I was on a plane watching a kid in the row ahead of me operate his phone like a slot machine.

Instagram. TikTok. Messages. Back to Instagram. His thumb moved so fast I couldn’t tell whether he was looking for something or trying to outrun it. The phone was very close to his face.

Just when I started to drop into a trance and zone out I see the warning screen pop up on his phone:

iPhone may be too close.

He pulled his phone away just far enough to close the warning and then went right back to doomscrolling. That warning stuck with me for a while.

His phone could tell when the screen was too close to his eyes, and somehow it had no idea when his feed was getting too close to his thinking.

Back in June, I wrote that AI is making smart people dull. I was writing about atrophy. I kept thinking about the other side of it: what are we supposed to train?

We live inside systems trained to give us more of what already held our attention. Spotify learns the songs we finish. TikTok studies what keeps us from swiping. Amazon predicts the next thing we might buy. Netflix rarely wants to surprise us enough to make us leave.

Now we have generative AI in the same loop. Any of us can ask ChatGPT, Claude, Gemini or whatever launches next week for a hundred ideas before the flight attendant reaches our row with the drink cart.

It feels like abundance. Sometimes it is.
But when we have more answers than we have time to form good questions, we inherit another problem.

And the output is getting more boring and lame. Same LinkedIn posts, same business ideas, the same five talking points rewritten 400 ways. People are flinging beautifully formatted word salad slop into every Slack channel and shared drive. There is a homogenized layer of lame settling over the internet, and the AI slop and AI sprawl are driving me a little insane.

The machines will never be less capable than they are today. What scares me is how capable they already are—and how predictable we’re becoming every day.

Claudefishing is only half the problem

In July, Substack launched an AI-text scanner and Chris Best gave the behavior it was built to expose a name: Claudefishing. A reader thinks there is a person on the other end of the words.

In reality, nobody did much thinking at all.

In August, Anthropic announced that future Claude models will watermark their text. The mark will be invisible to readers, but detectable with the right key.

I understand why both companies are doing it. Authorship matters. Trust matters. If I am asking you to pay attention to my words, you deserve to know whether I stood behind them or rented them by the token.

The scanner still can’t answer the question I care about most: Did a human see anything before the machine started writing?

You can disclose every tool, label every output and still publish a beautifully formatted pile of lame nothings. You can also use AI all day and produce something that could only have come through your particular mix of experience, curiosity, bad decisions, taste, scars and weird obsessions. The dividing line is what the human imagined before, during and after using the tool.

The price of intelligence collapsed

For most of history, intelligence was expensive to access.

Research took time. Analysis lived inside specialized people. A small company couldn’t put a strategist, researcher, analyst, copywriter and software developer to work on the same problem without spending serious time and money. Now some version of that team lives in a browser tab for the price of a streaming subscription.

Stanford’s 2026 AI Index found that the leading models are clustering closer together in capability while consumer access remains free or close to it. The 2025 report found that the cost of running a model at roughly GPT-3.5 performance had fallen more than 280-fold in less than two years.

Routine thinking and cognitive output are basically free at this point for machines. That change is where the premium needs to go, in my opinion.

Intelligence helps solve a problem you can name. Imagination notices that the name itself may be trapping you. It lets you see patterns, possibilities and futures that do not exist yet. It finds a door in a wall everybody else has accepted as load-bearing.

Ann Pendleton-Jullian, the architect and educator who wrote Pragmatic Imagination with John Seely Brown, describes imagination as the capacity to “see not only what is, but what could be” and to “discover unknown, undiscovered unknowns”. She treats it as something practical. A muscle for finding what we have not learned how to look for.

The longer version of that quote moves me closer to tears. It feels like a rallying cry in my nervous system. Creativity gets reduced to brainstorming, aesthetics or making a campaign look cooler. The part I care about goes much deeper. Imagination changes what becomes visible before anybody starts making.

That is why I have started thinking about AI as applied imagination. I got so obsessed with it I made an entire visual essay about it.

Applied Imagination

Are LLMS rotting our brains?

That is the tabloid version of the question, so I went looking for something better than a gut feel.

The study with the irresistible headline came out of MIT Media Lab in 2025: Your Brain on ChatGPT. Researchers split 54 people into three essay-writing groups. One used an LLM, one used search and one had to use the technology between their ears. They measured brain activity with EEG.

The LLM group showed the weakest neural connectivity. They reported less ownership of what they wrote and had more trouble quoting their own essays. The researchers called the potential buildup “cognitive debt,” which is a wonderfully academic way of saying the bill for easy thinking may arrive later.

Cue the brain-rot headlines.

Pump the brakes. The MIT paper is a preprint, not a peer-reviewed verdict. Only 18 people completed its final crossover session, and a later technical critique raised real questions about the sample, study design, reproducibility and EEG analysis. It does not prove ChatGPT causes brain damage. It does suggest that when the machine does the first thinking, less of you may show up in the memory, effort and ownership of the work.

The stronger evidence is less cinematic and more useful.

In 2025, researchers at the University of Pennsylvania ran a randomized field experiment with nearly 1,000 high school math students. Students using a standard GPT-4 interface performed 48 percent better during practice. Then the researchers took the tool away. On the unassisted exam, those students scored 17 percent worse than students who never had it.

The tool made them look smarter while it was in the room and left them less capable when it walked out.

There was a better version. A guarded AI tutor gave hints instead of coughing up answers. Practice scores rose 127 percent, and the damage to unaided performance was largely erased. The researchers watched the ordinary GPT group use AI as a “crutch.” The guarded group had to keep doing some of the lifting.

That is the whole game. Performance and learning are not the same thing. AI can help you finish today’s work while starving the skill that tomorrow’s work will require.

Adults appear to be making a similar trade at work. A 2025 CHI study surveyed 319 knowledge workers and collected 936 real examples of AI-assisted tasks. Higher confidence in AI was associated with less critical thinking. Higher confidence in one’s own abilities was associated with more. This was self-reported behavior, not a brain scan, but the pattern is familiar: the people who hand over judgment soonest tend to inspect the answer least.

The researchers describe a move “from task execution to oversight”. That sounds like an upgrade until you notice what can disappear in the handoff. Oversight may catch a bad answer. It cannot recover a thought you never formed.

These studies do not prove permanent brain atrophy. They give us something more actionable than a scary MRI headline. If the machine always takes the first rep, the skills underneath stop getting reps at all.

Good answers can block better ones

Psychologists call it the Einstellung effect. A familiar solution appears, attention locks onto it and a better solution becomes harder to see.

Researchers Merim Bilalić, Peter McLeod and Fernand Gobet watched this happen with expert chess players. The players were shown a position with a familiar move and a better, less obvious one. Even after they said they were searching for alternatives, their eyes kept returning to the squares involved in the first solution.

The paper’s title is almost the entire warning: Why Good Thoughts Block Better Ones. The first idea occupied their thoughts and directed what they could see. AI can create that condition on demand.

Ask for “the best campaign idea” and a polished answer appears before you have formed a view. Ask it to redesign onboarding and it immediately hands you the same familiar furniture: a welcome email, a checklist, a dashboard, a progress bar. The response feels useful because it is coherent. Coherence is exactly what makes the anchor hard to notice.

The effect has been measured in design work too. In a classic 1991 study, engineering students were asked to design a spill-proof coffee cup. One group saw an example with a straw and mouthpiece. The researchers explicitly pointed out that the example leaked and told the students to avoid those features.

They copied them anyway. The students who saw the example produced far more designs that resembled it than the control group. The bad example narrowed the field, even while sitting there with its flaws labeled.

Think about how most teams use AI. The machine shows the example first. Then the humans react to it. By the time the meeting starts, the field is already smaller.

The moment the room changed

I saw the reverse happen recently with a client.

They are a billion-dollar company that works with some of the biggest brands in the world. I was in a room with people who knew the domain far better than I did, working through a problem that involved AI, data and human behavior.

They were smart. They had the architecture, the inputs, the outputs and the data sources. They were moving through the problem exactly the way experts are trained to move through a problem.

And they were trapped inside it. I was the outsider, which gave me permission to be a little annoying. I started asking questions that did not seem to belong in the room.

What does this look like through the psychology of a dating app? What could a card game teach us about the interaction? Where is the tension a behavioral scientist would notice? What would happen if we treated the experience like entertainment instead of enterprise software?

I got the side-eye. Fair. They thought I was the weirdo, and they were right. The world needs more weirdos. Then somebody stopped and said, “Wait. Maybe we could try that.”

The room changed. People stopped explaining the system we already had and started leaning into a system that did not exist yet. New questions appeared. The shape of the problem moved.

That tiny moment is why I do this work. It is why I write this newsletter. I did not know more than the experts. I helped the room access knowledge it already had from a different direction. There is research behind that move.

In a set of experiments published in Cognitive Psychology, Mary Gick and Keith Holyoak gave people a difficult medical problem. A doctor needed to destroy a tumor with radiation without damaging the healthy tissue around it. Only about 10 percent found the convergence solution on their own.

Another group read a military story with the same hidden structure: a general divides an army into smaller groups and approaches a fortress from several directions at once. Without being told that the story mattered, roughly 30 percent transferred the idea. When researchers gave a simple hint to connect the two problems, the solution rate jumped to 75 percent.

The knowledge was available. They did not need a smarter doctor. They needed a story from somewhere else to change the shape of the problem. Crossing the distance between two things that look unrelated is imagination at work.

Use AI to widen your creative aperture

The usual AI workflow does the opposite. We describe the problem once, inside our current assumptions, and ask the model to solve what we just handed it. The model becomes an answering machine for our existing frame.

I have been testing a different sequence. The answer comes later. First, I use the machine to attack the shape of the problem.

Read more

I’m on the train back from the city, worn out from a day rolling out an AI solution for a leading QSR brand. I’m reading about psychology and anthropology, the kind of reading I do to shut my brain off. This time it did the opposite.

It was about the pieces of human life that never make the official record. Unconscious rituals. Things too ordinary to celebrate or photograph.

A quick-service restaurant brand was rattling around in my head at the same time. We’d been asking what “moments” the brand should notice, and the strategy brain kept giving the same answer. Lunch. Road trips. Game day. Family dinners. Hangovers.

You know the deck.

Then that book changed how I was looking at the problem.

I started thinking about the first fry you steal from the bag before you pull out of the parking lot. The person who swears they’re not hungry, then eats half of yours. The bag inspection before you leave the drive-thru, because you’ve been burned before. The pickle pulled off one burger and handed to the one person who actually likes pickles.

Nobody puts this in a focus group answer. You’ve done some version of it hundreds of times and never once put a sentence to it. That’s exactly why it’s interesting.

A customer can tell you what they think about a restaurant brand. They’re much worse at telling you about the three fries they ate between the window and the first red light.

I think there’s an enormous amount of value hiding in that gap. Stick with me. I’ll show you how to use AI to go find it.

The things people forget to tell us

Most strategies start with whatever people can articulate.

We ask customers what they want, employees what frustrates them, executives what keeps them awake, then we organize those answers until they start looking like insight.

There’s nothing wrong with that. I’ve spent twenty years doing versions of it.

But underneath the world people can explain is another one made out of habits, workarounds, weird little rituals and things that became normal so long ago nobody remembers choosing them.

It’s the spreadsheet that eats every Thursday afternoon. It’s the CMO who opens the same deck six times because something still feels wrong, even though they can’t tell the creative team why. It’s the employee who tells leadership that the new process is “working pretty well” while keeping the old spreadsheet open on their second monitor.

It’s also the executive who has been put in charge of AI transformation and then privately asks somebody half their age which AI tool they should be using. Or the founder who tells you they need a website when the real problem is that the company changed and they no longer know how to explain what it is.

These moments don’t look strategic when you first encounter them. Most of them look like noise. Until you notice the pattern.

The problem people give you is often downstream from the behavior they stopped noticing.

That idea has been following me around.

A six-week problem disguised as “just a report”

I was in Seattle recently working with a 3-billion-dollar events company on AI fluency. One of the people we worked with was a procurement analyst who had a report he built every week.

There was nothing particularly dramatic about it.

He collected information, moved it around, reformatted parts of it, checked things, assembled the output and sent it along. It took about four and a half hours every week.

He had been doing some version of it for years, which is what makes the story interesting. Nobody was storming into the executive meeting demanding action on the four-and-a-half-hour report crisis. There was no transformation initiative named after it. Nobody had even framed it as a problem.

It was just part of his job.

Four and a half hours every week works out to roughly 234 hours a year. Nearly six work weeks evaporated into a task everyone had accepted as “just a report”.

We sat with him and went through what he actually did. Where the data came from. Which steps were mechanical. Which parts required judgment. Where he checked himself. What happened when something was wrong.

Then together we built an agent around the work.

When he ran it and saw that it worked, he just stared at the screen for a second.

That stare stuck with me.

Of course there was an automation story there, but the AI wasn’t the part I cared about most. The interesting thing was watching someone discover that a piece of his life he had accepted as permanent wasn’t permanent at all.

We love giant transformation metrics because they look good on slides. Sometimes transformation is a guy realizing Wednesday afternoon belongs to him again.

You can’t see that KPI from 30,000 feet. You have to get close enough to notice it.

The company that exists when nobody is looking

I saw a different version of this while working through AI adoption with another organization.

We were trying to understand how people inside the company were actually using AI. Not what leadership believed was happening. Not what the policy allowed. What people were really doing when they had work to finish.

At first, assessments had names attached to them and the picture looked fairly clean. Once people could answer more freely, a messier company appeared.

People were experimenting outside approved tools. Shadow IT surfaced. Confusion showed up. So did fear. There were things people had been doing for weeks that barely existed in the official version of the organization.

That taught me something I keep coming back to.

There is the company described in the town hall, the operating plan and the leadership meeting. Then there is the company that exists at 11:23 p.m. when somebody has a deadline, the official process sucks and no one is watching.

The second company is usually much more useful to study.

That’s where you discover what people value enough to work around. You see which rules survive contact with reality and which ones become theater. You find the person everyone secretly depends on, the spreadsheet nobody is supposed to need and the unofficial tool that somehow works better than the expensive one everyone was told to use.

We tend to treat these things as exceptions to clean up.

Sometimes they’re messages.

A workaround is a tiny product brief written in behavior.

Sometimes the answer is hiding six years behind the question

A few days ago I was talking with the executive chairman of a company that has been around for generations. We were talking about AI and he asked me some version of the question everybody is asking right now.

What would you do if you were me?

I could feel my brain loading the usual ammunition.

Frameworks. Roadmaps. Operating models. Transformation plans.

I’ve got plenty of them.

Instead, I asked him what they did during COVID.

He looked at me sideways because we had been talking about AI and I had suddenly transported the conversation back to 2020.

But that was the point. There wasn’t a playbook for COVID either, especially for a company whose business lives heavily in physical events. They had already lived through a moment when the map disappeared.

He smiled and told me COVID almost killed them, then in another sense saved them. We spent the next stretch of the conversation talking about how the company had behaved when there was no obvious answer.

The thing he needed wasn’t sitting inside an AI framework. It was hiding inside his own scar tissue.

That changed the entire conversation for me. We’re often so busy importing expertise that we overlook the evidence sitting inside someone’s history. The capability looks invisible because nobody thought to label it.

The right question (and context) makes it visible again.

Maybe this is the work

I’ve been thinking about this in the context of my own work too.

For years people have come to me asking for things. A brand. A website. An app. A campaign. A strategy. More recently, an agent, an automation or some other flavor of AI thing.

I can do those things. My companies do those things every day.

But the conversation I love starts somewhere else.

Bring me the problem that keeps bothering you. Tell me the thing you’ve been circling for six months. Show me where your team keeps getting stuck. Tell me what feels wrong even though you don’t have the language for it yet.

Then give me a little time to look around.

A client recently asked me to spend time with his CTO and gave me one instruction: expand his aperture.

An aperture controls how much light gets into the frame. Open it and you don’t necessarily create something new. You allow yourself to see more of what was already there.

That feels closer to the work I care about.

Sometimes you zoom out. Sometimes you get painfully close. You look at a business problem like a psychologist, a technology problem like an anthropologist, a creative problem like an engineer.

Then something moves.

• The AI problem turns out to be a trust problem.

• The productivity problem turns out to be a prioritization problem.

• The request for a website is actually a CEO admitting, without saying it, that the old story no longer fits the business.

Problems rarely arrive wearing the right name tag.

That might be one of the most useful lessons I’ve learned.

Look at the five minutes around the thing

There’s a simple practice I’m using more often now.

When I’m trying to understand something, I stop staring directly at the main event and look at the edges around it.

What happens five minutes before?

An executive rewrites the first slide three times. Someone texts a colleague to make sure a number is right. A manager asks ChatGPT a question they didn’t want to ask their team. A customer checks their bank balance before hitting buy. A teenager turns down the music before pulling into the drive-thru so they don’t screw up the order.

That is all information.

Then look five minutes after.

Which slide gets screenshotted? What gets forwarded? Who stays behind after the meeting? Does the customer throw away the packaging or keep part of it? Does the employee use the new system tomorrow or immediately return to their spreadsheet?

We fetishize the main event because it’s easy to name. The purchase. The meeting. The presentation. The workflow. The campaign.

I’m becoming more interested in the wake it leaves behind.

That wake tells you what mattered.

The invisible moments scan

Once you start looking this way, there are a handful of places where the good stuff keeps showing up.

Watch the rituals. Pay attention to small things people repeat without discussing them. Repetition turns choices into furniture. Eventually nobody remembers the room without them.

Find the workaround. Whenever somebody has built their own spreadsheet, side channel, saved prompt, template, or tiny system, get curious before you clean it up. They may have solved a problem the official process refuses to acknowledge.

Look for contradictions. A company says creativity matters, then fills every available hour with meetings. A leader talks about empowerment, then rewrites every decision. An organization calls itself AI-first while people are embarrassed to admit they don’t know how to prompt. The gap between what we say and what we repeatedly do is usually where the interesting stuff is.

Follow the residue. Screenshots, receipts, abandoned drafts, open tabs, version names, Slack messages, crumbs in the car. Behavior leaves artifacts. The artifacts often tell a cleaner story than memory does.

Pay attention to embarrassment. The whispered question can be more valuable than the question asked in front of the room. People hide the places where identity and competence feel threatened. Those places are loaded with unmet needs.

