strategic co-pilots," and "proactive intelligence," transforming entire workflows. They see the market, the competition, the user's career trajectory.

The Lead Engineer dissects the technical aspects: API latency, data throughput, prediction model accuracy. They see the endpoint, the code, the infrastructure that actually has to work.

Both teams are smart and committed, and yet they're getting absolutely nowhere.

The feature sits stalled, the roadmap stays blocked, and every meeting ends in polite, mutual incomprehension. No one is disagreeing; rather, they're operating at completely different altitudes. One team flies at 30,000 feet while the other works at 3 inches.

This is where ambitious professionals get trapped. We dig deeper into our default view, convinced that more focus will crack the problem. Marketing leaders zoom out further. Engineers zoom in tighter. Both approaches fail.

Real breakthroughs don't come from staring harder at the problem. They come from deliberately shifting your perspective. The solution isn't "in the weeds" or "in the clouds"; it's actually in the systematic movement between both views.

Subscribe now

The Strategic Edge: Why Most Problem-Solving Fails

What separates strategic thinkers from everyone else is their refusal to get stuck at one cognitive altitude.

While others pick a lane and stay there, strategic operators move fluidly between the big picture and the tactical details. They zoom out to see the entire board, then zoom in to spot the winning move. We call this the Full-Spectrum StrAItegy, and it's your systematic approach to cutting through any deadlock.

This isn't about working harder. It's about thinking clearly, and AI is the perfect thinking partner to help execute this framework.

The challenge? Forcing yourself to change perspectives is difficult. Our brains get locked into departmental thinking. Marketing sees strategy; engineering sees execution. Personal incentives reinforce these silos.

AI breaks this pattern. An LLM has no turf to protect, no ego invested in particular solutions. It can adopt a strategic view one moment and tactical focus the next, without bias or politics. This makes it your ideal partner for running a structured thinking process that breaks deadlocks.

Here's exactly how strategic leaders are using AI to solve what everyone else can't.

The Framework: Step-Back Prompting Meets Predictive Forecasting

This approach combines two advanced prompting techniques that transform AI from task-executor into strategic partner.

The "Zoom Out" (Step-Back Prompting): Instead of diving into your specific problem, you force the AI to identify underlying principles, industry forces, and strategic context. Think film director pulling back from close-up to landscape shot. This activates abstract reasoning and reveals strategic patterns invisible when you're too close to the action.

The "Zoom In" (Predictive Forecasting): Rather than static problem-solving, you prompt the AI to project 2-3 possible futures. This moves you from reactive thinking to strategic scenario planning, just like storyboarding the next few scenes to see where the plot leads. It forces consideration of second-order effects and long-term consequences.

The result? Your next move isn't just a reaction; it's a deliberate step toward competitive advantage.

Your Strategic Playbook: Copy, Paste, Win

Here's the prompt framework, ready for immediate deployment. Use this when you're stuck and need to break through to strategic clarity.

Before You Start: Fill out the "My Problem" section at the bottom with your specific situation, industry context, and desired outcome.

For AI, context is king. Vague problems lead to vague answers, so provide as many details as you can: Name the teams, the product/feature, and the exact point of friction. The more context you provide, the more targeted, relevant, and effective the AI's guidance will be!

🚀 The Prompt:

# Persona
You are a Senior Strategic Advisor. Your expertise lies in cutting through complexity to provide clear, actionable frameworks for decision-making. You will guide me through the "Full-Spectrum StrAItegy" to unblock a problem I am facing.

# The StrAItegy: Full-Spectrum Analysis
I will provide you with a problem or blocker. You will guide me through the following three phases, asking me for input where necessary and pausing for my response before proceeding to the next phase.

**Phase 1: Zoom Out (The Wide Shot - Step-Back Analysis)**
Based on my problem statement, your first step is to analyze the broader context. Generate 3-5 powerful, abstract questions that force a "zoom out" from the immediate issue. These questions should probe the underlying principles, hidden assumptions, and external forces that might be at play.

Examples of question types:
- "What are the fundamental principles of [relevant field] that apply here?"
- "What larger trends in [user's industry] might be shaping this situation?"
- "If we re-framed this problem as an opportunity, what would it be an opportunity for?"

Present these questions to me and wait for my reflections.

**Phase 2: Zoom In (The Final Scenes - Predictive Forecasting)**
Based on my reflections from Phase 1, your second step is to project potential futures. Generate 2-3 distinct and plausible scenarios of what might happen in the next 6-12 months if this problem is (or isn't) solved. For each scenario, describe:
- The likely outcome.
- The key contributing factors.
- A potential "headline" that would describe this future state.

Present these scenarios to me.

**Phase 3: Synthesize (The Director's Cut - Actionable Path)**
Based on the insights from the "Zoom Out" and "Zoom In" phases, your final step is to synthesize this full-spectrum view into a clear path forward. Provide 2-3 immediate, pragmatic, and actionable next steps I should take to address my original problem. Each step should be grounded in the context we've built.

---
# My Problem

**My Problem/Blocker:** [Clearly state the problem you are stuck on. Be specific. e.g., "My engineering and marketing teams are constantly misaligned on the go-to-market strategy for our new feature, causing launch delays."]

**My Industry/Context:** [Provide brief context. e.g., "We are a B2B SaaS company in the competitive cybersecurity space."]

**My Desired Outcome:** [What would success look like? e.g., "A streamlined go-to-market process where both teams are aligned and we can launch features on time."]

The Framework in Action: Breaking Through Real Deadlocks

Let's return to our stalled SaaS feature. Here's what strategic problem-solving actually looks like when you run it through the Full-Spectrum framework:

Phase 1: Zoom Out (Strategic Questions Generated)

What fundamental shift in our user's workflow does this "strategic co-pilot" actually create?

What is the core promise we're making to users with this feature, and how does our engineering philosophy [speed, accuracy, reliability] deliver that promise?

If we reframed this misalignment not as conflict, but as productive tension between "market narrative" and "product reality," where's the most powerful intersection?

Phase 2: Zoom In (Future Scenario)

Headline: "B2B SaaS Disruptor Redefines Analytics Category with Launch of 'Proactive Intelligence' Co-Pilot"

Outcome: The feature launch becomes a breakout success, praised for cohesive user experience that delivers on "thinking ahead" for users. Competitors scramble to catch up, still competing on dashboards while your company owns the "proactive intelligence" narrative.

Contributing Factors: PMM and Lead Engineer aligned on single "Strategic Operator" persona. API technical specs prioritized delivering proactive alerts. Launch campaign told that specific story, avoiding generic AI talking points.

Phase 3: Synthesize (Your Strategic Path)

Mandate Joint Definition: Task PMM and Lead Engineer to co-author one-page "Promise & Proof" brief. Must map top 3 "proactive intelligence" promises from marketing directly to top 3 technical functions that prove them.

Launch 1-Week Use-Case Sprint: Run rapid sprint storyboarding five critical alerts the "Strategic Operator" receives from the feature. Each storyboard requires joint sign-off from both teams.

Establish Shared Metric: Implement one cross-team metric for launch success: "Weekly active usage of proactive alert feature by Strategic Operator cohort."

Notice the structured approach. This is core StrAItegy methodology: by guiding AI through systematic reasoning (Abstract → Predict → Act), you get strategic insights instead of surface-level suggestions.

Scaling Your Strategic Advantage: Framework Across Career Levels

This framework scales with your ambitions and responsibilities:

Individual Contributors: Focus on project blockers. Use it to find new angles on stalled tasks or gain clarity on your piece of larger initiatives.

Managers and Team Leads: Apply to cross-functional challenges. Diagnose process friction, resolve team misalignment, or pressure-test new workflows.

Directors and Executives: Deploy for strategic dilemmas. Evaluate market-entry risks, assess competitive threats, or reframe entire business unit approaches.

Three Traps to Avoid

• Getting Stuck in the Clouds: The "Zoom Out" phase is for perspective, not a permanent vacation from the details. Use the high-level view to identify leverage points, then commit to coming back down to act on them.

• Getting Stuck in the Scenarios: The "Zoom In" phase is for foresight, not fortune-telling. Don't chase every possible future. Choose the most plausible 2-3 scenarios to clarify the stakes, then move to action.