Find the tiny tax. Ten minutes doesn’t sound like a business problem. Ten minutes every week across 400 people for three years absolutely is. The most expensive processes in an organization don’t always hurt enough in a single moment to trigger change.

And look for the uncelebrated win. Going home on time. Feeling prepared walking into the room. Not needing to ask for help. Knowing you won’t look stupid. Getting four hours of your week back.

Companies measure outputs. People experience their lives in moments.

And both matter.

AI makes recorded knowledge cheap. Observation gets more expensive.

There is a reason I think this matters even more now.

AI is astonishingly good at working with what has been captured. Give it the research, the transcripts, the reports, the internet, the documents and the customer feedback, then it can compress, compare and remix that material faster than any human team I’ve ever managed.

But there is a catch.

AI can synthesize what has been recorded. Its blind spot is often the thing nobody thought to record

The best prompt in the world can’t retrieve a behavior nobody noticed.

A model can tell you a thousand things about fast food culture. It can even invent a perfectly plausible story about fry stealing.

What it cannot tell you, without evidence, is whether your customers actually do it often enough for you to build strategy around it.

Someone still has to notice.

It can summarize an employee survey. It can’t make an employee feel safe enough to tell you what they were embarrassed to write down.

It can generate twenty AI transformation frameworks for an executive.

It won’t know that this executive’s most useful AI lesson might be sitting in how his own company survived COVID unless someone brings that history into the room.

This is where human judgment starts getting more valuable, not less.

The machine gives us enormous pattern recognition across the things the world bothered to document. Our job is still to notice the thing sitting three feet away that everybody decided was too ordinary to matter.

That’s where I want to spend more of my time.

Go three inches to the left

The thing I read this week keeps rattling around in my head because it gave me language for something I think I’ve been doing instinctively for years.

Look for the undocumented. The unfinished. The uncelebrated. The unphotographed. The unphotogenic.

Look at the crumbs instead of the campaign. Look at Wednesday instead of the annual productivity report. Look at what the employee does after the training, not what they say on the survey. Look at the sentence somebody almost says and then swallows.

The obvious part of the problem is crowded. Everyone is staring at it.

Next time somebody hands you a brief, a strategy problem, a customer journey, a broken workflow or a giant hairy question, try looking three inches to the left.

You may find the thing everybody else walked past.

And that may be the whole advantage.

Steal this: The Invisible Moments Scan

I turned the idea into a prompt (and a skill/agent) so you can try it on something you’re working on right now.

I want you to run an Invisible Moments Scan on this:

[PASTE A BRAND, PRODUCT, AUDIENCE, BUSINESS PROBLEM, WORKFLOW, TRANSCRIPT, RESEARCH SET, OR SITUATION]

Do not solve the stated problem yet.

First tell me the official version of the situation. What would normally appear in the brief, dashboard, process map, customer journey, research report, or executive deck?

Then look around it for the behavior people are unlikely to mention because it feels too ordinary, unconscious, private, messy, embarrassing, unfinished, repetitive, or insignificant.

Specifically inspect:

1. The 5 to 10 minutes before the obvious event.
What are people checking, hiding, preparing, rehearsing, searching, arranging, delaying, or improvising?

2. The 5 to 10 minutes after.
What gets screenshotted, forwarded, cleaned up, saved, discarded, questioned, redone, celebrated, hidden, or abandoned?

3. Unconscious rituals.
What tiny behaviors happen repeatedly without anyone deciding to do them?

4. Workarounds.
Where have people invented spreadsheets, side channels, personal tools, templates, saved prompts, manual steps, or unofficial systems because the official system is not good enough?

5. Contradictions.
Where does stated belief differ from repeated behavior? Compare values with calendars, policy with practice, strategy with spending, and stated preference with actual choice.

6. Private behavior.
What changes when nobody is watching? Look for embarrassment, fear, guilty pleasures, quiet confusion, hidden tools, whispered questions, shortcuts, and behavior that would never appear in an advertisement or case study.

7. Residue.
What evidence does the behavior leave behind? Look for browser tabs, receipts, screenshots, drafts, sticky notes, wrapper crumbs, calendar patterns, version names, Slack messages, duplicate spreadsheets, saved prompts, and open loops.

8. Micro-costs.
What tiny repeated tax has become so normal that nobody calls it a problem anymore? When the evidence supports it, translate that tax into weekly and annual time or money.

9. Uncelebrated wins.
What small change would matter deeply to the person but never appear in the normal KPI dashboard? Think leaving on time, feeling prepared, avoiding embarrassment, getting an afternoon back, trusting the answer, or not needing to ask for help.

Give me 15 Invisible Moments.

For each one, include:

MOMENT:
Describe a specific scene I can picture happening.

VISIBLE BEHAVIOR:
What could actually be observed or verified?

HIDDEN TENSION:
What desire, fear, identity, tradeoff, frustration, or emotion might sit underneath it?

WHY IT MATTERS:
What could this tiny moment reveal about the larger problem?

OPPORTUNITY:
What could a brand, product, leader, creator, or organization notice, remove, amplify, redesign, reward, name, or build around?

EVIDENCE STATUS:
Label it Observed, Supported, Hypothesis, or Needs Evidence.

Do not fabricate customer truths.

When evidence is missing, call the idea a hypothesis and tell me exactly what I would need to observe to prove or kill it.

Then score every moment from 1 to 10 for:
- Surprise
- Emotional truth
- Strategic value
- Specificity
- Ownability

Choose the five strongest.

Finish with four things:

THE INVISIBLE TRUTH
One sentence explaining what everybody else studying this problem is likely to miss.

THE OPPORTUNITY
The most useful strategic or creative move created by seeing it.

THE STORY
A one-sentence hook that could become an article, campaign, product pitch, presentation opening, or social post.

WHAT TO WATCH NEXT
Three real-world behaviors or artifacts I should observe next to test whether this insight is actually true.

Prefer mundane specificity over clever language.

I don’t want trends.

I want the thing everybody has seen and nobody thought mattered.

If you want to turn this into a skill easily you can ask claude to “create a skill from this” or you can use our free skill creator MakeMySkill.

I’m driving home from a client meeting with three ideas in my head and 35 minutes of road ahead of me.

It’s the middle of the afternoon and my phone rings. Dave.

He’s not calling to say hello. He’s in one of those frustrated founder panics where five thoughts arrive before the first sentence is finished.

“How the hell are you getting so much done with AI?”

Then he starts telling me how good he’s gotten at using it.

I love Dave. He was my first cofounder. He’s a serial entrepreneur and one of the smartest people I know. I also don’t have the heart to tell him he’s using billion-dollar alien technology like a glorified Google search.

So I tell him my actual secret. It isn’t a better prompt or the flux capacitor agentic widget of the week.

It’s the spaces in between.

The 20 minutes before pickup. The drive home. The gap between a client meeting and dinner. The parts of the day where I’m technically unavailable to sit at a computer but my brain is very much still working.

Dave immediately pushes back.

“That’s my problem. I’m driving all the fucking time.”

Exactly. For years, my car was a coffin. Not because it was dangerous, but because that was where usable time went to die. I’d leave a meeting with three ideas in my head, get in the car, drive 30 minutes, get home, get ambushed with family sh*t to do, and maybe remember one of them the next morning.

If I was disciplined, I’d record a voice memo.

Then six months later I’d find it beside a recording of my son singing about chicken nuggets (the ones shaped like dinosaurs).

My problem wasn’t that I had no time. It was that I didn’t have a good interface for the time I already had. So I started building around that.

And because Dave is Dave, our conversation eventually turned into me building him a ridiculous version of the same thing.

He can call a phone number and have a conversation with his own agent. It sounds nearly indistinguishable from Scarlett Johansson. It knows enough context to help him think, capture ideas, orchestrate tasks, book and rebook meetings, remember commitments, and keep track of the things he would normally have to sit down at a computer to deal with.

It’s custom.
It’s complicated.
It’s expensive.
It’s luxury software.

And it works because Dave understands the power and the cost of luxury software - he understands that the expensive part isn’t always the software. Sometimes the expensive part is dropping the ball.

But you do not need my setup. You don’t need Vapi. You do not need to build an agent. You don’t even need a Tesla.

You can get 70 percent of the interesting part with the phone already sitting in your pocket.

And once that happens, a bunch of places we used to think of as unusable suddenly become interfaces.

This is bigger than dictation

One of my friends, Tanay Kothari, co-founded Wispr Flow. If you haven’t used it, Flow takes the messy way people actually speak and turns it into clean text inside the apps they already use.

The interesting part isn’t just transcription quality. It’s what happens to behavior once speaking becomes good enough to replace typing.

Wispr currently says Flow can turn speech into text at roughly 220 words per minute versus about 45 words per minute on a keyboard. In March, the company said Flow had reached hundreds of thousands of daily active users and was growing around 40 percent month over month. Wispr has also reported that after six months, its average user produces nearly three-quarters of their characters through voice. These are Wispr’s own numbers, but the behavioral shift behind them is what interests me.

People aren’t just occasionally dictating an email. Their default input method is changing. That matters because keyboards created an invisible boundary around serious computer work.

A screenshot of my actual stats from Flow: (insane, I know)

You could think anywhere. But if you wanted to turn the thought into something useful, eventually you needed to sit down.

AI changes the second half of that equation.

Now you can speak badly. Rambling is fine. You can change direction halfway through a sentence. You can say, “Wait, that isn’t actually the problem.” The machine can follow the thread, ask the next question, bring in context, organize what happened, and leave you with something usable.

That’s not voice-to-text.

That’s my new interface for thinking.

You can build the cheap version today

The version we built for Dave is the Millennium Falcon.

Don’t start with the Millennium Falcon. Start with a bicycle.

For this tutorial, I’m going to use Claude because something important has changed: Claude Voice can now use the same model families available in text chat, switch between voice and text inside the same conversation, and use connected tools such as Gmail, Google Calendar, Google Docs, and Slack while you’re talking.

I still prefer ChatGPT for certain kinds of pure voice conversation. But there’s an important limitation as I write this: standalone ChatGPT Voice currently does not support connected apps. You can talk, use memory and web search where available, move between voice and text, and continue inside the same chat, but connected apps aren’t available inside standalone Voice yet.

So Claude currently gives us the cleaner path for the system I’m about to show you.

The basic version has four parts:

• A voice interface

• A model that understands how you think

• Access to the right pieces of your context

• Very clear boundaries around what it can do

That last one matters.

I’m building a command center. Not a digital chauffeur with access to my bank account.

Build the shell

Open Claude and create a new Project.

Projects are now available across Claude plans, including free accounts. Free users can currently create up to five.

Call yours whatever feels natural:

• Personal Command Center

• Drive Mode

• Thinking Partner

• The Cockpit

• Rolling Brain

The name matters less than you think.

In fact, Claude’s own documentation points out something slightly weird: Claude does not automatically receive the Project name or description as context. Anything you actually want the model to know needs to live in the Project Instructions or knowledge base

Create the Project.

For now, leave the Project Instructions blank.

Most people do the opposite. They start by dumping a giant autobiography into that box. That usually gives you a bloated pile of biography, temporary problems, preferences you don’t actually care about, and contradictions you forgot you wrote.

I’d rather have the machine interview me first.

Teach it how you think

Your job title is almost useless context. Knowing that I’m an entrepreneur or creative strategist does not tell a model how I actually make decisions.

It doesn’t know that I tend to think out loud. It doesn’t know that I often discover the real problem halfway through explaining the surface-level one. It doesn’t know whether I want reassurance or an argument. It doesn’t know which problems in my life are permanent and which ones are just this week. It doesn’t know when I want a sharp answer, when I want a sparring partner, and when I need it to shut up and let me finish the thought.

So we teach it.

Start a new chat inside your Personal Command Center Project and paste this:

You are the architect of my Personal Command Center.

This will become a voice-first AI workspace that I use from my phone and car to think, plan, prepare for conversations, make decisions, capture ideas, and turn raw thoughts into useful work.

Your first job is not to advise me.

Your first job is to interview me deeply enough to understand how I work, think, communicate, and make decisions.

## Interview behavior

- Use the AskUserQuestion tool when it is available.
- In regular Claude chat or voice mode, ask me one question at a time and wait for my complete answer.
- Ask between 12 and 18 primary questions.
- Do not show me the entire question list in advance.
- Start with easy factual questions, then move toward patterns, contradictions, preferences, fears, blind spots, and examples.
- Borrow the mechanics of strong long-form interviewers such as Katie Couric, Howard Stern, and Joe Rogan. Use warm entry, close listening, precise follow-ups, and curiosity about contradictions. Do not imitate anyone’s personality or speaking style.
- Ask for specific examples whenever my answer is abstract.
- Follow interesting threads instead of rigidly completing a questionnaire.
- Distinguish between stable traits and temporary circumstances.
- Do not turn the interview into therapy.
- Do not flatter me or tell me that every answer is insightful.
- After every four or five questions, briefly summarize what you believe you have learned. Ask me what you misunderstood or missed.
- Preserve memorable phrases I use naturally.
- Do not write the final instructions until I say: “Build my command center.”

## Topics to understand

Learn enough to understand:

- The roles I play in my professional and personal life
- The work I am responsible for each week
- The decisions that consume most of my attention
- My current goals, commitments, and pressures
- How I naturally solve problems
- How I respond to uncertainty
- Whether I think best through talking, writing, examples, debate, diagrams, or data
- What makes advice feel useful or useless to me
- How directly I want to be challenged
- The mistakes or avoidance patterns I repeat
- How I communicate with clients, coworkers, family, and collaborators
- My preferred level of detail
- My preferred tone and language
- The tools where my calendar, meetings, documents, and tasks live
- The information I consider sensitive
- The actions an AI should never perform without explicit confirmation
- What I want this system to help me accomplish during a typical drive

## Final deliverables

After I say “Build my command center,” produce the following:

### Cognitive profile

Write a concise profile explaining:

- How I appear to think
- How I make decisions
- Where I create the most value
- Where I may get stuck
- How an AI partner can be most useful to me

Do not diagnose me. Do not invent facts.

### Project Instructions

Create one clean markdown block that I can paste directly into Claude Project Instructions.

Keep the instructions under 5,000 characters.

The Project Instructions must define:

- Claude’s role inside this Project
- My relevant context
- My goals for the Project
- How Claude should ask questions
- When Claude should challenge me
- How Claude should handle uncertainty
- How Claude should preserve my voice
- My preferred response length and structure
- How Claude should behave during voice conversations
- Which actions require explicit approval
- Which actions Claude must never take
- How Claude should separate facts, assumptions, and recommendations
- How Claude should maintain and update durable context

Include a dedicated Drive Mode.

In Drive Mode:

- Keep spoken responses under 90 seconds unless I request more depth.
- Lead with the most useful point.
- Ask one question at a time.
- Never ask me to read, tap, compare visual options, or look at the screen while driving.
- Do not overwhelm me with long lists.
- Allow me to finish long thoughts without interrupting.
- When I am thinking out loud, do not rush to solve the problem.
- Summarize important decisions and ideas for review after the drive.
- Require explicit confirmation before using any tool that changes, sends, deletes, purchases, publishes, or schedules something.
- When a task needs visual attention, capture it for later rather than walking me through it while I am driving.

### Starter commands

Give me 10 short voice commands I can use immediately inside the Project.

Include commands for:

- Starting Drive Mode
- Receiving a daily brief
- Preparing for a meeting
- Thinking through a decision
- Capturing an idea
- Developing an article
- Reviewing commitments
- Ending a drive
- Producing a post-drive summary
- Updating my Project context

### Assumption check

List three things Claude still does not know about me and should never assume.

Don’t do this interview while driving. Do it at home. Take your time.

Correct the model when it gets you wrong.

Those corrections are often more valuable than the original answers because you’re teaching the system the difference between something that sounds plausible and something that is actually true about you.

When you’re done, say:

Build my command center.

Claude will produce your Project Instructions.

Copy them. Open the Project.

Select Set project instructions.

Paste them in.

Save.

Claude also lets you add files and other material to the Project knowledgebase, and those materials are available across chats in that Project.

Even with newer memory features arriving in Claude, I still treat important context like infrastructure.

If I need the system to know something tomorrow, I put it somewhere durable.

I don’t cross my fingers and hope it remembers.

Turn on Drive Mode

Do this before the car moves.

Open your Project and select the model you want.

For most normal drives, I use Sonnet. It’s fast enough to feel conversational and capable enough for serious work. For a hard strategic problem, delicate negotiation, or deeper creative session, I’ll sometimes move up to Opus.

Claude Voice now uses the same model families available in text chat, and you can change models during a voice conversation.

Then tap the waveform icon.

Claude supports both hands-free and push-to-talk voice modes. Hands-free listens for natural pauses; push-to-talk is useful when the environment is noisy.

Because you already built the behavior into your Project Instructions, you shouldn’t have to repeat a giant prompt every time you get in the car.

Just give it the context it needs and start talking:

I'm driving. Let's use Drive Mode. Here's what's on my mind...

That’s it.

The Project Instructions should already tell Claude to keep responses short, let you think out loud, ask one question at a time, avoid pushing you toward the screen, and get explicit approval before taking external actions.

If you find yourself having to repeat those rules every drive, don’t create a longer voice prompt. Go back and fix the Project Instructions. The goal is for this to eventually feel like calling someone who already knows how to work with you.

Claude can carry the same conversation between voice and text, so when the drive ends you can open the thread and keep working without starting over.

That matters more than it sounds. Your drive can be where you discover the thing. Your desk can be where you finish it.

One warning because apparently we still need warnings about two-ton pieces of metal moving at 70 miles an hour:

Hands-free does not mean distraction-free.

NHTSA says 3,208 people were killed in crashes involving distracted drivers in 2024.

Set the system up before moving.

Don’t review spreadsheets in traffic.

Don’t start an emotionally nuclear conversation while merging onto I-95.

Don’t be an idiot.

Five ways I use my mobile command center

Once the system knows you, the magic isn’t giant prompts.

It’s having a few reliable openings that put you into the right kind of conversation.

The morning runway

Before the first meeting:

Look at my calendar for today.

Give me a 90-second brief covering the three moments that deserve the most preparation, any conflicts or unrealistic transitions, and the one outcome that would make today successful.

Do not read my entire calendar back to me.

This is much more useful than asking: “What do I have today?”

The calendar already knows what exists. I want the system to make a judgment about what matters.

The thinking drive

When something is bothering me:

I need to think out loud.

Do not solve the problem yet.