• Outsourcing Your Judgment: The AI is your strategic partner, not your boss. Use its output to challenge your assumptions and reveal blind spots, but you still own the final decision.

The Professional Transformation: From Operator to Strategic Leader

You now have a systematic approach to getting unstuck.

The feeling of being trapped in a problem is a signal; not that you need to try harder, but that you need to shift perspective. By systematically zooming out then zooming in, you're not just solving immediate problems. You're building the strategic muscle that separates operators from leaders.

This is how you move from executing tasks to directing outcomes. From being constrained by immediate challenges to seeing the full spectrum of strategic possibilities.

While others stay trapped in their default altitude, you move fluidly between strategic vision and tactical execution. This flexibility becomes your competitive edge, making you indispensable when complex problems need solving.

You have the playbook. Your next challenge is waiting.

Take that issue sitting on your desk, run it through this framework, and see the board differently.

What kind of insights surfaced for you? Let us know in the comments.

Leave a comment

Ready for More AI Strategy Insights? Let's Keep You Ahead of the Curve.

We at StrAItegy Hub are working towards building tools and frameworks that turn AI complexity into competitive advantage. If you're serious about staying ahead in the AI-enabled workplace, we've got more for you.

Weekly strategic AI briefings and actionable frameworks

Subscribe now

Let's continue to learn and innovate, together.

Here's what nobody tells you about AI writing tools: they're incredible at cranking out content, terrible at capturing your voice. So you end up in this frustrating loop of prompting, editing, re-prompting, and editing again just to make it sound remotely human.

And you were told AI was supposed to make you more efficient.

Lucky for you, we've built a solution that actually delivers on that efficiency promise. Total time investment: 15-20 minutes to build your toolkit, then 2 minutes per email for the rest of your career.

Subscribe now

What We're Building: Your Authentic AI Voice Toolkit

Getting things done quickly matters. But the moment your emails sound AI-generated, you've lost more than you've gained.

Nobody wants colleagues thinking, "Did ChatGPT write this?" The fix isn't avoiding AI. It's teaching AI to write like you actually write.

This workflow uses Google Gemini to analyze your actual Gmail communication patterns (works with both personal Gmail and Google Workspace business accounts). You'll walk away with three powerful tools: a detailed Style Guide that captures your voice patterns (your personal reference document), Custom Gemini Gem instructions that create a one-click AI assistant trained specifically on your writing style (like having a personal AI copywriter who knows exactly how you communicate), and a versatile manual prompt template you can use with any AI model to instantly generate content in your voice. Together, they'll eliminate the guesswork from AI writing and generate emails and communications that sound authentically you, whether you're updating stakeholders or negotiating project timelines.

A Note on Platforms

This guide focuses on Gmail/Google Workspace with Google Gemini, but we know not everyone lives in the Google ecosystem. We're currently developing versions of this workflow for:

• Claude users who want to analyze their communication patterns

• Microsoft 365 teams using Outlook and Copilot AI

If you're in one of those camps, this Gmail/Gemini version will still teach you the core methodology. The follow-up guides with platform-specific instructions are coming soon.

Who This Works For (And Why You Need It)

Whether you're cranking out project updates as an IC, managing team communications, or crafting board-level messaging, authentic voice matters at every level.

Individual Contributors: Stop spending 45 minutes crafting the "right tone" for stakeholder updates. Get AI that writes clear project communications and professional cross-team emails without sounding like you copied corporate boilerplate.

Managers & Team Leads: Maintain your leadership voice across dozens of daily communications. From team check-ins to executive reporting, your AI clone handles the heavy lifting while keeping your authoritative-but-approachable tone intact.

Directors & VPs: Scale your strategic communication without losing impact. Whether you're aligning departments or briefing executives, generate high-stakes messages that carry your full authority and expertise.

Executives: Preserve your personal brand across board communications, company announcements, and external stakeholder messages. Delegate the drafting, keep the gravitas.

The common thread? Everyone needs to communicate more, faster, without their voice getting lost in AI generic-speak.

We've tested this method with executives, ICs, and everyone in between. The results? AI writing that consistently passes the "did I actually write this?" test.

Before You Start: Quick Setup Notes

A few prerequisites to sort out first:

1. You'll need Google Gemini access and your Google Workspace connected. Think of it as giving your AI assistant access to your communication filing cabinet.

2. Head to Gemini settings, navigate to the "Apps" section (link), and toggle "Google Workspace" to ON. This lets Gemini analyze your Gmail or content.

   

3. Critical success factor: Keep everything in one Gemini chat session. Gemini builds context like a detective gathering clues. Switching sessions means starting from scratch.

4. Model Selection: Use Gemini 2.5 Pro if you can (it's free as of this writing). Built for complex analysis work, so you'll get sharper results. Gemini 2.5 Flash works great too if you hit access limits.

   

Ready to stop planning the meeting about the meeting and actually get to work? Let’s go

Step 1: Teaching Gemini Your Communication DNA

Time to turn Gemini into your personal communication apprentice. We're analyzing your recent sent emails to map out your unique writing patterns.

Your move:

Start a fresh Gemini chat and paste this prompt. Grant access to Gmail if asked.

Copy this prompt:

Analyze my Gmail emails to understand my writing and communication style. To ensure a robust analysis, please:

1.  **Focus on emails I have *sent*.**
2.  **Analyze a substantial and representative sample:**
    * Aim for at least 50-100+ sent emails from the past year, if available.
    * Prioritize variety in recipients (e.g., manager, direct reports, clients, external peers) and communication purposes (e.g., requests, updates, feedback, persuasion, informational).
3.  **Identify and describe my communication style in detail.** Note key characteristics such as:
    * Overall tone (e.g., formal, informal, enthusiastic, measured)
    * Common formality levels and how they might shift with audience/purpose.
    * Typical sentence structure, length, and complexity.
    * Frequently used words, phrases, jargon, or idioms.
    * Commonly used openings, closings, and calls to action.
    * Use of questions, emojis, humor, or storytelling.
    * Implicit communication habits (e.g., things I consistently *avoid* doing, like using slang with executives, or patterns in response times if discernible).
4.  **Act as a world-class AI prompt engineer and create initial *drafts* of the following three items based on your analysis:**
    * **a. Basic Style Rules:** A preliminary set of rules summarizing my communication style.
    * **b. Draft Gemini Gem Instructions:** A draft set of instructions suitable for a Gemini Gem, designed to help generate communications in my style. These should be actionable and guide the Gem on key stylistic elements.
    * **c. Draft Manual AI Prompt Template:** A draft manual prompt template that includes:
        * Clear placeholders for essential task inputs, such as: `{{objective}}`, `{{recipient_name}}`, `{{recipient_role}}`, `{{key_message_points}}`, `{{desired_action_or_outcome}}`, `{{overall_tone_preference (e.g., default, more formal, more casual)}}`, `{{context_for_communication}}`.
        * A section like "[ASSISTANT INSTRUCTIONS: Maintain the following style derived from [Your Name]'s emails: (Summarize core style here)]" that you will pre-fill with a concise summary of my style.

Please present the style description first, followed by the three draft assets.

What you'll get:

Gemini returns its first analysis of your style plus initial drafts of your Style Rules, Gem Instructions, and Manual Prompt. Don't expect perfection here. These are rough sketches we'll refine into precision tools.

Step 2: Deep Style Training and Reality Check

Now we push Gemini from casual observer to method actor. It's going deep on your communication style, then proving it learned correctly by writing something new in your voice.

Your move:

Stay in the same Gemini session. Use this prompt template, replacing [Your Name] with your actual name and describing what you want Gemini to write in the <new_piece_to_create> section. Pick any business communication task as your test case to see how close the output is to your actual voice.

Copy this prompt:

<instructions_for_gemini>
    You are an expert "voice-cloner" and writer, tasked with deeply understanding and replicating [Your Name]'s communication style.