Listen for assumptions, contradictions, emotional reactions, and parts of the situation I may be avoiding. Reflect back what you hear and ask one useful question at a time.

This is probably the closest I get to using AI like a therapist, although I don’t think AI is a therapist.

I use it as a mirror.

Most AI systems answer too quickly. They hear version one of your problem and race toward a solution. People do this too. The first version of the problem is often the polite version.

The real problem shows up around minute seven.

The meeting approach

On the way to an important conversation:

I am preparing for my meeting with [person or company].

Use my calendar, connected meeting notes, and Project knowledge.

Remind me what they care about, what I committed to, what remains unresolved, and what I should not forget to ask.

End with a 30-second opening I can use to start the meeting.

This is where voice stops being a novelty.

I’m not asking a chatbot for generic negotiation advice.

I’m trying to walk into the room with my own memory returned to me at the exact moment I need it.

NOTE: Voice mode cannot access connectors at this time. For setups like mine & Dave’s these are custom apps.

The decision room

When I’m stuck between options:

Help me think through this decision.

First, tell me what decision you believe I am actually making.

Then identify the assumptions behind each option, the likely second-order consequences, and what information would change the recommendation.

Do not recommend anything until I confirm that you have framed the decision correctly.

That first instruction is the trick. A lot of bad decisions happen because we spend all our energy solving the option set in front of us without asking who created the option set.

Sometimes the most useful thing AI can do is tell you you’re answering the wrong question.

The shutdown drive

At the end of the day:

Help me close the day.

Review the commitments, decisions, and ideas that came up in our conversations today.

Separate them into:

What I completed  
What I owe someone  
What I am waiting on  
What deserves deeper thought  
What can be discarded

Keep the spoken version short. Save the detailed summary for me to read later.

This isn’t about squeezing another eight units of productivity out of the drive home.

It’s about arriving home without carrying twelve unfinished conversations around in your skull.

Add memory without handing it the keys

A generic model can help you think. A connected model can help you remember.

That’s where this becomes dramatically more useful. It’s also where it becomes more dangerous.

My rule is:

Context in. Consequences out.

I am much more comfortable giving an AI selected information that helps it recall, prepare, compare, and draft than I am giving it the ability to send, delete, purchase, publish, or move money.

Start boring.

Calendar first

Calendar access lets the system understand where your attention is supposed to go.

It can prepare you for the day, identify unrealistic schedules, find open thinking time, and remind you about upcoming conversations.

Claude’s Google Calendar connector can also create, update, and delete events, so set a stricter rule inside your Project Instructions:

You may read and analyze my calendar.

Before creating, changing, deleting, accepting, or declining any calendar event, repeat the exact proposed action and wait for my explicit approval.

The redundant approval is intentional.

I like seat belts and airbags too.

Meeting memory second

This one changed the usefulness of the system for me.

I use Fathom as meeting memory.

Fathom now has an MCP integration specifically for tools like Claude and ChatGPT. It lets those assistants search meetings you already have access to and use that context to create things like follow-ups, briefs, plans, and summaries without manually copying transcripts around.

Granola offers a similar MCP connection for meeting notes and transcripts, with access depending on your plan and workspace configuration.

Once your meetings become searchable context, you can ask questions like:

• What did I promise this client?

• How did they describe the problem in their own words?

• Which objections keep showing up across sales conversations?

• What ideas have I repeated in three meetings but never written about?

• Where did somebody react strongly to this idea?

That last category is gold.

Your meetings contain years of language you’ve already paid for with your time.

Most of us let it rot in transcript graveyards.

Use whichever meeting system you already trust.

I don’t care who wins the notetaker feature war.

I care whether the conversations become usable.

Documents third

Connect the place where your current plans, proposals, drafts, and research live.

For a lot of people, that’s Google Drive.

Claude’s current Google Drive connector can search and retrieve files, read common document formats, add Drive files to Projects, and save generated files back to Drive when file creation is enabled.

Do not connect your entire digital life on day one.

Start with one folder.

Imagine you hired a thoughtful human assistant this morning.

What folder would you comfortably hand them first?

Start there.

Email last

Email is valuable context wrapped in radioactive material.

It contains:

Private information.
Automated garbage.
Emotional landmines.
Requests designed to hijack your priorities.
Old conversations you forgot existed.

And at least one newsletter you swear you unsubscribed from four times.

So I connect email last.

Claude’s Gmail connector can currently search and read messages, work with labels and threads, and create drafts in Gmail. It cannot send those Gmail drafts on your behalf.

I like that boundary.

Let the machine find the thread.
Let it remember what happened.
Let it prepare the response.
Let the human decide what leaves the building.

I use Superhuman for my email and disable the ability for it to send emails (it can only draft). AI makes mistakes. Plan accordingly.

The guardrails I would not skip

Before I connect personal or company systems, I put something like this into the Project Instructions:

## Action boundaries

You may retrieve, summarize, compare, organize, and draft.

You must receive my explicit approval before you:

- Send or publish anything
- Create, change, or delete a calendar event
- Contact another person
- Delete or move a file
- Change a system of record
- Make a purchase
- Move money
- Submit a form
- Accept terms
- Share private information
- Connect a new service

Never request, store, repeat, or expose passwords, authentication codes, private keys, banking credentials, or account recovery information.

When passengers are present, do not read private emails, confidential meeting details, financial information, health information, or personnel matters aloud.

When uncertain, stop and ask.

I’d also spend five minutes looking at the privacy controls for whichever plan you use.

Anthropic says inputs and outputs from commercial products such as Claude for Work are not used for model training by default. Consumer accounts such as Free, Pro, and Max have separate privacy controls that let users choose whether chats and coding sessions may be used to improve Claude.

And remember that third-party connectors have their own terms and security practices.

The AI might have a beautiful voice.

It is still a new assistant.

Don’t hand the new assistant every key you own on Monday morning.

NOTE: When building agents, having deterministic blockers and hooks is critical.

Turn your car into a content studio

This is where it gets fun.

Most of my best content doesn’t begin with writing.

It starts with me noticing something. Getting annoyed by something. Hearing somebody say a phrase in a meeting. Seeing two things that shouldn’t be connected and wondering if maybe they are.

The car is where those half-formed ideas start talking.

The mistake is asking AI to turn the first thought into a LinkedIn post.

That kills it.

You compress the idea before the idea has earned its shape. You get a competent pile of sentences where there might have been something sharp.

I don’t want AI to write for me. I want it to help me see better before I write.

So I separate the work into stages.

Raw Capture

Use this when you have the beginning of a thought but don’t understand it yet.

Enter Raw Capture.

I am going to speak freely about an idea.

Do not interrupt unless I become completely stuck. Do not outline, summarize, improve, or write content while I am talking.

Listen for:

- Phrases that sound distinctly like me
- Moments where my energy changes
- Stories or examples I mention
- Questions I keep circling
- Claims that need evidence
- Contradictions that may contain the real idea

Wait until I say “Cut the tape.”

Then return:

- A clean transcript
- My five strongest verbatim lines
- The question underneath the idea
- The tension or contradiction
- What still feels generic
- Three directions worth exploring

Do not write the article yet.

Lens Finder

Once you have raw material:

Run Lens Finder on the Raw Tape.

Find the perspective shift that would cause a smart reader to see the problem differently.

Identify:

- The common interpretation
- Why that interpretation is incomplete
- The hidden assumption
- The new lens
- The five-second moment where the idea becomes visible
- The consequence for a founder, operator, or creator
- The action the reader should take differently

Give me three possible lenses.

Make them meaningfully different. Do not give me three rewrites of the same idea.

Story Miner

Once your meeting memory is connected:

Search my connected meetings and Project knowledge for true moments that could support this idea.

Look for:

- A client describing the problem in their own words
- A moment when I changed my mind
- A mistake, conflict, or surprising result
- A repeated pattern across several conversations
- A concrete scene with people, pressure, and consequences

Do not invent dialogue or combine separate events into one story.

For each possible story, provide:

- What happened
- When it happened
- Why it supports the argument
- The exact source
- What details still need verification

This is where accumulated experience starts compounding.

Your meeting archive becomes a private library of stories, language, objections, arguments, failures, and moments when somebody’s energy changed.

That is a much more interesting database than “meeting transcripts.”

Argument Room

Do this before you draft.

Enter Argument Room.

Treat this idea as if it must survive an intelligent, skeptical reader.

Test:

- Whether the central claim is actually new
- Which claims are observations versus facts
- What evidence is missing
- Where I am overstating the case
- What a strong critic would say
- Whether the conclusion follows from the story
- Which part sounds clever but says very little
- What practical value the reader receives

Do not rewrite the idea to make it safer.

Help me make it truer.

End the drive

At the end of any creative session, say:

End the drive.

Create a post-drive work packet containing:

- The cleaned transcript
- My strongest verbatim phrases
- The central lens
- The proposed article thesis
- A narrative outline
- Stories or examples to investigate
- Claims requiring verification
- Questions I still need to answer
- The single best next action

Do not write the final draft unless I explicitly ask.

That last instruction protects the idea from premature polish.

I want the machine helping me excavate before it starts building.

Turn the workflows into real Skills

You can run every prompt above manually.

Once you find yourself repeating one, package it as a named Skill.

Claude’s custom Skills are now available across Free, Pro, Max, Team, and Enterprise plans when code execution is enabled. A Skill can contain a few lines of instructions or a larger package with files, examples, and executable workflows. start with four:

• /raw-tape

• /lens-finder

• /story-miner

• /argument-room

Do not build twenty Skills because Skills are exciting.

Build one when you have repeated the same process enough times to know what good looks like.

A Skill should capture judgment, not merely save keystrokes.

My actual stack

The basic version of this system requires almost nothing:

• A phone.

• Bluetooth or headphones.

• A Claude Project.

• A good set of Project Instructions.

That’s enough to discover whether voice actually changes the way you work.

The next version adds calendar, meeting memory, a focused document set, and the content workflows above.

My personal setup goes much further. It sits on something called an agent harness and has a bunch of third party connectors that gives it my full context.

My favorite pure conversational voice experience is still ChatGPT.

For custom systems, I also use Vapi to build voice agents that can follow specialized workflows, talk to my own systems, and do things the consumer products aren’t designed to do.

That’s the world Dave’s setup lives in.

I would not begin there.

Technical people have a terrible habit of building the Millennium Falcon before proving they need a bicycle.

Build the stupid-simple loop first:

Talk. Think. Capture. Review.

Use it for a week.

Notice where it helps. Notice where it interrupts you. Notice which questions change your thinking. Notice what you keep having to explain.

Then build around your actual behavior.

A note for the haters: This isn’t about me squeezing work out of every minute

There is a horrifying version of this article where every shower, walk, commute, meal, school pickup, and bathroom break becomes another opportunity to optimize your output. Sounds like a Black Mirror episode, doesn’t it?

Hard pass.

Some drives should have music. Some drives should have silence.

Some drives should contain my five-year-old explaining why Bass Pro Shops have the coolest trucks and boats.

My goal isn’t to remove every idle moment from your life. The lens shift I’m offering is much, much smaller than that.

We assume some periods of the day are “dead time” because the interface we associate with productive work isn’t available.

No keyboard. No giant screens. No desk.

So nothing important can happen.

But your brain never agreed to that arrangement.

Some of the places where we can’t type are exactly where our minds loosen up enough to think.

• Walking.

• Driving.

• Cooking.

• Getting coffee.

• Staring through a windshield after a hard meeting.

Then build an interface around that place for yourself.

For me, that place happens to have four wheels. My car used to be the gap between places where work happened. Now it’s sometimes where I prepare for the work.

Where I process what just happened.

Where I catch the commitment I would have forgotten.

Where I hear the sentence that becomes the article.

Where I find the real problem underneath the one I was busy trying to solve.

And occasionally, where I hear myself think for the first time all day.

You don’t need my command center to try this. On your next drive, don’t ask AI to write something. Or for it to generate ten new ideas. Just start with:

I need to think out loud. Don’t solve it too fast. Help me hear what I actually think.

If you want to go deeper, here are some links for you:

1. Claude Voice Mode — https://support.claude.com/en/articles/11101966-use-voice-mode

2. ChatGPT Apps — https://help.openai.com/en/articles/11487775-connectors-in-chatgpt

3. Claude Projects — https://support.claude.com/en/articles/9517075-what-are-projects

4. Claude Project setup — https://support.claude.com/en/articles/9519177-how-can-i-create-and-manage-projects

5. Claude Google Workspace connectors — https://support.claude.com/en/articles/10166901-use-google-workspace-connectors

6. Fathom MCP — https://help.fathom.video/en/articles/11497793

7. Granola MCP — https://docs.granola.ai/help-center/sharing/integrations/mcp

8. Claude Skills — https://support.claude.com/en/articles/12512180-use-skills-in-claude

9. Anthropic commercial model training policy — https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training

10. Anthropic consumer model-improvement controls — https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training

11. Wispr Flow speed claim — https://wisprflow.ai/use-cases/google-docs

12. Wispr Flow growth + usage — https://wisprflow.ai/post/the-master-plan

13. Wispr six-month voice usage — https://wisprflow.ai/new-funding

14. NHTSA distracted driving — https://www.nhtsa.gov/risky-driving/distracted-driving

I’m standing in my backyard, staring at the koi pond that came with my house.

I didn’t buy the pond. I bought the house. The pond was already there, sitting there in the backyard, looking expensive and serene and completely self-sufficient. At least that’s what I thought.

Then I learned that koi ponds have pumps, filters, chemistry, seasonal care, predators, algae, and a person who knows what all of those things mean. There is always a guy. The pond looked like a feature of the property. But what I actually inherited was a small living ecosystem with a maintenance schedule.

A friend of mine had a custom pizza oven built into his patio. Another bought the kind of car that makes grown men lower their voices when they say the model name out loud. Both got exactly what they wanted. They also got the special fuel, the unusual replacement parts, the weird service interval, and the guy you call when the thing you love starts making a noise you don’t understand.

This is the part we leave out when we talk about luxury. We talk about the first drive, first pizza, first summer night next to the pond. We don’t imagine ourselves three months later trying to figure out why the water looks strange and smells like a sewer.

Every luxury needs a mechanic.

And I think a lot of companies are accidentally filling themselves with luxury goods right now.

I love vibe coding. I’ve been building software since I was a kid, and the fact that I can describe an idea to Claude, ChatGPT, Cursor, or whatever comes next and watch a working version appear still feels a little bit like magic. The distance between imagining something and touching it on a screen has collapsed.

For builders and tinkerers, that’s intoxicating. For executives who have spent years trying to explain what they want to product teams, it’s even more powerful. They can finally shape the software around the way they think instead of learning to think like the software.

You can build the dashboard your vendor refuses to give you. You can create a personal CRM that remembers who you met, why they mattered, and the thing you promised to send three weeks ago. You can make a digital chief of staff that lives across your calendar, documents, messages, customer data, and the open loops that normally stay trapped inside your head.

You can build software for one person. And that person can be you.

There’s something beautiful about designing for one. Most commercial software has to survive roadmaps, customer segments, support tickets, compliance reviews, quarterly targets, and the competing needs of thousands or millions of users. Personal software can be strange. It can be particular. It can use your language, follow your habits, and arrange information in the exact order your brain wants it.

That’s why some of the best products started as a founder scratching their own itch. Someone had a problem. The existing tools weren’t good enough. So they built something for themselves and eventually discovered that plenty of other people had the same pain.

This game isn’t new, but the number of people playing it are first-timers.

You don’t need to be a traditional software engineer to make the first version anymore. But you do need to have enough taste to know what good looks like, enough clarity to describe what you want, and enough patience to keep shaping the output when the computer gives you something that looks like your cat threw up on the screen.

The internet is obsessed with the build. People post videos of apps they made over a weekend. Founders show off internal dashboards. Executives start creating little agents between meetings and hand them to their teams on Monday like they came home from Home Depot with a homemade espresso machine.

Sometimes the thing is genuinely good. That’s important. A sharp person inside a business can build a tool that better fits the company than a platform charging $30,000 a year, because that person actually understands the work.

Then something funny happens.

Nobody cancels the vendor.

The invoice still clears every month. The contract auto-renews in the spring. Finance approves it because it looks like infrastructure.

The internal tool stays too. Four people start depending on it because it works better. But only one person knows how it works, why it works, where the keys are, which assumptions are buried in it, and what to do when the system on the other end changes.

So the company ends up paying twice and depending on one.

That’s the position I keep seeing, and most organizations don’t realize they’re in it because both halves look like wins. The enterprise platform looks safe. The internal tool looks resourceful. Together they create duplicate cost and concentrated risk.

You probably already have one of these inside your company.

The handoff is where the truth shows up.

The person who built the shiny new tool has spent forty hours arguing it into existence. They understand every quirk because they created the quirks. Then they pass it to the team, and the team inherits a system nobody understands wtf is going on.

There’s no documentation because the maker never needed any (and probably couldn’t create it if they tried). There are no tests because it worked when they ran it. There’s a hardcoded API key buried somewhere anyone can use to run up your AI bill, and an assumption about the sales pipeline that stopped being true back in March.

The software rarely breaks in a cinematic way. It degrades one day at a time.

A model gets retired. A connector changes. An authentication flow gets updated. A field inside HubSpot is renamed and the chain starts passing empty values and erroring silently. One Monday the report comes out wrong, but it looks exactly as right as it did the Monday before. It looks the same but there’s no sign of truth to be found.

Six weeks later, the team has stopped using it. And nobody tells the executive because they love their prized creation, they whipped up between bourbon old-fashioneds.

That doesn’t make vibe coding bad. It means we need a better category for what it is.

Vibe coding is a whiteboard that runs.

It’s a wireframe that can answer back. It’s a sketch you can click. It lets someone with taste make an idea tangible enough for other people to react to, and that is incredibly valuable because most people are awful at responding to abstractions.

Show them a deck and they nod.

Show them a working interface and suddenly they have seventeen opinions.

Good. Now you have something to work with.

Treat vibe coding as the fastest thinking tool we’ve ever had and it becomes a superpower. You can skip weeks of vague debate and put a working idea in front of someone by the afternoon.

Treat the same prototype as production software and you’ve committed the company to maintaining a document of one person’s Tuesday.

The prototype answers one question: could this exist?

A product has to answer much harder questions.

Should this exist?

Can people trust it?

Who owns it?

What happens when the data is wrong?

What happens when the person who built it leaves?

What happens when fifty people use it in ways the original builder never imagined?