    <step_1_analyze_style_foundation>
        Reference the comprehensive analysis of **[Your Name]'s** communication style. This analysis was derived from all emails previously analyzed earlier in THIS SAME CHAT SESSION. Do not refer to external knowledge or new email examples; rely entirely on the prior analysis stored in our current conversation context.
    </step_1_analyze_style_foundation>

    <step_2_build_style_dna>
        From the comprehensive style analysis (from step_1_analyze_style_foundation):
        * Identify and list the core, recurring patterns defining **[Your Name]'s** Style DNA. This should include, but not be limited to: tone, sentence length/variability, preferred vocabulary/phrasing (including any unique expressions or jargon), typical pacing, use of humor (if any), formality levels and common shifts, structural habits (e.g., how arguments are presented, paragraph length), common openings/closings, and any other defining stylistic elements.
        * **[AI, determine these patterns based *only* on the prior analysis in this conversation].**
    </step_2_build_style_dna>

    <step_3_draft_and_refine_piece>
        <task_definition>
            Write the requested piece specified in `<new_piece_to_create>` (see <inputs_for_task> below) using the Style DNA identified in step_2_build_style_dna.
        </task_definition>

        <iterative_refinement_process>
            Let's set a confidence meter (0-100%) for how closely the draft sounds like **[Your Name]**, based on the comprehensive style analysis and the Style DNA. Initialize it.

            For each refinement round (minimum 50 rounds, or until confidence is >= 95% and stable for 3+ consecutive rounds):
            1.  **Self-Critique:** Compare the current draft against the **[Your Name]'s** Style DNA (from step_2_build_style_dna). Provide 1-2 sentences of specific feedback. Examples: "The tone is slightly too formal compared to the Style DNA's typical approach for this context," "Sentence structure needs more variability," "Incorporate more of the typical [phrase type] identified in the Style DNA."
            2.  **Adjust Style Rules (Internal):** Mentally (or in your internal scratchpad) adjust the application of the Style DNA based on the self-critique.
            3.  **Rewrite Anew:** Rewrite the piece from scratch to incorporate the feedback and better align with the Style DNA. Do not simply patch the previous version; aim for natural flow.
            4.  **Update Status:** Increment the iteration version number and update the confidence meter. Log the iteration number and confidence.

            Repeat this iterative process thoroughly.
        </iterative_refinement_process>
    </step_3_draft_and_refine_piece>

    <step_4_deliver_final_outputs>
        Once the iterative_refinement_process is complete (either 50+ rounds achieved OR confidence is >= 95% and stable for 3+ rounds):
        1.  Present the **final version of the <new_piece_to_create> *only***.
        2.  Immediately following the final piece, provide a debug block clearly labeled "DEBUG: FINAL STYLE DNA AND RULES" containing:
            * The fully refined **[Your Name]'s** Style DNA (as determined and refined through step_2_build_style_dna and step_3_draft_and_refine_piece).
            * A concise list of the final, actionable style rules that produced the final piece.
    </step_4_deliver_final_outputs>

    <constraints_and_mindset>
        * Your primary goal is to sound *exactly* like **[Your Name]** as represented in the analyzed emails, not like a generic AI.
        * Do not stop the refinement process prematurely. Rigorous iteration is key.
    </constraints_and_mindset>
</instructions_for_gemini>

<inputs_for_task>
    <new_piece_to_create>
        [User: Clearly describe the new email, message, or text you want Gemini to write in your style here. Be specific about the recipient, purpose, and key points to include. Example: "Write an email to my manager, Sarah Lee, asking for a brief 30-minute meeting next week to discuss the Q3 project plan. Mention I have a few preliminary ideas to share."]
    </new_piece_to_create>
</inputs_for_task>

<developer_note>
    In the past, models have sometimes undershot on achieving deep stylistic similarity. To ensure you don't do this, rigorously follow the iterative_refinement_process for at least 50 rounds OR until your confidence meter is 95%+ and has been stable for at least 3 rounds. For each round, internally note the iteration number so you don't lose track. Do not return a response until this refinement threshold is met. Ensure the final output strictly adheres to the identified Style DNA. The comprehensive style analysis from the initial email review (earlier in this chat) is the *only* source for style understanding.
</developer_note>

What you'll get:

A piece of writing that should sound distinctly like you. Gemini's been through intensive training at this point. The real test: does reading it make you think, "Yeah, I would totally write this"?

If Your Step 2 Results Aren't Perfect

Don't panic if your test writing doesn't sound exactly like you on the first try. Here's what usually helps:

• Too formal? Your Gmail might skew toward work emails. Mention this to Gemini: "My analysis might be too formal since it's mostly work emails. Adjust for a more conversational tone."

• Missing your personality? Add a follow-up prompt: "The writing captures my structure but misses my personality. Analyze for humor, storytelling, or unique phrases I use."

• Not enough email history? If you have fewer than 50 emails, try: "I have limited email history. Focus on the strongest patterns you can identify and note areas where you're making educated guesses."

Remember: Step 3 is where we turn these insights into foolproof daily tools.

Step 3: Building Your Final Arsenal

Gemini understands your voice. Now we convert all that learning into practical, everyday tools.

Your move:

Still in the same session (seeing a theme here?), run this prompt right after Step 2 finishes. Replace [Your Name] with your actual name.

Copy this prompt:

Excellent work on the style refinement. Now, using all insights gathered from the initial email analysis AND the intensive refinement process (including the Style DNA and rules from the previous XML prompt's debug output), please:

1.  **Synthesize a Comprehensive Style Guide for [Your Name]:**
    * This document should be detailed and actionable, serving as a go-to reference.
    * Structure it with clear sections. I suggest including:
        * **Overall Voice & Persona:** (e.g., helpful expert, friendly collaborator, formal authority).
        * **Core Communication Principles:** (e.g., clarity first, action-oriented, maintain positivity).
        * **Tone & Formality:** (Default tone, how it adapts, formality levels for different audiences).
        * **Vocabulary & Phrasing:** (Commonly used positive/negative words, unique phrases, jargon to use/avoid, preferred alternatives for common words).
        * **Sentence Structure & Pacing:** (Typical length, complexity, variability, use of short/long sentences for effect).
        * **Openings & Closings:** (Standard and alternative ways to start/end communications).
        * **Calls to Action:** (How requests or desired next steps are typically framed).
        * **Formatting Preferences:** (e.g., use of bullet points, bolding, paragraph length, email signatures).
        * **Emoji & Humor Usage:** (Guidelines on when/how, if applicable).
        * **Context-Specific Adaptations:** Crucially, detail how my style adapts across different communication contexts identified from my emails (e.g., Formal to CEO, Semi-Formal to clients, Informal to team, Task-Oriented updates, Persuasive requests). For each context, highlight the key stylistic adjustments.
        * **Things to Consistently Avoid:** (e.g., specific clichés, passive voice, excessive hedging).
2.  **Compare and Verify (Self-Correction Step):**
    * Briefly compare these newly synthesized comprehensive style rules and the Style Guide outline against the *initial* style description and *initial* draft AI prompt components that were generated when we *first* analyzed my emails (before the XML refinement prompt).
    * In a short paragraph, highlight 2-3 key refinements, new insights, or increased nuances gained through the more intensive XML prompt refinement process. (This helps confirm the value of the refinement step).
3.  **Generate Final, Polished Style Assets:**
    * Using the comprehensive Style Guide from point #1, please provide the following three outputs, formatted clearly and ready to save/use:
        * **Output 1: "[Your Name]'s Communication Style Guide"** (The full document as synthesized above).
        * **Output 2: "Updated Gemini Gem Instructions for [Your Name]"** (A concise, actionable set of instructions suitable for inputting into a custom Gemini Gem. This should distill the most critical elements of the Style Guide, with a focus on actionable guidance for achieving the style and adapting to context).
        * **Output 3: "Updated Manual AI Prompt Template for [Your Name]"** (An updated version of the generic manual prompt template. The style guidance section, e.g., "[ASSISTANT INSTRUCTIONS: Maintain the following style...]", should be pre-filled with a concise distillation of my core style principles, common phrases, and key context-adaptation rules derived from the comprehensive Style Guide. Ensure placeholders like `{{objective}}`, `{{recipient_name}}`, etc., are still present).

Ensure all three final outputs accurately and richly reflect the final, refined understanding of my writing style, incorporating the depth from the iterative process.