Personal software becomes organizational software the moment another person starts depending on it.

That transition is where most of the cost (and pain) lives.

I’ve spent twenty-three years watching software move from idea to reality. First, I built things for brands. Then I invested in early-stage technology companies and helped shape products that had to survive the jump from founder instinct to real paying customers. Now I work with founders, executives, and creative leaders who are among the best in the world at what they do.

They know what information matters. They understand the nuance inside their business. They can look at a dashboard and immediately see what is missing. They know which question opens the real problem and which answer is technically correct but practically useless.

For most of their careers, they had to squeeze that judgment into someone else’s software. Now they can build around it or on top of it.

That’s the part of personal software I find genuinely exciting. We’ve entered a world where the best people in a field can begin encoding more of their taste, judgment, memory, and decision-making into the systems around them.

A founder can have a digital chief of staff who follows them across the tools they already use. They can be driving, walking through an airport, or sitting in the five-minute gap between two meetings and talk through what’s happening. The system can surface the person they forgot to call, prepare context for the next meeting, dispatch an action to someone on the team, or explain what changed in the business without making them click through six dashboards and a spreadsheet called FINAL_v7_USE_THIS_ONE.xls

A creative leader can have a portable thinking environment. An idea arrives in the car or in bed at midnight. They capture it in their own language. The system connects it to earlier ideas, client context, research, and work already in motion.

That’s more than automation. It’s personal infrastructure.

And infrastructure changes the standard.

Software can be disposable. Infrastructure has to be dependable. Infrastructure needs ownership, uptime, documentation, permissions, monitoring, and a budget. The moment your custom tool becomes part of how decisions get made, customers get served, or money moves through the business, it stops being a clever experiment.

It becomes part of the company.

This is why I’m less interested in whether an AI system looks impressive and more interested in three brutally practical questions.

Does it save real time or money you can point at?
Does it create customers, margin, capacity, or better decisions you can trace?
And can you trust it tonight, when the person who built it isn’t in the room to check it?

Most vibe-coded internal tools pass the first question, ignore the second, and fail the third without telling anyone.

Trust is the killer. A system can save time and still be useless because nobody believes the output. It can generate beautiful work and create more anxiety than value because no one knows where the answer came from. It can be right ninety-five percent of the time and still be catastrophic when the missing five percent involves money, legal exposure, customer promises, or someone’s career.

The number of apps shipped is a lousy measure of progress. The real test comes after the prototype, when somebody has to own it, fix it and pay for it. Smart companies will make the hard call early. Some ideas deserve a real product. Others should stay sketches. And plenty were spreadsheets all along.

They’ll design the mechanic at the same time they design the luxury.

That mechanic may be an internal person. It may be a small team. It may be a partner who watches the models, maintains the integrations, checks the outputs, documents the system, and helps it evolve as the company changes. We run a company that does this, so if you need it, DM me.

The form matters less than the accountability.

Someone has to own the pond. Fix the pizza oven and service the mid-life crisis mobile.. cough.. I mean sports car.

This doesn’t mean HubSpot, Salesforce, Google, Microsoft, and the rest of the software world disappear. Their systems still hold the records. They manage permissions, history, workflows, security, and all the boring infrastructure that becomes very exciting the first time it stops working.

The shift is that those platforms may stop being the place where we spend most of our time.

For twenty years, humans adapted themselves to software. We learned the menus, fields, workflows, and strange internal logic of tools built to serve everybody.

Now software is beginning to adapt itself to us.

Software now lives where we work and how we work

You may eventually interact with one familiar interface that understands you, then talks to dozens of systems behind the scenes. You ask for what you need in your own language. The system translates your intent into the rigid structures the underlying software requires.

The interface becomes personal. The infrastructure remains shared.

For years, companies bought giant software platforms and used a fraction of the features. Now we may build dozens of tiny custom tools because they’re cheap and easy to create, then discover we use a fraction of those too.

The cost moves. It doesn’t disappear.

Instead of paying for bloated software, we start paying for abandoned software. Half-finished agents. Broken connectors. Personal dashboards that were exciting for three weeks and now send error emails to an inbox nobody checks.

We trade feature debt for maintenance debt.

I’m not telling anyone to stop building. I love this stuff. I build these systems for a living, and I’m not slowing down.

Personal software is absolutely the future. For the right person, built the right way, it can create an advantage that off-the-shelf software never could. It can help a founder see the business more clearly. It can give an executive access to the context they normally lose between meetings. It can make the taste and judgment of someone exceptional more available to the people around them.

This stuff is a new layer of business. And the mistake I keep seeing play out in the boardrooms and C-suite meetings I’m in is buying the fantasy without budgeting for the reality.

So the next time a software renewal crosses your desk, ask one question before you approve it:

Who here has already built a version of this?

Sometimes the answer will be nobody. Sign it and move on. Sometimes the answer will be a name.

Now you know what you may be paying for twice. More importantly, you know who you may be depending on alone.

Then ask the questions people usually wait too long to ask.

Who understands it?
Who checks it?
Who maintains it when a model changes?
What happens the week that person is gone?

Those questions aren’t meant to kill the experiment or rain on your innovation party. They’re how these experiments earn the right to become infrastructure you grow your company on.

I still love my koi pond. It’s beautiful. I’m glad it came with the house.

I just understand what I own now.

The fish are the visible part. The system is what keeps them alive. Thanks, Bob (my fish guy).

Your company is building a lot of beautiful new things right now. Tools shaped like the judgment of one sharp person, running on models that shift underneath them.

Someone in your family dies.

The world doesn’t skip a beat. It just hands you a checklist of new shit to do.

Flights. Family calls. Death certificates. Dates. The grief is real, but the logistics get there first. You move through all of it in a fog because somebody has to.

Airline policy is the last thing you have the energy to think about.

So when Jake Moffatt’s grandmother died, he did what millions of us now do.

He asked the system.

He opened Air Canada’s website. The fancy AI chatbot told him he could book a flight, travel, and apply for a bereavement discount within ninety days. The answer was calm and more confident than a tech bro after three espressos at Philz.

It appeared on the company’s website. Why wouldn’t he trust it?

Moffatt booked the ticket and took the trip. Air Canada later refused the refund.

Then the story got weird.

Air Canada argued that it should not be responsible for the chatbot’s answer. The British Columbia Civil Resolution Tribunal said the argument amounted to treating the chatbot like a separate legal entity responsible for its own words.

The tribunal wasn’t buying it.

The chatbot was sitting on Air Canada’s website, speaking to an Air Canada customer about an Air Canada policy. The company owned the answer, even though the answer was wrong.

Moffatt received 650.88 CAD in damages, plus interest and fees. The money barely matters. The useful part of the case is the sentence now slipping into meetings everywhere:

Those four words sound harmless. They describe where an idea came from. They also move responsibility out of the room.

Used often enough, they become an alibi.

People are making mistakes left and right because we’re getting intellectually lazy. The machine gives us something clean, confident and plausible. We move it into the next meeting before anybody has wrestled with the thought.

I think we’re becoming intellectually bulimic.

We binge on information, skip the digestion and throw up bad thoughts at machine speed. Board decks. Pricing decisions. Performance reviews. Strategy documents. Half-digested ideas wearing nice fonts.

AI didn’t invent bad thinking. It put bad thinking on an assembly line. And it’s not a bug, it’s actually a feature of how the systems are designed. Wanna nerd out on how LLMS works? Here you go

The alibi that lets everyone walk away clean

Inside a company, responsibility rarely disappears all at once.

A model flags a high-performing employee as underperforming because someone used a free HR skill they downloaded on GitHub without actually testing it first. A pricing system suggests a 7% increase. An AI agent drafts the board report and wraps the numbers in charts clean enough to survive a skim, but it made up all the numbers.

Work speeds up. Meetings shrink. The recommendation feels more credible because it arrives formatted and scored.

Then the language softens.

“The system recommended it.”

“The model flagged it.”

“The AI suggested it.”

That wording turns a judgment call into something that feels technical and inevitable. Numbers create distance. A confidence score makes a choice look like a finding. Following the recommendation feels safer than owning the decision.

Then something breaks.

A strong employee gets fired. A loyal customer leaves after the 7% increase. The board report contains a number that looked plausible and turned out to be wrong.

Somebody asks the question everyone should have asked earlier: Who made the call?

By then, ownership has gone missing. People reviewed the output. Someone clicked approve. The work shipped. Nobody can explain who truly decided.

“Nobody” works fine until accountability shows up.

A human in the loop can still be asleep at the wheel

Most companies believe they have covered this problem with a human reviewer. Their checklist asks whether a person looked at the output before it went live.

Someone did. Great.

Looking is cheap. Ownership costs something.

Ask yourself: Which human actually owns the judgment?

That person should be able to explain the decision in plain language. They should know what the system produced, where it could be wrong and why they trusted it in this case. They also need enough authority to pause the work or reject the answer.

A warm body next to an approve button gives you ceremony. It doesn’t give you ownership.

This is where accountability and traceability meet. Accountability puts a name on the call. Traceability lets that person show the source, the AI output, the human judgment and the reason the work shipped.

A wrong answer can be corrected. A decision with no chain of custody becomes a board problem, an audit problem, or a lawsuit.

Sometimes you get the combo meal.

AI makes weak thinking look finished

The alibi usually starts with convenience.

“The AI drafted it.”

A proposal that took three hours now takes five minutes. A dense report becomes a clean summary that skips the meat that only a human brain with real wisdom can see. A junior employee produces something that looks senior enough to pass at a glance.

That last part is dangerous. When an output looks finished, people treat it as finished.

You skim language you didn’t write. You trust a chart you didn’t build. The model’s confidence starts replacing your own scrutiny.

Spreadsheet mistakes often look suspicious.

They have headings, clean labels and a chart that seems to know what it’s talking about. Polish becomes camouflage.

We’ve been doing a lot of work lately helping finance and business teams adopt AI to reimagine their work. One thing people don’t understand about finance is this:

If you’re 95% right, you’re 100% out of a job.

I recently read that at one company (not my clients), AI-generated board reports carried incorrect data for an entire quarter. The reports looked finished. The charts were clean. The analytics sounded confident.

One small problem. The numbers were absolutely wrong.

A pretty chart with a bad number is just a confident lie in brand colors.

Once that number lands in a board deck, nobody cares that the model hallucinated. They want to know where the number came from, who checked it and why it stayed wrong for three months.

Nobody has ever saved a board meeting by saying, “To be fair, the AI sounded extremely confident.”

The model produced the error. The workflow waved it through.

Everybody touched it. Nobody owned it.

The decision exists. The trail is missing.

A June 2025 Resume Builder survey asked 1,342 U.S. managers how they were using AI. Among managers who used it for people decisions, 78% used it for raises, 77% for promotions, 66% for layoffs and 64% for terminations.

More than one in five said they often or always allowed AI to make final decisions without human input. Only 32% had received formal training.

This was a commercial survey, not a peer-reviewed study. I wouldn’t treat it like scripture. I would treat it like a smoke alarm.

General-purpose AI tools are already sitting inside decisions that change people’s lives. Most companies haven’t built a decision trail to match.

That gap landed in federal court this month.

Twenty-six Meta employees alleged that internal AI-assisted systems helped select workers for layoffs and disadvantaged people who had taken medical or family leave. Meta denied the claim and said humans made every layoff decision.

On July 17, 2026, a federal judge declined to pause the layoffs. He said the plaintiffs lacked evidence showing whether and how AI had been used improperly. Reuters captured the bind on July 22. The workers were not in the rooms where the decisions happened. Their lawyers said Meta held nearly all the relevant information.

The company says humans decided. The people affected can’t see how. The evidence lives inside systems they don’t control.

The decision exists. The trail is missing.

The machine has better paperwork

There is another reason people follow the model. It shows up with receipts.

Your best operator has spent fifteen years learning which deals go bad. She hears a pitch and catches something small. The buyer’s urgency doesn’t fit the budget. The room goes quiet at the wrong moment. Something smells off.

She says, “I don’t like this deal.”

The model arrives with a risk score, contributing factors, a confidence interval, and a recommendation.

One person has fifteen years of scars. The machine has a deck.

A 2024 working paper by Hossein Nikpayam, Mirko Kremer and Francis de Véricourt found that people making delegated decisions were more likely to follow an algorithm to protect themselves from blame. As of June 2026, the paper was still under review at Management Science. That makes it a useful signal, not settled science.

The behavior will still feel familiar to anyone who has survived a corporate meeting. A bad decision with documentation is easier to defend than a good decision built on judgment you can’t explain.

The machine leaves timestamps, scores and neat little boxes. Experienced human judgment often lives inside someone’s head.

The model wins because it filed the paperwork.

Companies need to capture the human judgment too. Ask the operator what felt wrong. Write down the pattern she recognized. Record why she overrode the score.

Otherwise, your best judgment disappears while the machine’s worst judgment gets archived.

Someone still absorbs the impact

Researcher Madeleine Clare Elish coined the term “moral crumple zone” for the human who absorbs blame when an automated system fails, even though that person had limited control over it.

You can see how this plays out inside a company.

A senior leader approves the system. An analyst uses it inside a workflow they didn’t design. The system helps produce a bad decision. The analyst clicked approve, so the blame rolls downhill.

The executive keeps the alibi. The analyst gets the wreckage.

Putting a name on a decision works only when that person has enough context to understand it and enough power to reject it.

A name without authority is theater.

If you touch it with AI, you own it

I build AI systems for a living. I want teams using them. I want the hours back and the better work that can come with them.

I also know what happens when speed outruns ownership.

At my companies, we introduced a simple rule:

If you touch it with AI, you own it.

The person using AI to write code owned the code. The person building a prototype owned the prototype. Anyone carrying AI-generated analysis into a meeting owned every number on the page.

The rule killed the ghost employee in the room.

“AI” couldn’t own the code. It couldn’t own the forecast. It couldn’t sit in the boardroom and defend a recommendation.

The person carrying the work had to do that.

People slowed down just enough to check assumptions and challenge weak outputs. They stopped treating the model’s answer as pre-approved.

Quality went up because scrutiny did.

Review is a task. Ownership follows you into the room after the decision goes bad.

NIST’s AI Risk Management Framework uses formal language about roles, oversight and accountability. The operating rule can stay plain: put one name on the decision.

Run the five-minute replay test

Take one meaningful decision your company made with AI in the room. Use a hire, a forecast, a pricing change or a board recommendation.

Put this block at the top of the document:

Decision:
AI contribution:
Source data:
Human judgment:
Owner:
Reason for approval:
Override trigger:

Ask the owner to replay the decision in five minutes.

Can they show where the information came from? Can they explain what the AI did? Can they point to the moment human judgment entered the process? Can they tell you why they signed off?

Use a person’s name. “Leadership” can’t sit in front of the board. Neither can “marketing,” “finance” or “the AI team.”

Every blank field is a hole you’ll wish you had filled when the number gets challenged.

Three prompts that catch bad AI work before it ships

No prompt makes a model truthful. These three add friction at the moment clean-looking work tries to outrun judgment. Each comes from a method tested in peer-reviewed machine learning research.

The chain-of-verification prompt

Chain of Verification, usually shortened to CoVe, forces the model to draft verification questions,, answer them independently and revise the original response. Researchers reported fewer hallucinations across list questions, closed-book question answering and long-form generation.

Use this after the first answer. A fresh chat is better because the model is less likely to defend its original wording.

Audit the answer below using Chain of Verification.

1. Extract every factual claim, number, date, named entity and causal claim.
2. Write the questions that would expose each claim if it were wrong.
3. Answer those verification questions independently. Do not use the original answer as evidence.
4. Check primary sources when browsing is available. Link each source.
5. Label every original claim VERIFIED, CONTRADICTED, UNSUPPORTED or UNCLEAR.
6. Rewrite the answer using only verified claims. Mark anything that still needs human review.

Answer to verify:
[PASTE THE ANSWER]

The important move is independence. “Check your work” often produces a confident rereading of the same mistake. CoVe makes the model build questions that can prove itself wrong.

The cite-it-or-strike-it prompt

RARR stands for Research and Revision. The method searches for evidence, adds attribution and edits unsupported claims while trying to preserve the original answer.

This is the one I would run on a report, strategy memo or article before it leaves the building.

Build a source ledger for the draft below.

Break the draft into atomic factual claims. For each claim, return a table with:

Claim | Best primary source | Exact support | Status | Required change

Use VERIFIED, PARTIAL, CONTRADICTED or UNSUPPORTED for status.

A citation counts only when the source supports the exact claim. Do not use search snippets, another AI summary or a source that merely discusses the same topic.

After the table, revise the draft. Keep verified claims. Fix partial or contradicted claims. Delete unsupported claims unless they are clearly labeled as opinion.

Draft:
[PASTE THE DRAFT]

This catches citation theater. A blue link is not evidence. The page behind it has to support the sentence in front of it.

The prove-it-with-a-tool prompt

The CRITIC method gives the model external feedback from tools such as search, a calculator or code, then asks it to revise. That last part matters. A model grading its own homework from memory is still the same kid holding the answer key.

Use this on forecasts, pricing, analytics and anything headed toward a board deck.

Treat every number and factual claim below as untrusted until checked.

1. List each number, formula, factual claim and source assumption.
2. Recalculate every number with a calculator, spreadsheet or code tool. Do not do arithmetic in prose.
3. Verify factual claims against primary sources.
4. Return a reconciliation table:

Item | Original value or claim | Tool or source used | Verified result | Variance | Owner action

5. Stop and flag any item that cannot be reproduced.
6. Revise the work only after the reconciliation is complete.

Work to check:
[PASTE THE WORK]

The dumbest useful rule in this whole article might be this: make the machine show the math outside the paragraph.

These prompts can lower the error rate. They can’t own the answer.

You still do. So when the board asks where the number came from, the prompt is not your defense.

Research and receipts

• Moffatt v. Air Canada, 2024 BCCRT 149, case summary and award details

• Resume Builder survey of 1,342 U.S. managers, June 30, 2025

• Reuters report on the Meta employees’ lawsuit, July 22, 2026

• Madeleine Clare Elish, “Moral Crumple Zones,” Engaging Science, Technology, and Society, 2019

• Nikpayam, Kremer and de Véricourt, “AI-Assisted Decisions in Organizations,” working paper

• NIST AI Risk Management Framework

• Dhuliawala et al., “Chain-of-Verification Reduces Hallucination in Large Language Models,” Findings of ACL 2024

• Gao et al., “RARR: Researching and Revising What Language Models Say, Using Language Models,” ACL 2023

• Gou et al., “CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing,” ICLR 2024
  

I’m sitting in a swanky restaurant in Los Angeles, trying to look like I belong there.

I don’t.

The room has that particular L.A. glow. Low light, beautiful people, people wearing t-shirts that cost more than my first car. Across from me is the showrunner of a television series you definitely know.

I knew his work long before I knew his name. I’d watched it as a fan and studied it as a storyteller. His work had been inside my head for years. Now he was sitting a few feet away, casually deciding what to order.

I was doing my damn best to appear normal.

Inside, I was seventeen years old again.

I could feel every credential I’d ever earned falling out of my pockets. Twenty years building brands. Thousands of ideas. Companies built. Rooms won and lost. None of it seemed to matter.

My brain had reached its verdict.

Right before the panic fully set in, my friend Aaron introduced me.

“Pete is one of the best storytellers and brand builders I know.”

The showrunner looked up and smiled.

“Oh,” he said. “I thought you were just an AI automation guy.”

Four seconds.

That was all it took.

He wasn’t trying to insult me. He was telling me what he saw.

And what he saw was the version of myself I’d been advertising.

For months, I’d talked about AI constantly. Adoption, agents, automation, transformation. I was so busy proving that I understood the future that I’d buried the twenty years that made me useful in the first place.

The stories and scars. The taste. The pattern recognition. The ability to sit inside a messy problem long enough to see a way through it. I am a maker, marketer, entrepreneur, and most importantly, a human with a unique point of view.

And I’d taken everything I’d earned and hidden it behind a tool anyone could rent for less than the coffee sitting on the table.

The comment hurt because it was true.

I’d made the machinery visible and the craft invisible.

The rented layer

A lot of smart people are making the same mistake.

Founders and experienced operators are reducing themselves to the tools they use. Strategists are becoming prompt people. Writers are becoming content machines. Agency owners are selling speed when their real value has always been judgment.

I think of this as tech sharecropping.

You build your identity and your value on rented land. The platform owner can change the price, improve the machinery, or give the same equipment to everyone else in your market.

You may become faster. You may make more things. That doesn’t mean you’re building an advantage.

The AI conversation makes this easy to forget because it’s obsessed with the Platform du jour or best AI model this week. Every week brings a new comparison, leaderboard, or declaration that one model has made the previous one obsolete.

I use these tools every day. I test them, push them, and sometimes break them. This isn’t an argument against the tech.

It’s an argument against mistaking access for advantage.

Your competitor can access the same model. So can the freelancer underbidding you. Your client can open the same blank chat window and type a version of the same request.

The model can’t be the whole moat because you don’t own it.

What you own is everything you bring to it.

The prompt was never the whole point

The AI world has started calling this context engineering.

Strip away the technical language and the idea is straightforward. Models produce better work when they receive the right information for the job. That includes source material, examples, constraints, decision rules, and a clear picture of what good looks like.

The answer is rarely a more theatrical prompt.

For the last few years, people have treated prompting like spell casting. Find the right phrase. Assign the model an impressive role. Tell it to breathe deeply, channel a genius, and produce something extraordinary.

Sometimes those tricks help. They don’t replace substance.

You can tell a model to act like the best strategist in the world. It still doesn’t know why you rejected the last three campaign ideas. It doesn’t know the phrase a customer used when they finally admitted what was wrong. It has never seen the proposal that won the biggest deal in your company’s history or the one that died in procurement.

Without those things, it’s guessing.

The most useful distinction isn’t between good prompts and bad prompts.

It’s the latent space between the request and the brief.

A request tells the model what you want.

A brief gives it enough of your world to make intelligent choices.

Your business is leaking memory

Here’s the strange part. You probably already possess most of the context you need.

It’s scattered across your call transcripts, proposals, project notes, emails, Slack threads, voice memos, and postmortems. Some of it’s trapped inside the head of the person everyone waits for before the work can continue.

It’s hiding in ordinary moments.

A buyer stops giving you the polished answer and admits what they’re afraid of.

A founder rejects a headline and explains why it feels wrong.

A project falls apart, then teaches the team something that never makes it into the next kickoff.

A senior person makes a tiny decision that changes the entire direction of the work. Everyone else sees the decision. Nobody captures the reasoning behind it.

That material isn’t just digital exhaust. It’s the memory of the business, and that’s where the magic lives.

Most companies lose pieces of that memory every day. Meetings happen, decisions get made, and people move on. The founder gives the same correction six weeks later because the first correction never became a standard.

Then the company points an AI model at a pile of files and wonders why the answer feels generic.

The model may be working perfectly. The business hasn’t made its knowledge usable.

Your context isn’t the number of documents you can upload. It’s the record of how your company notices things, makes choices, learns, and communicates.

That record becomes more valuable each time you use it.

The four vaults

I use four vaults to organize this material.

Not because every business problem needs a four-box diagram. The vaults are simply a way to keep four different kinds of value from collapsing into one large folder called “knowledge.” or worse a “company brain” that everyone is selling you on X or LinkedIn but falls apart after using it for a week.

Decisions

The Decisions vault holds your judgment.

This is where taste becomes teachable.

Taste often sounds mystical because experienced people make decisions faster than they can explain them. They know a strategy is weak before the evidence is neatly organized. They can sense when an offer is carrying too much weight or when a clever headline is hiding an empty thought.

A model doesn’t begin with your judgment. It begins with patterns gathered from other people’s work.

Each time you correct the model, you reveal part of your own decision system.

“That claim needs proof.”

“We are explaining our process before the buyer cares.”

“This sounds polished, but it has no tension.”

“This idea is interesting. It isn’t useful.”

Don’t throw those corrections away after fixing the draft. Save the reason behind them.

Over time, they become a taste file. Not a vague manifesto filled with words like bold and human. A useful taste file contains rules, examples, and the reasons those standards matter.

Records

The Records vault holds what actually happened.

This includes sales calls, lost deals, customer interviews, project outcomes, client feedback, and the language people use when they stop performing.

I record my calls because I want the market’s description of the problem, not the version I invented in a conference room.

Buyers rarely speak in positioning statements. They say things like, “I don’t wanna look stupid in front of my team.” They admit that everyone is busy but the business is standing still. They tell you that the company bought the tool, trained the staff, and watched nothing change.

That language has weight because it came from a real moment.

Your customers are often better copywriters than you are. They’re closer to the fear.

The value isn’t only in collecting quotes. The larger prize is the pattern that appears across conversations. What triggers the search for help? Which concern keeps returning? What do people need to believe before they move?

The records turn your opinions into evidence.

Process

The Process vault holds how the work really happens.

Not the polished case study where the project moves from discovery to strategy with suspicious ease. I mean the decisions, checks, shortcuts, and course corrections that make the work good.

Most experienced people carry this knowledge in their bodies. They notice something and change direction. They ask a question that seems obvious to them and completely changes the room.

For many founders, this knowledge lives in the worst possible place.

Their head.

That feels like expertise until the entire team needs access to it at once. Then expertise becomes a traffic jam.

AI can’t scale a process you’ve never explained. Your team can’t follow a standard that exists as an UNdocumented feeling.

Record yourself doing the task. Narrate the parts you usually skip when explaining it. Talk about what catches your attention and what causes you to abandon one direction for another.

The parts that feel embarrassingly obvious are often the parts everyone else needs.

Voice

The Voice vault holds the shape of your thinking.

Voice isn’t a list of favorite words. It isn’t adding a joke to the second paragraph or asking the model to sound punchier.

Your voice includes the things you notice, the tension you’re drawn toward, and how long you allow an idea to develop before you name the lesson.

Most AI writing isn’t terrible. That’s part of the problem.

The sentences are clean. The structure is visible. The conclusion arrives on schedule. Nothing is technically broken, but nobody seems to be home.

There’s no pressure inside the writing. No risk. No sense that a person has lived through the idea and might have something at stake.

A good voice file should help a model understand how you think, not merely how you decorate sentences. It should contain real writing, rejected language, beliefs, rhythm, and examples of work that feels unmistakably yours.

The model can produce words.

Your context gives those words a reason to exist.

Homework for Work

Matthew Dicks has a practice called Homework for Life.

Each day, he records one small storyworthy moment—the kind that would otherwise disappear if you didn’t catch it. Not the whole story. Just the moment.

Over time, those entries become a record of a life that memory would have erased.

I became so obsessed with this that I wouldn’t stop. So obsessed that I ended up helping him build it into an app so everyone could do it too.

I think businesses need a version of that practice.

Call it Homework for Work.

At the end of each week, capture the moments that revealed something useful about the business.

What did a buyer say that changed your understanding of the problem?

Which decision did you make that would be difficult for a less experienced person to explain?

What correction did you give that you never want to give again?

Where did the work change direction, and what caused it?

This isn’t a meeting summary. Most summaries preserve what was discussed and lose what mattered.

Homework for Work saves the moments when the truth became visible.

Those moments are the raw material for the vaults. A buyer’s admission belongs in Records. A creative correction may become a Decision rule. A change in direction might expose part of the Process.

The companies that become meaningfully better with AI won’t simply have access to better tools.

They’ll become better at remembering what they learn.

The flywheel

Once you treat context as memory, the system isn’t complicated.

Capture something that happened. Turn it into a usable form. Apply it to real work. Correct what comes back. Save the correction.

A sales call improves the buyer-language bank. A difficult project adds a warning to the process. A rejected draft makes the taste file more precise.

Most people use AI like a vending machine. They enter a request, receive an answer, and start from zero again the next morning.

A context system carries knowledge forward.

The proposal you write next month should benefit from the objection you heard today. A new employee should have access to the thinking behind your standards, not only the finished examples. The AI shouldn’t repeat an error you corrected three times already.

A competitor can buy access to the same model tomorrow.

They can’t buy the history of your decisions.

What this changes for service businesses

AI didn’t make your experience worthless. It just exposed how little of that experience had been captured.

When judgment lives only in your head, the technology feels threatening. It appears to be learning the surface of a skill you spent years developing.

When your judgment becomes context, the relationship changes.

The model can prepare the work, compare options, find patterns, or create a first draft using the standards you’ve already defined. The expert remains responsible for the decisions that matter. The system simply stops asking that expert to rebuild the starting point every time.

That changes the economics of a service business.

You’re no longer using AI only to do ordinary work faster. You’re allowing your best thinking to reach more work without turning every output into a pale imitation of the founder.

The goal isn’t to remove the expert. It’s here to stop wasting the expert’s time.

That’s also how you avoid the end-shittification of everything AI touches. You don’t fight generic output with louder instructions. You give the system more truth, stronger standards, and real examples.

Build the first version this week

Don’t turn this into an enterprise knowledge program or a company brain like all the fancy tech-bro CEOs are talking about. You don’t have their resources, and you don’t need them to just get started right now.

Pick one recurring piece of work. It could be a proposal, a client brief, a sales follow-up, or a strategic review.

Then gather the context that work deserves.

Start with evidence. Find the calls, outcomes, and examples connected to the task. Add the decision rules the work must respect. Document the process just far enough that another person could follow it. Include real samples of your voice.

Then create a brief.

The best prompt engineers aren’t searching for magic sentences. They’re designing a clear job for the model. They define the source of truth, show it examples, establish constraints, and explain how the result will be judged.

The prompts below are built around that discipline.

The Context Audit

Use this to find the knowledge your business already owns.

I am building a context system for this recurring business task:

[DESCRIBE THE TASK]

Review the material I provide and organize the useful information into four categories:

DECISIONS
Standards, judgment calls, approvals, rejections, and tradeoffs.

RECORDS
Buyer language, calls, proposals, outcomes, objections, project history, and proof.

PROCESS
Steps, handoffs, decision points, quality checks, and places where human judgment matters.

VOICE
Point of view, recurring language, rhythm, stories, metaphors, humor, phrases to avoid, and examples that sound like us.

For every finding, include:

- The source
- The exact evidence
- Why it matters to the task
- Whether it appears current, outdated, contradictory, or uncertain

Do not create rules from a single example unless you label them as a hypothesis.
Do not invent missing information.
Do not treat an opinion as a fact.

Finish with:

- The ten most valuable context assets
- The most damaging gaps
- The first three things we should capture next

MATERIALS:

[PASTE OR ATTACH MATERIALS]

The Buyer Truth Miner

Use this with recorded sales or customer calls.

Analyze these sales or customer call transcripts.

Use only the transcripts as evidence. Include the exact quote and source call for every conclusion.

Identify:

- The language buyers use to describe the problem
- The personal or professional stakes behind the problem
- The event that made the problem urgent
- What buyers need to believe before they will act
- The objections that slow the decision
- The moment when their level of interest changes

Separate repeated patterns from rare but revealing comments.

Do not polish the buyer’s language.
Do not invent motivations.
Do not treat one person repeating a concern as evidence of a wider pattern.

Then create:

- A buyer-language bank containing the strongest quotes
- Messaging ideas grounded in those quotes
- Claims our current marketing should stop making
- Questions we should ask on future calls

TRANSCRIPTS:

[PASTE OR ATTACH TRANSCRIPTS]

The Taste File Builder

Use this to turn your feedback into standards another person can apply.

I will provide examples of work I approved, work I rejected, and the feedback I gave.

Identify the judgment rules behind my choices.

Create a rule only when it is supported by multiple examples or by an explicit statement from me.

For every rule, include:

RULE
Write it as a specific “We always” or “We never” statement.

PURPOSE
Explain the value the rule protects or the problem it prevents.

EVIDENCE
Reference the examples that support it.

GOOD EXAMPLE
Show what following the rule looks like.

FAILURE MODE
Show what happens when the rule is ignored.

CONFIDENCE
Label the rule confirmed, probable, or uncertain.

Also identify contradictions in my feedback and preferences that appear specific to one project.

Do not flatter me.
Do not use vague words such as bold, premium, authentic, or strategic unless you define what they mean in practice.

APPROVED WORK:

[PASTE]

REJECTED WORK:

[PASTE]

FEEDBACK:

[PASTE]

The Process Extractor

Use this after recording yourself completing a task.

Turn my narrated workflow into a process another person could follow.

For each step, document:

- The goal
- The input required
- The action taken
- The decision being made
- The rule guiding that decision
- The quality standard
- The common failure
- Where human judgment is required
- How AI could prepare, compare, check, or draft
- The output created

Then identify:

- Steps I skipped because they feel obvious to me
- Places where the process contradicts itself
- Decisions that rely on undocumented taste
- Handoffs likely to create delay or rework
- Work that should remain human-led

Do not make the workflow cleaner than my description.
Do not add generic steps I did not mention.
Flag missing information instead of filling it with assumptions.

Finish with a one-page checklist and the questions someone would need answered before completing the work without me.

NARRATION:

[PASTE TRANSCRIPT]

The Context Brief

Use this before asking the model to produce important work.

TASK

What needs to be created or decided?

[DESCRIBE THE TASK]

AUDIENCE

Who is this for?
What do they care about?
What are they likely to doubt?

[DESCRIBE THE AUDIENCE]

STAKES

Why does this work matter?
What happens when it succeeds?
What happens when it fails?

[DESCRIBE THE STAKES]

SOURCE OF TRUTH

Use the following records as evidence.
Do not make claims beyond them.

[PASTE RESEARCH, QUOTES, DATA, NOTES, OR PROJECT HISTORY]

DECISION RULES

Use these standards when making choices.

[PASTE TASTE FILE]

PROCESS

Follow this workflow.

[PASTE PROCESS]

VOICE

Use this voice guide.

[PASTE VOICE FILE]

REFERENCE EXAMPLES

These examples demonstrate the quality or behavior I want.
Explain which qualities you are carrying forward. Do not copy their surface language.

[PASTE EXAMPLES]

CONSTRAINTS

Include:
[LIST REQUIREMENTS]

Avoid:
[LIST BOUNDARIES]

DEFINITION OF DONE

The result succeeds when:

[LIST CLEAR CRITERIA]

OUTPUT

Return:

1. The finished work
2. Any unresolved factual

I’m sitting in my car on the hottest day of the year. I’m doing something behind the wheel that would have gotten me arrested a few years ago.

My hands aren’t on the wheel. They’re doing something they absolutely shouldn’t be doing.

And to make this even worse, my coworker is sitting next to me in the passenger seat. I was supposedly treating him to lunch, but I had already checked out. My eyes were glued to my phone, where an AI army was doing the work of a dozen people while a real human waited for me to look up.

We were headed to one of those little Mediterranean spots that looks forgettable from the outside and then ruins you for every polished fast-casual bowl place pretending to have soul. The kind of place where the hummus tastes like somebody’s grandmother still has veto power.

My car was driving itself. That was strange enough, but you haven’t heard anything yet. The more insane thing was what was happening on my phone. One AI agent was chasing research. Another was shaping copy. Somewhere else, a product idea was getting cross-referenced with analytics data and drafting product spec docs. The kinda work that used to require a big team, a meeting, a brief, another meeting to clarify the brief, and someone named in the calendar invite as optional but secretly essential.

Now it was happening while the car looked for parking and I stayed glued to my phone.

For a second, I felt the private thrill of it. This is the founder fantasy. The Iron Man console without the billionaire lighting package. A 10- or 15-person team, or at least a convincing shadow of one, running behind a few subscriptions and a phone screen.

Then the feeling turned. Not into guilt exactly. Something smaller and worse.

Shame.

Because I was sitting next to a person I liked and respected, trying to be present enough to call it lunch, while some part of me was still inside the command center. The machine was driving. The agents were working. I was checking whether the synthetic workers had come back with something useful.

I wasn’t resting.

I was supervising a company from a motor vehicle.

That’s the strange truth I keep swirling around my ADHD brain: AI has made me more productive. Obviously. It has also made work more available, more tempting, more endlessly expandable. It has taken the ceiling off my ambition and replaced it with a refresh button.

The old promise was that AI would give us our time back.

Some days it does.

Most days it gives our appetite better tools.

The productivity trap isn’t failure. It’s expansion.

The lazy critique is that AI is overhyped.

I don’t think that is the most interesting version. The more uncomfortable version is that AI is useful enough to get us into trouble.

It makes the first step cheaper. It makes the next step visible. It makes the ugly middle less embarrassing to enter.

You don’t open ChatGPT or Claude or an agent workspace and think, today I will ruin my boundaries and turn into a f*cking work-a-holic.