Expected Output (Your Personalized AI Voice Arsenal – Delivered!):

This is the moment you've been working towards! Gemini will provide you with three final, personalized assets, all embedded with your personal voice:

• Your Comprehensive [Your Name]'s Style Guide

• Your Updated Gemini Gem Instructions

• Your Updated Manual Prompt Template, now hardcoded with your unique style nuances

Time to Deploy Your Voice Arsenal

You've built something powerful here. A system that lets AI write like you actually write, not like AI thinks you should write.

So, what’s the bottom line for your Monday morning and beyond?

• Save Your Assets: First things first, save those invaluable outputs from Step 3. These are your new productivity weapons.

• Activate Your Gem: Use the “Gemini Gem instructions” deliverable to create your custom Personal Writing Gem here. Gems are Google’s version of Custom GPTs, and creating this gives you one-click access to your writing style within Gemini.

• Go Manual When Needed: Keep that Manual Prompt Template handy. It’s perfect for using with Gemini when you want more granular control, or if you’re trying to bring your writing style to other AI models (ChatGPT, Grok, etc).

• Consult Your Guide: Your Style Guide is your personal reference. Store it somewhere safe (not in your downloads folder) so you can ensure all of your AI-generated content is perfectly aligned.

This toolkit is designed to significantly boost your communication efficiency while ensuring every AI-assisted message still sounds genuinely like you.

Maintenance note: Your communication style evolves. If your role changes significantly or you feel your writing has shifted, re-run this process. Think of it as updating your voice firmware.

Found This Guide Valuable? Share it with a colleague who's still wrestling with robotic AI writing. Help them join the ranks of professionals who've cracked the code on authentic AI communication.

Share

Ready for More AI Strategy Insights? Let's Keep You Ahead of the Curve.

We at StrAItegy Hub are working towards building tools and frameworks that turn AI complexity into competitive advantage. If you're serious about staying ahead in the AI-enabled workplace, we've got more for you.

Daily insights and sharp takes on AI in corporate environments

Follow StrAItegy Hub on X

Weekly strategic AI briefings and actionable frameworks

Subscribe now

Let's continue to learn and innovate, together.

The AI conversation is everywhere, but it sounds like it's coming from that overly enthusiastic consultant who only speaks in buzzwords. People are nodding along because they can see the potential, but nobody knows how they're actually supposed to apply it practically into their life.

You might be crushing it at work. There's strategy to execute, teams to lead, and deliverables that won't write themselves. But between the headlines about job displacement and the endless tool launches, suddenly "prompt engineering" lands on your skill wishlist, right next to "surviving another 6 AM stakeholder call."

The AI hype machine is extremely loud, especially with tools seeming to multiply faster than meeting invites. Everyone's promising revolution, but the roadmap? That's still in development.

Here's where StrAItegy Hub comes in.

We're not another voice shouting into the AI void.

We decode hype into how-to’s. We turn shiny new tools into systems you can trust. We make AI useful for people who think, lead, and build.

What We're Building for You

This isn't just content. It's your strategic enablement platform:

• StrAItegy Edge: Real workflows for real work. Prompt packs that solve actual corporate challenges. Think IKEA instructions for assembling your AI advantage but without the leftover screws.

• StrAItegy Pulse: Weekly AI intel without the noise. We'll tell you what matters and why, so you can stop sifting through endless tech blogs and get back to executing.

• StrAItegy Lab: We test so you don't have to. Unbiased tool comparisons, honest experiments, and straight answers about which model actually handles that urgent report better.

• Signal Boost Podcast: Real professionals, real stories, zero guru nonsense. Your peers in strategy, product, marketing, and ops sharing what worked, what bombed, and the lessons that stuck.

This Is for Doers, Not Dreamers

Corporate strategist juggling competing priorities? We get it. Consultant managing multiple clients? Been there. IC trying to make your mark? We see you. Leader navigating the future of work? We're building this for you.

Our goal: Make you not just AI-aware, but AI-advanced in how you think, operate, and lead.

Ready to Stop Waiting for AI Clarity?

Subscribe now. Get first access to everything we build. Help shape what StrAItegy Hub becomes.

We're not aiming for perfection from day one. We're aiming for relentless utility.

Subscribe now

Let's build the future of smarter work together.

The StrAItegy Hub Team