You think I will clean up this one thought before the meeting.

Then the thought turns into a doc. The doc suggests a product angle. The angle wants a landing page. The landing page implies a tool. The tool needs onboarding, copy, QA, a launch note, maybe a small experiment because now the thing exists enough to make not testing it feel irresponsible.

That’s how a thirty-minute save becomes a new country on the map.

This isn’t laziness. It’s productive compulsion: the feeling of progress without the old stopping points that used to constrain the work.

Visual proof: AI changes work patterns

Why the AI machine feels so good

There is a psychological reason this feels so sticky.

AI gives you progress in small, fast doses.

A sentence clicks. An outline stops wobbling. A bug you were avoiding suddenly has a plausible fix without spending hours on Reddit or StackOverflow chasing answers. The pile on your desk doesn’t disappear, but for 30 seconds it looks organized enough to forgive you.

Teresa Amabile and Steven Kramer found that progress in meaningful work is one of the strongest drivers of positive inner work life. AI manufactures that feeling at absurd speed. It doesn’t just help you work. It keeps handing you evidence that work is happening.

And the reward isn’t predictable.

Sometimes the answer is garbage. Sometimes it’s close enough to annoy you. One line or weird punctuation that you catch and give Claude the side eye as if it can see you looking at it with scorn.

That variability matters. The dopamine story is less about simple pleasure and more about reward prediction: the brain learning from the gap between what it expected and what it got. AI creates those gaps all day.

You prompt, wait, judge, tweak. The next run might save the hour. The next branch might be the breakthrough. Or it might send you down the kind of expensive little rabbit hole where someone erases the production database because the grown-up with scar tissue got replaced by a subscription.

Now your work tool has the emotional rhythm of a casino and the costume of productivity. A tiny CEO psyop you run on yourself.

AI removes friction, including the friction that protected us

Friction gets treated like the enemy because a lot of it is.

Bad forms. Slow handoffs. Meetings that exist because nobody wants to write the decision down. Whole afternoons lost to formatting a thing nobody will read twice.

Of course we want less of that.

But some friction was doing important invisible work for us.

The time it took to brief another human forced the idea to become clearer. The cost of asking a designer for another round made you decide whether the round mattered. The pain of building a prototype made you choose which idea deserved to become more than a note in your phone.

AI collapses those costs, and that collapse feels like freedom right up until every impulse starts wearing a little badge that says urgent.

That tiny urgent badge is the whole party trick.

When every idea can have a first draft faster than your standards can get their pants on, the first draft starts asking for a second.

When every product thought can become a spec, the spec starts lobbying for a roadmap. So now your taste has to work harder because polish no longer proves much.

Visual proof: The ironies of generative AI

Your old team has now become a software subscription

This is the part that still feels a little obscene to say out loud.

For a few subscriptions a month (and a shitload of tokens), I can simulate pieces of a team that used to require salary, management, meetings, context transfer, and the emotional cost of asking another human to care about my half-baked ideas.

The simulation isn’t the same as a great human team. It has no taste until I lend it taste. It has no judgment until I give it standards. It has no real accountability unless I build the accountability around it.

But it’s good enough to change the emotional math.

The old bottleneck was simple: do we have the team for this?

The new bottleneck is weirder: do I have enough attention to supervise all the synthetic help I can summon?

That’s the chart I would put on the wall. Labor used to slow me down. Now judgment has to. And judgment is more annoying because you can’t delegate it without risking abdication.

And if you are a founder, builder, control freak, ADHD brain, creative strategist, or high-agency operator, this is both intoxicating and slightly poisonous. You can finally get the thing in your head into the world.

The problem is that everything in your head now wants a turn.

Working backwards is the brake you’ve been looking for

This is why the lens matters. Most AI sessions start with the tool’s appetite.

What can it make? What can it automate? What can I spin up while I’m waiting for my coffee, sitting in the car, pretending I’m only checking one thing?

That feels practical. It’s also how the machine becomes the center of the room.

Working Backwards changes the order.

Amazon teams have used a strange ritual before building: write the future press release and FAQ first. Not the roadmap. Not the task list. The ending.

That sounds corporate until you realize how human it is. It forces the idea to walk into the room as if it already happened and answer for itself.

Who changed? What got easier? What fear went away? What would make someone care enough to switch, pay, trust, or tell another human? What breaks if we’re honest? What has to be true before this idea deserves a team, a launch plan, or even an agent run?

The document isn’t the magic. The sequence is. If it was good enough to keep Amazon from building every shiny object that wandered through the room, it’s good enough for those of us now surrounded by machines that can make every shiny object look strangely plausible.

Don’t start with the machine. Start with the future that would make the machine worth using.

Visual proof: The productivity paradox

Use AI as a challenger, not a butler

I’m not arguing for some artisanal, anti-AI purity ritual. I use these tools constantly. I build with them, write with them, argue with them, and use them to pull half-formed things out of the fog inside my head. Sometimes they even become money-making machines.

The change is smaller and harder. The nuance is where the magic gets made: stop making the machine your butler before you’ve made it your challenger.

Before you ask it to make the thing, ask yourself whether the thing deserves to exist.

Before you ask it for more options, ask what future would make one option obviously better.

Before you ask it to make something sharper, define what sharpness would actually mean. Sharper for whom? Sharper against what fear? Sharper in service of what decision?

This is the difference between using AI to avoid judgment and using AI to sharpen judgment.

Visual proof: Provocations restore critical thinking

The hungry ghost with a productivity app

Here’s the mirror I want you to feel.

You’re using AI all day and somehow ending the day further behind.

You have more output, but less closure.

You keep opening loops because starting has become almost free.

After a long AI session, you don’t feel rested. You feel bright-eyed and overclocked, like your brain stayed late at the casino and drained your bank account.

You’re not only asking AI to help you move faster. Sometimes you’re asking it to postpone the moment where you have to choose.

That’s why better prompts, agents, flux capacitors, and Willy Wonka widgets alone won’t save us. We need better stopping rules. We need a way to decide what deserves the machine before the machine makes everything feel deserving.

Before you open your AI slop machine again, do this instead

The lesson isn’t that AI is bad. That would be too easy, and honestly too false.

The next time I feel that little surge, the one that says, just spin up one more agent, just run one more draft, just see what comes back, I want a better question waiting for me.

What future am I trying to make true?

If I can’t answer that, the machine waits.

If I can, then AI becomes leverage instead of appetite with a login.

That is the whole move: decide the ending, then invite the machine.

Steal these prompts before the machine steals your afternoon (and soul)

Prompt 1: The Working Backwards Brake

Don’t create the output yet. First help me work backwards.

Here’s the thing I’m considering:
[describe the product, offer, article, campaign, feature, workflow, or decision]

Before making anything, define the future that would make this worth doing.

Answer in this order:

1. Who is changed by this?
2. What are they struggling with before?
3. What can they do, feel, avoid, trust, or become after?
4. What is the plain-language promise?
5. What would a skeptical customer, teammate, buyer, operator, and critic ask?
6. What has to be true for this to work?
7. Is this worth building, writing, sending, or automating yet?
8. What is the smallest next test?

Be direct. Don’t flatter the idea. Challenge my original thinking before helping me produce the artifact.

Prompt 2: The Productive Compulsion Audit

I’m using AI on this workstream:
[describe workstream]

Help me identify where AI is expanding the work instead of reducing it.

Look for:
- new loops created
- unnecessary second drafts
- tasks that feel easy but don’t matter
- places where I’m using AI to avoid choosing
- outputs that create review burden
- missing stopping rules

Then give me three rules for when to stop.

Prompt 3: The Friction Filter

I want to remove friction from this workflow:
[describe workflow]

Before suggesting automation, separate the friction into three categories:
1. Waste friction that should be removed
2. Judgment friction that protects quality
3. Relationship friction that creates alignment

Then tell me what AI should automate, what humans should keep, and what should happen before the machine starts.

Try the Working Backwards lens

If you want to try the lens on a real idea, use my free Working Backwards tool. It walks you through the Amazon-style press release, FAQ, verdict, and PRD sequence before you let AI start multiplying the work:

I bought a car that made me worse at driving.

That is not a sentence Tesla will be putting in a commercial, but it is unfortunately true.

Since getting my new one, I use Autopilot for probably 95 percent of my driving. Highway. Traffic. Lane changes. The tiny micro-adjustments my hands used to make without asking permission from my brain.

At first it felt like magic. Then it felt normal. Then one day I actually had to drive.

Not sit in the driver’s seat.

Drive.

Hands on the wheel. Eyes scanning. Body making calculations. How fast is that car coming up on my left? Do I have the gap? Is this guy drifting? Why does this lane suddenly feel like it was designed by someone who hates me?

And I felt anxious.

Not because I forgot how to drive. I have been driving for decades. But something in me had gotten softer. A skill I used to trust without thinking suddenly wanted supervision.

That scared me a little.

Not in a dramatic way. In a stinging, embarrassing, “oh, this is how atrophy happens” kinda way.

First the machine helps you. Then the machine handles it. Then one day the thing comes back into your hands and your hands feel less like yours.

That is what I keep thinking about with AI.

We aren’t just outsourcing tasks now. We’re outsourcing first drafts, first thoughts, first questions, first interpretations, first attempts at language, first attempts at judgment.

The danger is not that AI does things for us.

The danger is that it does the wrong things first.

Because the first pass is where you find out what you think.

The Three Words That Flatten You

The most dangerous AI prompt might be three words long.

Make this sharper.

I see it everywhere. A founder has a sales email that feels soft. A strategist has a deck that almost says something. A writer has a post that sounds fine, which is usually the first sign that it is not fine at all.

They type the magic little sentence.

Make this sharper.

The machine obeys. It trims. It polishes. It adds confidence. It makes the thing sound more like something a competent person would post on LinkedIn after two coffees and one mild identity crisis.

Then the person reads it and feels the tiny hit.

Better.

Except it is not better. It is smoother.

And smooth is not sharp.

AI does not know what sharper means. It knows shorter, cleaner, warmer, more formal, more casual, more structured, more polished.

Sharper requires a standard. Sharper requires taste.

Sharper requires knowing what the work is supposed to do, who it’s for, what truth it is carrying, what must be removed, and what line should make the reader sit back because it found the bruise.

That is not the machine’s job. It’s yours.

AI is the gear. You are the guide.

Gear can help you climb faster. Help you carry more weight. But gear does not choose the mountain. Gear does not know your standards. Gear does not know what kind of person you are trying to become.

Hand the gear the map, the compass, and your sense of self, and it will quietly walk you toward the middle.

The Middle Has Gravity

The middle is where most AI output wants to go.

Models are trained on a giant pile of language, habits, patterns, cliches, templates, and repeated thoughts. They predict what usually comes next.

That is the miracle. It’s also the freaking trap.

Your best work rarely lives where the most probable next word (token) lives. Your best work lives in the scar. The weird angle. The personal rule. The earned belief. The thing you noticed because you were actually in the room when the room went quiet.

The machine can help you find that. Left alone, it often pulls you away from it.

It will make you sound reasonable. It will make you sound polished. It will make you sound like everyone else who asked the same model to improve the same kind of thing.

The research is starting to catch up. A Microsoft and Carnegie Mellon study of 319 knowledge workers found that higher confidence in GenAI was linked with less critical thinking, while higher self-confidence was linked with more critical thinking. The people who trust themselves use AI better. The people who over-trust the machine hand over the wrong layer. Microsoft Research

Another study found that GenAI improved individual creative output, but AI-assisted stories became more similar to each other. AI helped people make better work, but the work started clustering toward sameness. arXiv

A 2026 meta-analysis from Tilburg University looked across 19 studies and 61 effect sizes and found a small but real homogenization effect in human-AI co-creation. Their language is careful. Mine is simpler. AI can pull the crowd toward the same polished middle. Tilburg University

That is the uncomfortable truth.

AI can raise the floor and lower the ceiling at the same time.

It can make weak work decent. It can make slow work fast. It can make blank pages less terrifying. Used lazily, it can also sand down the edge that made the work worth making in the first place.

Microplastics For The Mind

I have called this CrapGPT before. Microplastics for the mind.

A little generic email here. A little AI-polished post there. A little “make this sound more professional” before lunch.

None of it feels dangerous in the moment. That is the problem. It feels useful. It feels productive. It feels like friction went away.

Then one day your work has no pulse.

Your voice sounds borrowed.

Your taste goes quiet.

You start accepting words because they are clean, not because they are true.

That is how the flattening happens. A little residue at a time.

The answer is not to stop using AI. I use this technology every day. I build with it, think with it, argue with it, test ideas against it, and use it to move faster through the fog.

But speed without standards is just faster drift. And the new skill is not knowing which model to use.

The new skill is knowing what part of yourself should never be delegated.

Sharper Is Not A Vibe

AI cannot make something sharper until you define what sharper means.

Sharper might mean more specific. More emotionally honest. More commercially direct. More useful to the reader. More dangerous. More restrained. More evidence-backed. Less performative.

Those are different jobs.

A machine can’t choose between them without borrowing a value system from somewhere.

So give it yours.

Before I ask AI to improve anything now, I define the standard.

For this piece, sharper means the idea must be easy to repeat. The enemy must be clear. The science must support the claim without turning the piece into homework. The reader must feel slightly implicated. The payoff must be useful enough to save. The voice must sound like a person who has lived the problem.

That is a real brief.

“Make this sharper” is not.

The Spike Stack

That last row is the whole game.

The best AI use does not erase your fingerprints. It makes them harder to fake.

Flat AI use makes you sound like the meeting notes from a company that has lost the will to tell the truth.

Spiky AI use makes you more specific, more discerning, more useful, more yourself.

Use AI As A Challenger

Most people use AI like a butler.

Do this. Write that. Summarize this. Make me look prepared.

That is fine for low-stakes work. I am not here to shame anyone for using AI to clean up a calendar invite.

But the work that matters needs a different posture.

Use AI as a challenger.

Ask it to find the weak claim. Ask it to name the assumption. Ask it to show where you are hiding behind language. Ask it what a skeptical buyer would reject. Ask it what sounds like everyone else. Ask it what only you could say.

A 2025 arXiv study tested “provocations,” short AI-generated critiques and alternatives added to AI suggestions. The researchers found that these provocations could trigger critical and metacognitive thinking.

Translation: AI got more useful when it stopped simply obeying and started making the user think. arXiv

Good AI should not replace your reps.

It should make the reps better.

It should not remove the friction.

It should add the right friction.

Because friction is not always failure. Sometimes friction is the gym.

The Lens Before The Prompt

This is why I have been thinking so much about lenses.

A lens is the frame that decides what becomes visible.

Backwards asks: what future would make this worth doing?

Inversion asks: how would this fail?

Subtraction asks: what can we remove without losing the soul?

Compounding asks: what gets stronger every time we repeat it?

Proof asks: what evidence would make this undeniable?

Most people ask AI for an output.

Better users ask AI for a lens.

A prompt says, “Write me the answer.”

A lens says, “Show me what I am not seeing.”

The prompt is downstream of the lens.

That is why prompt libraries feel useful for about twelve minutes. They give you words to paste, but they do not give you better perception.

And perception is the scarce thing now.

The ability to see what matters before the machine multiplies what does not.

Steal This Prompt

Use this before you ask AI to improve anything that matters.

Do not rewrite yet.

First, identify the lens I am using.

Then define what "sharper" should mean for this piece based on:
1. The audience
2. The emotional job
3. The business goal
4. The point of view
5. The standard the work must meet

Then give me five possible definitions of sharper:
- More specific
- More emotionally honest
- More commercially direct
- More contrarian
- More useful to the reader

For each definition, explain what would change.

Then recommend the best sharpening strategy.

Challenge my original thinking before helping me create the output.

Only rewrite after the standard is clear.

That prompt works because it makes you guide the machine before the machine starts generating.

It puts your taste back in the driver’s seat.

Drive Or Be Driven

The old divide was access.

Who had the software. Who had the team. Who had the budget. Who had the technical skill.

That divide is shrinking. And the new divide is posture.

One person uses AI to avoid thinking. Another uses AI to think better.

One person uses AI to finish. Another uses AI to understand.

One person uses AI to sound smart. Another uses AI to become harder to copy.

Same tool. Different human. Different future.

The risk is not that machines think.

The risk is that we stop noticing when we are not.

AI should make you more spiky, not flatter. It should make your taste sharper, your questions better, your standards clearer, your thinking more alive.

It should help you find the sentence only you would write, the offer only you could make, the strategy that comes from your lived experience, not the internet’s average opinion in a nicer jacket.

The machine can make words. Code. Images. Options. Plans. Summaries. Strategies.

But it can’t decide what’s worth becoming.

Drive the machine, or the machine will drive you toward the middle where dull and dumber lives.

The office still had the wood paneling from the seventies. Dark, heavy, a little scuffed at the corners. The kind of room where deals used to close on a handshake and a glass of something brown. I was in the leather chair across the desk, the one that has probably held a hundred nervous salespeople, and I will be honest, I was a little nervous too. I did not yet know if I could help this man.

I have done this work for twenty years. So believe me when I tell you that what happened in this company over the next two weeks is the clearest proof I have ever seen that the way we build things just changed for good. By the end of this you will know how to do it yourself, and how to not get burned doing it. But first you have to meet him.

The guy across the desk built this place. Fifty million in revenue. Hundreds of people. He started here in sales more than twenty years ago and worked his way up, and he has run the whole thing as CEO for over fifteen. He can walk the floor and tell you what every person does and which one to trust on a bad day. He knows this company in his bones.

He also had no idea what was happening inside it anymore. Not the day to day. He said it himself, low and a little raspy, the way you say a thing you do not want overheard. He was staring at what he called an AI extinction event, and he knew two things for certain. He had to adopt this stuff. And he did not have the people to pull it off.

A former client told him to call me. So there I was, in the wood paneling.

I should tell you where I come from, because it matters for what happened next. I wrote my first line of code when I was twelve. I have spent my life building software, and I wrote about this exact shift back in early 2025, when old-school coders like me were still half-laughing at it. So I am not the guy who is amazed this works. I am the guy who watched the wall get built and can tell you, with total certainty, that it just came down.

I did not teach him with a slide deck. I taught him the only way that actually sticks. I asked him what he does all day. What he wishes he understood about his own company. The question that keeps him up at 2 am when his whole family is sleeping soundly in their beds.

Then I put the thing in his hands. Because that is the secret nobody tells you about teaching anyone anything. You do not explain it to them. You let them play with it. We had his team drop Claude onto his laptop, and we just started.

Two weeks later he called me, and he was talking fast. I could almost hear his heartbeat beating fast over the phone.

His meetings had changed. He could not stop describing it. People used to walk into his office and have conversations about conversations. They would circle back. They would align on the alignment. They would book a follow-up to discuss the thing they were already discussing.

Now they walked in with prototypes.

Someone took a real customer complaint and mocked it up as a clickable thing, a little Claude artifact you could actually use. Someone else built a tool to show a broken process instead of describing it for the ninth time. The sales guy who became a CEO, who had never written a line of code in his life, was sitting in a room where his people had stopped talking about the work and started showing it.

He told me the strangest part was his own silence. For fifteen years he had been the loudest voice in every room, the man with the answer. Now he just sat back and watched his people show him things he never thought to ask for. That was the turn. Not the software they bought. The day the guy at the head of the table stopped performing certainty and started building alongside everyone else.

That is what this piece is about. Not the tools. The shift. How you go from talking about the work to showing it, and how you do it without getting somebody hurt.

Because both things are true at once. You can now build software by describing it. And a lot of people doing that are building themselves a disaster and calling it progress.

The difference is not talent. It is knowing which jobs you do yourself in the garage, and which ones go up on the lift.

The barrier just fell, and nobody sent the email about it

Here is what actually happened in the last two years, in plain terms.

For your whole career, making software meant hiring someone who could code. That was the wall. You had the idea, they had the keys. You waited, you paid, you hoped they understood what you meant.

That wall is gone. Not “getting shorter.” Gone. The new requirement for building a working tool is that you can describe what you want clearly. That is it. People call this vibe coding, which is a goofy name, but the goofy name is hiding something serious.

This is not the first wall to fall. It happened to photography when the phone killed the darkroom. It happened to publishing, to video, to music. I have called it the Great Democratization Cycle, and software was always going to be next. It just took longer, because the wall was higher.

It means the bottleneck moved. It used to be “can you build it.” Now it is “do you know what you actually want, and do you know when you are in over your head.”

Which is good news for you, because knowing what you want is literally your job.

Here is the line I want stuck in your head for the rest of this piece. Prototypes are the new spec doc. You used to describe what you wanted in a forty-page document nobody read. Now you build a clickable version in an afternoon and let people poke at it. Showing beats telling, every single time. A thing people can click settles an argument that a document would have dragged out for three meetings.

Red light, green light, don’t get slapped

This is the whole safety system. If you remember nothing else, remember this part. I am going to talk like a mechanic now, because that is the right frame.

Some jobs you do yourself in the garage. Low stakes, forgiving, nobody gets hurt if you get it a little wrong. Some jobs go up on the lift and get a professional, because if you get them wrong, someone bleeds. Same toolbox. Completely different risk.

Green light. Build it yourself, today, no fear.

Internal tools. The stuff only your team touches. A tracker, a calculator, a little dashboard. The audience is forgiving because the audience is you.

Mockups and prototypes. The clickable version of an idea you want feedback on before you spend real money. This is the highest-value thing on the list. Build the fake one, watch people use it, learn what is wrong before it costs you.

Personal automations and throwaways. The one-off thing you need once. Build it, use it, delete it.

The simple test for green light is this. No login. No payments. No real customer data. If your thing has none of those three, you can ship it without a second thought.

Red light. Do not ship the thing you vibe-coded. Call an AI mechanic.

The moment real customers touch it at scale. The moment money changes hands. The moment it holds personal data, health data, financial data, anything you would hate to see leak. Anything in a regulated industry. Anything that has to keep running for years without you babysitting it.

Here is why, and it is not snobbery. AI-built code can look completely correct and be wrong exactly where you are not looking. It will invent a function that does not exist. It will leave a door unlocked. It will handle the normal case perfectly and fall apart on the weird one. In the garage, that’s a funny story. Pointed at your customers and their data, that is a headline.

One company learned this the public way this year. A site built entirely by vibe coding leaked over a million login tokens, because the AI wrote the lock and nobody checked that it actually locked.

Everybody draws exactly the wrong lesson from a story like that. The lesson is not that the AI screwed up. Of course it did. People screw up too, all day long. A few years back an intern at HBO Max fat-fingered a blank test email to the entire subscriber list. The subject line said “Integration Test Email #1.” It became a meme for a week. And then it was completely fine. You know why it was fine? Because there was a human in the loop. A name on the mistake. A manager who said yeah, that was us, we got them. There was an accountability trail.

Vibe coding can erase that trail. When the thing that made the call was a model running at 2am with nobody watching, there is no intern to coach and no manager to own it. There is just a hole in your business and a shrug.

IBM understood this in 1979. There is a slide from an old internal training deck that still makes the rounds. It says, in plain beige letters, “A computer can never be held accountable. Therefore a computer must never make a management decision.” Forty-six years ago, in a room full of mainframes, somebody already saw the whole thing. The machine can do the work. It cannot hold the bag.

That is the AI mechanic's whole job. The mechanic keeps you on the road. The mechanic keeps you safe. The mechanic is the line of defense, the name on the work, the human who holds the bag so a machine never has to pretend it can.

Yellow light, since real life lives here. You want to actually sell this thing. Fine. Vibe code it to prove the idea and win the first few users. Then, before it becomes your real product, get a human who knows what they are doing to look under the hood. Use the speed to find out if you are right. Do not use the speed to skip the seatbelts.

The whole map on one grid

If you want the cheat version, it comes down to two questions. How bad is it if this breaks, which is your blast radius. And can you undo it when it does, which is whether there is a rollback, a backup, a trail back to safe. Everything else just slides you around those two.

The things that decide which box you land in: sensitive data, actions you cannot reverse, whether there is an audit trail, what it costs every time it runs, who gets access, how reliable the code has to be, and whether anyone will have to extend or maintain it later. Check a few of those and you have slid out of the top-left, where you should be playing, and into the bottom-right, where you should be calling for help.

The simplest possible version: the Claude artifact

Before we get into the full build, I want to give you a win you can have in the next ten minutes, with zero setup, zero accounts, zero scary words.

It is called an artifact. When you ask Claude to build something interactive, it does not just describe it to you. It builds the actual working thing right there in the conversation, in a panel next to the chat. A calculator. A little app. A dashboard. A game for your kid. You watch it appear.

And here is the part that surprises people. You can hit publish and get a link. Send that link to anyone. They do not need a Claude account. They do not need to pay anything. It just works, whether ten people open it or ten thousand. That part costs you nothing.

You change it by talking. “Make the button bigger.” “Add a column for the date.” “I hate the blue, make it warmer.” It is the most forgiving sandbox ever built, because there is nothing to break and nothing to set up.

Go do this today. Pick one annoying little thing in your week and ask Claude to build you a tool for it. That is the on-ramp. Everything after this is just turning that magic into something more permanent.

Now the real thing, step by step

Okay. You have felt the magic. You want to build something that sticks around, lives at a real web address, and does not vanish when you close the chat. Here is the whole journey in plain language. Five steps. No computer science degree required.

Step one. Write the spec, so you are not building blind.

The clearer you are about what you want, the better the thing you get back. Garbage in, confident garbage out. So spend a few minutes turning your rambling idea into a tidy one-pager before you build.

I built a free tool for exactly this, because I got tired of watching smart people start with mush. It is at makemyprd.com. You dump in your messy idea, the problem, who it is for, what it should do, and it hands you back a clean structured spec. Goals, scope, the features, the edge cases you forgot. No cost, no card. Use it or write your own on a napkin, I do not care, just do not start from nothing.

Remember the line. The spec is the brief. The prototype you are about to build is the proof.

Step two. Build it with Claude.

Start in the artifact. From MakeMyPRD download the PRD and paste your spec, or just talk it through. Describe the thing like you are explaining it to that smart intern. Then iterate by talking. Watch it, react to it, fix it out loud. This loop is the entire skill. You are not coding. You are directing.

When your little thing outgrows the chat window, when it has real screens and real logic, that is when you ask Claude to give you the actual files, the real version you can host somewhere. You are graduating from the sketch to the building.

Step three. Give yourself an undo button. This is Git.

Git is a scary-sounding word for a simple thing. It is a save-history time machine for your project. Every time you save, it remembers. Made a change that broke everything? Roll back to yesterday like it never happened.

GitHub is just the place that history lives online, and a free account is plenty. Why does a non-technical person care about this? Three reasons. You stop losing work. You can always undo. And the day you do hand it to a real developer, everything they need is already in one place, with a record of how it got there.

You do not need to learn scary commands. The modern tools, Claude included, will set this up and save to GitHub for you while you click buttons. You just need to know what it is and that you want it.

The nerdy bit, skip if you want. Git tracks the changes, GitHub stores them in the cloud. “The cloud” is just someone else’s computer that never sleeps. That is the whole mystery.

Step four. Put it on the internet. This is deploying.

Deploying is a fancy word for moving your thing from your laptop, where only you can see it, to a real web address, where anyone can. It used to be the part that made grown engineers cry. Now it is mostly clicking a button.

The free options, in plain terms. GitHub Pages is the simplest if your thing is basic, and it is free. Netlify and Vercel are two popular services that connect to your GitHub (neither of them is paying me to say this btw), and when you save a change, they put the new version online automatically. Both have free tiers that are more than enough to start. You pick one, you connect it, you get a link. That is it.

Now your thing has an address. It is real.

Step five. Know your limits. This is the mechanic part.

Before you let a single real human near it, do the boring safety pass. This is where you stop being the excited builder and put on the mechanic’s hat for five minutes. We will get into the exact checklist in a second.

And here is the honest rule. If this thing is about to touch customers, money, or real data, this is where you stop and bring in a pro. Being your own mechanic means doing your own oil change. It does not mean rebuilding the brakes yourself the night before a road trip with your family in the car.

The plain-English jargon decoder

People throw these words around to sound smart. Here is what they actually mean, so nobody can flame you again.

API. Think of a waiter. You do not barge into the kitchen and cook. You tell the waiter what you want, they bring it back. An API is how one app asks another app for something without needing to know its messy insides. Your weather app does not own the weather. It asks a weather API, like ordering off a menu.

Connector. A pre-made adapter. Somebody already built the bridge between Tool A and Tool B so you do not have to. The little integrations in Zapier or Make are connectors. You do not wire them. You flip them on. Sites like Apify make it stupid simple to connect different technologies. And the best part you just ask your LLM to read the docs and it gets it right most of the time.

MCP. This is the newest one, and the easiest to picture. It is USB-C for AI. It used to be that every tool needed its own special cable to talk to an AI. MCP is the one universal port. It lets your AI plug into your calendar, your files, your Slack, without custom wiring for each one. In plain terms, it is how an AI stops just talking and starts doing things for you. Anthropic, the folks behind Claude, started it, and now most of the field uses it.

Deployment. Already covered, but to lock it in. Moving your thing from your garage to the road.

You do not need to master any of these to build your first ten things. You just need to stop nodding politely when someone says them.

Starting is easy. Continuing to run is the real job.

Remember the wood-paneled office. The sales guy who became a CEO, who never wrote a line of code, whose company changed the day he stopped talking about the work and started showing it.

Here is the part I most want you to walk away with. Vibe coding is not the finish line. It is the starting gun. Anybody can get to a prototype now. That is the easy eighty percent. The hard part is the last stretch, the one where the thing actually ships, holds up with real users, and keeps running after the excitement wears off. That gap, from a cool prototype to something your business can lean on, is where most companies’ AI goes to die.

That last stretch is the whole reason I keep talking about an AI mechanic. Vibe coding gives you the spark. The mechanic builds the engine that survives the trip. The one who ships it, services it, and keeps it alive when the model changes underneath it, when the bill creeps, when the edge cases show up at the worst possible time. Starting got easy. Staying alive is the whole job now.

And if you are an executive reading this with a knot in your stomach, let me take that off you. You do not have to become technical. You do not have to bet the company on a vibe. You start small. You let your people build the cheap, safe stuff and feel a real win for once. Then, when one of those wins earns the right to become real, you bring in a mechanic to ship it and keep it running. That is the whole path. Small enough to start Monday. Safe enough to sleep at night.

When I wrote about all this back in early 2025, I ended on a question I could not answer. When the AI cannot fix what it broke, who do you turn to? This whole piece is my answer. For the garage jobs, you turn into the builder. For the lift jobs, you call a mechanic.

So do not wait. This week, pick one annoying thing in your job and build a Claude artifact or Claude Code app for it. Publish it. Send the link to one person. Feel what that CEO felt.

Then, the day your little thing earns the right to grow up, you will know exactly what to do. And if you want help making it stick, shoot me a DM. I am happy to point you to a good AI mechanic. Some days that is me.

Read more

Somewhere in your company there’s a number that gets a little bigger every month. It’s your AI spend. And if you’re honest with yourself, you can’t actually explain what it’s buying you.

This is the story of the night I learned why. And the three questions that tell you whether you’ve got the same problem.

Last month I sat across from a CEO at dinner. Eight-figure company. Sharp guy. Twenty years in his industry. And he said something I’ve thought about almost every week since.

He said: “Pete, I’m scared that if this AI thing actually works, I’m going to have to fire people who’ve been with me since the beginning.”

That was not the conversation I walked in expecting to have.

I had walked in expecting to close a deal. We’d been going back and forth for months. He wanted what we were building. His team wanted it. The numbers worked. And then we sat down to dinner and I realized the real conversation had never happened yet.

I want to come back to what he said and what I said back. But first I need to tell you something about the bill.

Pull up your AI spend for the last 90 days.

Whatever the number is, write it down.

Now answer three questions. Take your time.

One: Can you name three decisions your company made differently because of that spend? Not “we use AI for drafts.” Not “the team loves it.” Three specific decisions. With outcomes you can measure.

Two: Do you know what your people are actually doing with it? Not in theory. Which workflows. Which departments. Specifically.

Three: When renewal comes, and it’s coming, can you walk into that room and explain what you got?

If you answered no to any of those, I need you to understand something before we go any further.

Your AI budget is not a strategy. It is a bet you have not evaluated.

That’s not a disaster. But it is a clock.

I’ve sat in enough of these rooms to know the script by heart. I can tell you the ending before the lights go down. Here’s how it goes.

The executive makes the bet. Announces it. Distributes the licenses. This is the visible act of leadership. The thing that gets said in the all-hands. The thing that goes in the board deck.

Then the platform goes live. And people use it the way people use any tool they haven’t been trained on. They clean up an email. They summarize a document. Things they could have done without it.

The bill grows. Not because anyone is doing anything wrong. Because nobody built a system to connect the spend to anything that actually matters. The tokens run. The licenses sit there. And the number on the invoice gets a little bigger every month.

Then someone asks about ROI.

There’s no number. There’s a story about the future.

And the contract renews.

I spoke to a leader last week whose company has over 1,800 Claude licenses. She said: “The part I can account for, I know exactly what they’re doing. The rest of it, I have no idea what’s happening there.”

Sit with that for a second. She runs the company. She signs the invoice. And she just told me she’s flying half the plane blind.

The tool works. It always works. What nobody built was the bridge between a license and a result. That bridge is the whole job. And almost nobody builds it.

It’s the most common version of this I see right now.

A few weeks ago Axios reported that one company spent $500 million on Claude in a single month. They never set a limit on the licenses, so their people just kept running. The bill arrived and no human being had ever approved that number.

Five hundred million dollars. In thirty days. By accident.

That’s the loud version. The quiet version is everywhere. Salesforce told the world its Anthropic bill will land near $300 million this year. Meta started capping what its own employees can spend because the internal meter was climbing toward the billions. There’s a word for it now. Tokenmaxxing. Burn as many tokens as you can and call it momentum.

I’ve been calling it something else for two years. But fine. The market finally caught up.

Here’s the part that should keep you up. Those companies have whole finance teams watching the money, and they still got caught. You probably don’t. And the labs are not going to eat the real cost of this forever. The day a token stops being subsidized and starts costing what it actually costs, your invoice doubles overnight. On that day, the gap between what you spend and what you can explain stops being awkward and becomes the line your board circles in red.

So the real question is whether your company is running a smaller, quieter version of the same mistake.

Most are. They just haven’t looked at it that way yet.

Let me tell you how the dinner ended.

So I said something to him. I want to be clear, I had not planned to say it. It just came out.

I said: I’m not your hatchet man. That is not what this is. I’m going to give you wings. What you do with them is up to you. You can fly to the moon. You can use it to fire people. That is your call. I’m just going to give you the thing you keep telling me you want.

He didn’t answer right away. He turned his glass a slow quarter turn on the white tablecloth, the way people do when they’re actually deciding something and not just stalling. The waiter started toward us, read the table, and quietly backed off. I let the silence sit.

Then he looked up and said: okay.

Then we ordered dessert and I closed the deal.

But here is the thing I’ve kept thinking about since that dinner. He was not afraid of the technology. He was afraid of what the technology would reveal. That the work some of his longest-tenured people did, the work he’d built a company around, could be compressed by an order of magnitude. That the math would then make a very specific argument he’d have to answer.

Most executives carrying an AI bill they cannot explain are afraid of the same thing.

Not that it doesn’t work.

That it does.

In the paid section, I’m digging into how we solved this step-by-step

Read more

Going paperless had been a running joke at the company for two years. Nobody believed it would happen.

This was a 100-year-old, 750-person CPG distributor, and I was three weeks into helping them put AI to work. The way I do that is the opposite of what consultants sell. I don’t walk in with a deck and a list of tools to install. I sit with the people who actually do the work, one at a time, and I hunt for the one who has already cracked it. Because in almost every company, somebody has. My whole job is to find that person, understand exactly what they did, and turn it into something the rest of the company can latch onto.

That morning, I found him. The guy who plans the truckloads.

For years his desk had looked like a paper avalanche. Printed orders, sorted by region, paper-clipped into piles so he could build efficient trucks by hand. That was the job everyone agreed AI could never touch.

He shared his screen, and the paper was gone. All of it.

“As of last week,” he said, “I took about two days with Claude, and we completely eliminated paper from our processes.”

Two days. For a problem the company had laughed about for two years.

I went quiet for a second. Then I told him we should hold a moment of silence for what was on the screen.

Because nobody would call this guy technical. He was the furthest thing from a “disruptor” or “early adopter” or the other cringy buzzwords you read about.

He didn’t open a chatbot. He sat down and wrote it a job description.

Not a prompt. A job description.

Three years at the company. Came up through operations. The last tool they handed him was Copilot, and in his words, he “didn’t have much luck.” Same guy, same desk, same kind of AI everybody else got handed. He just did one thing with it that almost nobody does.

So I asked him how.

He didn’t use it. He hired it.

What he told me next changed how I talk about this with every operator I work with now.

He didn’t open a chatbot and start asking it questions. He sat down and wrote it a rulebook.

“I want to filter this, trim it down, give it a set of rules that it can follow and execute on a daily basis.” That was the whole move. He took everything he knew about building a good truckload. Consolidation rules. The cutoff dates. A holding spot for the one-pallet orders that don’t have a home yet. He wrote it all down and handed it to the AI as a set of standing instructions.

Then he tested it like you’d test a new hire on their first week.

“I thought I had it all figured out the first two times,” he told me. “Then about a hundred tries later, we got it to a pretty comfortable place.”

A hundred tries. He didn’t trust it on day one. He checked its work, caught where it guessed wrong, and corrected it. He even taught it to flag its own uncertainty. His words: “I’ll say, wait, don’t assume anything. If you’re going to assume, let me know before you do something.”

That is not how people use software. That is how good managers onboard people.

And it worked. The paper’s gone. His error rate dropped. He’s saving four or five hours a week. He told me, almost surprised, “it just empowers you, you can do three things at once now.”

The Copilot didn’t fail because Copilot is bad. It failed because nobody onboarded it. He didn’t win because Claude is magic. He won because he treated it like a teammate he was responsible for.

The same building. The opposite result.

A few doors down, I sat with someone who’d been at the company seven years. She runs container logistics. Every day she traces dozens of shipments across a pile of port and steamship websites, screenshots the arrival times, color-codes the changes in red inside a spreadsheet, then hands it to her team to retype into the system one record at a time.

It is hours of careful, manual, high-stakes work. If an arrival time is wrong, a container gets missed at the port and the company eats real money in fees.

She had the exact same AI tool installed on her machine. The same one the truckload guy used to clear his desk.

When I asked if she’d tried it, she said, “I have no idea what you’re talking about.”

Same company. Same tool. One person cleared two years of paper.
The other didn’t know the tool existed on her own laptop.

The big new orange icon on her taskbar.

The variable was never the software. They had identical software. One of them got onboarded and the other one got a login link.

This is about to be everyone’s problem

For two years “AI at work” mostly meant a chatbot you opened when you wanted it. That era is ending fast.

In just the last few weeks, Microsoft shipped an always-on agent that lives inside your email and files. Perplexity put an agent on a little computer that runs your tasks around the clock whether you’re there or not. Meta is testing a personal agent some people will pay two hundred dollars a month for. Accounting agents now close the books and reconcile the accounts.

The thing you’re being sold is no longer a tool. It’s a worker. It shows up, it has access, it does things while you’re in meetings.

But there’s a number nobody prints on the launch slide. More than half of employees admit they already bypass the AI tools their company paid for and quietly do the work by hand. A third have never touched them at all. Companies are hiring digital coworkers by the thousand and handing almost none of them a clear freaking job description.

You’d never do that with a person. You’d never give a new hire your bank login, point at the month-end close, and walk out of the building. But that is exactly how most teams are deploying agents. Then they act surprised when the team doesn’t trust it and quietly routes around it.

An agent that nobody onboarded doesn’t blow up. It just fades. It sits there, unused, until someone cancels the contract and nobody notices it’s gone.

The fix is not a better tool. The fix is the thing the truckload guy did on instinct. Write the job description first.

Below is the four-step version I give every operator, plus a prompt you can paste into Claude that builds the whole thing with you, one question at a time.

Read more

I’m sitting in my home office on a video call. The kind where everyone nods along and waits, and you spend half of it trying not to stare at your own face while the Zoom psychosis kicks in.

Four minutes in, this guy goes somewhere else. He hasn’t said a word. He just has a look on his face like he’s about to pass out like a crazy Sabrina Carpenter fan at Coachella.

He’s a longtime friend and a client. An executive at a streaming platform you almost certainly pay for and have been binging for the last decade. He spent fifteen years building a feel for what people actually wanna watch. The kind of taste and discernment most people never develop in a lifetime. He’s in the one percent of the one percent at it.

He’d called me for career advice. Somebody outside his own company who’d give it to him straight. Somewhere in the call I told him what I was building at GPC. That’s when he went quiet.

When he finally spoke, it was the voice of someone who’d already run these lines in the mirror a dozen times while brushing his teeth in the morning.

“I already know I’m training this AI for it to one day replace me. You’re the one with the job security, not me. Can you slow down, please?”

I felt offended. I got defensive. I told him that’s not what I do. I amplify people. I’m not here to cut headcount or shave a few hours off a timesheet. I help companies create real impact and real shareholder value by giving their people superpowers. By elevating what a person is capable of, not reducing them to a stack of tokens and AI slop. Most people get this backwards. A company like his doesn’t need less intellectual horsepower right now. It needs more. And more doesn’t come from AI instead of people. It comes from people plus AI.

He answered slowly.

“The whole purpose of my job now is to tell the algorithm what’s good and what’s bad. When the algorithm can do that without me, I’m gone.”

I didn’t have a good answer.

I gave him the kind of answer that’s true and not useful. He listened politely. We finished the call.

I drove home with the radio off.

That call replayed for a week. I’d walk into a meeting and his line would surface mid-sentence. I already know I’m training this AI for it to one day replace me. I couldn’t shake it because he was right about something I hadn’t admitted to myself yet.

I’d been walking into companies selling adoption and ROI. The numbers that make a CEO nod. What I was actually doing, without seeing it, was feeding the fear my friend was naming. Showing up to make the algorithm better at his job.

Five seconds. That’s how long it took to see it.

I’m not just here to be an EBITDA driver. I’m here to elevate what people are capable of.

Happier people build healthier companies.

Healthier companies create more shareholder value than any AI rollout ever will.

Here’s what I should’ve said to my friend on the Zoom.

Most of the layoffs you’re reading about aren’t really about AI.

Last year, AI got blamed for under five percent of the million-plus US job cuts. Economic conditions and over-hiring got blamed for the rest. A lot of these companies staffed up too fast in the Covid boom, and now they need a reason to cut that doesn’t sound like a mistake. AI is the perfect cover story. Even Sam Altman has said some companies are blaming AI for layoffs they would have done anyway.

And if you dig into the data, AI isn’t even subtracting work. It’s adding it. Somebody sends you an AI-drafted report that looks right, and two hours later a human is still untangling what it actually meant. Every agent you deploy needs a person to babysit it, catch it when it breaks, and own what it ships. The superpower comes with a leash.

You’ve seen the headlines about a new job called the Forward Deployed Engineer. You probably haven’t seen why every AI lab on earth started pouring billions into it in the same quarter. That reason is the whole story. It’s also the answer I owed my friend.

This article is what I wish I’d said to my friend.

Why Forward Deployed Engineer is suddenly the hottest job in tech

Every major AI lab on earth stood up the same thing in the same quarter.

OpenAI launched a four-billion-dollar services arm in May to send engineers into customer environments. Every other lab is racing to match it. Job postings up more than a thousand percent in a year, median base comp near a hundred and seventy thousand, and at Palantir, where the role was born a decade ago, the most senior FDEs clear six hundred thousand a year.

The role isn’t new. McKinsey built its name sending consultants to eat lunch with steel mill foremen back in the 1930s. Anytime a powerful institution needed to land a complex idea inside someone else’s company, the move was the same. Send a person. Have them learn the local way. Build what works for that customer specifically.

What’s new is every AI lab on earth figured this out at the same time. And they all ran into the same wall.

The model alone doesn’t change anything.

A frontier model dropped into your company is a Formula 1 engine dropped in your parking lot. Incredible machine. Goes nowhere until somebody builds the car around it and teaches your people to drive it. The labs learned this the expensive way. Building the model was never the hard part. Getting it adopted inside a real company, by real people, inside the actual workflows, is the whole job. That’s why they’re spending billions to put humans back in the room. MIT looked at hundreds of corporate AI rollouts and found ninety-five percent of them never moved a dollar on the books. Not because the models were weak. Because nobody built them into how the work actually gets done.

So the bottleneck was never the model.

The bottleneck is the senior person on your team who’s gone quiet in the workshops and stopped opening Claude after the first week.

What your CEO peers are missing

I see this pattern in every company I walk into now.

At a client earlier this year, three of the most senior people in the building each found a reason to get me alone. Same week. The CFO is the one I remember. He waited until the room emptied out, walked over, and closed the door with the careful quiet of somebody who’d decided something on the drive in. He sat down across from me. Looked at the wall behind my head like he was reading something written there.

He’d watched me show his leadership team how to do half of his job in an afternoon. He wasn’t angry. He was doing math.

“I’m not afraid they’ll fire me. I’m afraid they’ll move me to fractional. Keep me at half my salary because they can.”

The CEO of that company had told me the day before that the team was excited about the technology. He wasn’t lying. He was on the other side of the door, where executives never hear what their best people are feeling.

The fractional fear is the new version of the layoff fear. It’s more specific and more believable than replacement. It’s the conversation your best senior person is having with their spouse right now.

You think the fear is being fired. It isn’t. It’s being kept around at half of who they used to be. Being asked to teach the machine the parts of their judgment that took twenty years to build, then thanked for their service.

They’ve watched it happen. So they go quiet in the workshops. They use the AI you bought them like a vending machine. They’re protecting what they have, and they’re protecting it from you.

Mark Cuban said something that stuck with me. There are generally two types of people using these tools. The ones who use it to learn everything, and the ones who use it so they don’t have to learn anything. Your scared senior person has become the second kind. Not because they’re lazy. Because nobody ever gave them a reason to be the first kind.

Everyone’s selling you subtraction. The real math is multiplication.

Walk into any AI consulting conversation in 2026 and you’ll hear the same pitch. Cut headcount. Free up budget. Move senior people to fractional. The whole industry is selling subtraction.

They’re doing arithmetic. The real opportunity is geometry.

Subtraction asks how few people you can get away with. Multiplication asks how much more each person can create, and how much better you can connect with the customers you already have. One shrinks the company. The other grows it. Bain checked the math in February. Companies that modernize workflow and workforce together deliver 2.3 times the total shareholder return of the ones that don’t. Same technology. Different scoreboard.

Your most valuable AI asset isn’t the latest model du jour from a frontier lab. It’s the senior strategist who’s been with you since the early days. The principal consultant whose pricing carries the firm’s reputation. The head of client services who knows which account is wobbly six weeks before anyone else does.

Right now what they know lives in their head. When they leave, it walks out the door with them. You watched it happen with the last two partners who got poached by someone who could pay more than you could. One of them was the relationship your biggest account trusted. The other had eighteen years of pricing intuition that wasn’t written anywhere. The numbers you used to bill those clients live at a competitor’s firm now. You promised yourself you’d get it written down. You never did. Nobody ever does.

AI is the first tool that can fix this.

Not by replacing them. By hiring them twice.

Once in person, the way you already have. And once again as something that carries their judgment into the rest of the firm. So when a new account director shows up in week one, they get your principal’s actual frame in their inbox. So when a junior writes a deck at midnight, they have a tool that catches what your senior strategist would have caught. So when your senior person retires or leaves, what they knew stays.

Archimedes said give me a lever long enough and I’ll move the world. The lever in your firm has been sitting at the same desk for fifteen years. AI just made it long enough to move every workflow in the building.

The question is whether your senior person trusts you enough to put their hand on it.

Two ways to do it. Train your own people to run it, or bring in someone who’s done it. Either way it’s four steps, two weeks, and tools you’re already paying for. Step one starts at one specific desk.

First, find your team on the ladder. Six rungs from “smarter search bar” to “team running work in your senior person’s frame.”

R0 Chatter. You ask. Claude answers. Faster than Google. Doesn’t know your business yet.

R1 Prompter. You feed it examples and context. Same Claude, suddenly sharp.

R2 Builder. You teach Claude your business once. Projects, knowledge, memory. It starts to stick.

R3 System Thinker. Skills, checklists, playbooks. Claude runs work in your senior person’s frame.

R4 Connected Operator. Drive, Slack, your CRM. Claude stops asking. It sees what your team sees.

R5 AI Manager. Code, agents, automations on a leash. The firm gets smarter every week.

Most companies stall on rungs one and two. The hire-them-twice play lives on three through five. The teams that get there aren’t the smartest. They’re the most consistent.

The leaders who hear what their senior people actually feel get an unfair advantage. Here’s the full step-by-step playbook below.

Read more

I’m standing in my garage on what I thought was just another Monday when my heart gets ripped out of my chest.

His empty car seat stares at me. Sitting on the ground, discarded.

The same car seat I carried him home from the hospital in. The same car seat that sat in my wife’s SUV for years. The same car seat that once held this tiny little human whose entire body fit between my forearm and wrist.

Now he’s five. And the seat sits there like an artifact from another civilization. Plastic. Lifeless. Crumbs. Sun-faded fabric. Tiny scratches from a life already lived as I gazed at the relic.

That moment hit me harder than I expected. Not because of the car seat.

Because somewhere between building companies, obsessing over craft, chasing growth, surviving downturns, learning AI, rebuilding myself every few years, and trying to stay useful in industries that reinvent themselves every twelve months I felt like that carseat.

My son kept growing. The world kept changing.

And now all of us are standing in this strange new moment where the things we spent decades becoming good at suddenly feels unstable.

Designers. Writers. Developers. Marketers. Artists. Insert your vocation here.

We’re in this together. We’re all having the same existential conversation in our heads right now: If AI can do the thing I do, then what exactly makes me valuable?

That question sits underneath every conversation happening right now, whether you admit it or not. And I think most people are answering it incorrectly. Because the real value was never the artifact.

It was judgment. For years, we confused execution with expertise.

We thought the magic lived in the pixels, the code, the presentation, the copy, the spreadsheet, the deliverable. But the deliverable was never the thing. The thinking was the thing.

The discernment was the thing. The taste was the thing. I know this because I lived it.

I started my career as a designer. And I mean real obsessive designer energy. The kind where you’re nudging typography one pixel left and right at two in the morning because something feels off even though nobody else on earth would notice.

The kind where you spend six hours making a shadow feel invisible. The kind where your girlfriend at the time thinks you’ve completely lost your mind because you keep muttering things like the spacing feels heavy. And honestly?

Maybe we were insane. But that obsession trained something deeper than technical skill. It trained taste.

When I was younger, one of the ways I taught myself design was by recreating work from designers I admired. I’d find a beautiful layout online. Fade the image out.

Then rebuild it from scratch. Not to steal it. To understand it.

To feel the decisions. Why this font? Why that spacing?

Why does this feel expensive while something else feels cheap? Why does this composition create tension? Why does this headline pull your eye first?

That process mattered. Because every great artist starts as an apprentice. Musicians play covers before they write albums.

Chefs recreate dishes before inventing new ones. Painters imitate masters before developing style. Writers unconsciously mimic voices before finding their own rhythm.

Nobody starts original. First you imitate. Then you understand.

Then you remix. Then eventually, if you stay with it long enough, you become yourself. That’s mastery.

Or at least the beginning of it. But somewhere along the way, modern work broke the apprenticeship model. Now everybody wants seniors.

Nobody wants juniors. Nobody wants to train. Nobody wants to slow down long enough to develop people.

And I think that’s a catastrophic mistake. Because what companies are accidentally doing right now is destroying the future supply of masters. Especially now.

Because AI amplifies this problem dramatically. A lot of younger workers are skipping the deep practice phase entirely. They’re consuming outputs without studying the process.

Optimizing for shortcuts before developing discernment. Generating before understanding. And the dangerous part is that the outputs can still look impressive.

That’s the trap. A slick-looking answer is not the same thing as wisdom. Generation is not mastery.

Prompting is not thinking. Access to intelligence is not intelligence. You can outsource execution. You cannot outsource discernment. That distinction is going to define the next decade. A weak thinker with AI becomes dangerous.

A strong thinker with AI becomes extraordinary. And I think we need a new category for this emerging type of person.

Not a specialist. Not a generalist. Generative expert.

Someone who can move fluidly between systems, psychology, storytelling, product, operations, design, technology, behavior, economics and synthesize something useful. Not because they memorized everything. Because they understand how things connect.

That’s the real leverage now. Synthesis. The ability to distill patterns across disciplines and turn them into action.

Honestly, I think this is why some founders feel almost superhuman in certain rooms. It’s not that they’re the best designer. Or the best engineer. Or the best salesperson. It’s that they’re exceptional combiners. They can pull from multiple mental models simultaneously.

A little psychology here. A little systems thinking there. A little storytelling. A little product instinct. A little timing. A little taste.

And suddenly, they see opportunities others miss entirely. That synthesis layer is becoming exponentially more valuable in the AI era. Because AI changes the economics of value creation forever.

The bottleneck isn’t production. The bottleneck is taste. And taste is becoming the last real moat.

And taste is just a compressed experience. That’s why I think we’re entering a renaissance of thought. Not because AI magically made everyone smarter.

But because AI is forcing us to confront how we think. The internet gave us infinite information. AI gives us infinite generation (if you can afford the tokens).

Neither guarantees wisdom. In fact, abundance often destroys discernment. Most frontier models today are intellectually obese.

They contain brilliance, propaganda, genius, mediocrity, insight, spam, manipulation, and contradiction, all blended together into one giant probabilistic soup. Which means the opportunity is no longer information access. It’s signal distillation.

That’s the unlock. The people who win this next era will not be the people with the most information. They’ll be the people with the ability to spot the cleanest signal.

The best information diet. The strongest discernment. The clearest thinking.

The best ability to separate timeless from trendy, insight from engagement bait, understanding from performance, wisdom from noise. That’s why I increasingly believe the future belongs to people who can distill.

Distill ideas. Distill expertise. Distill frameworks. Distill organizational memory.

To distill entire fields of knowledge into reusable systems. That’s mastery. Not the flashy kind. The quiet kind. The kind where a surgeon’s hands know before her brain does. Where a chef tastes once and adjusts three things at once.

Compression. Curation. Connection.

A great teacher compresses decades into moments. A great strategist compresses complexity and jargon into felt-sense clarity. A great storyteller compresses truth into a narrative that moves you.

A great designer compresses emotion into form. And now, AI allows us to do something humanity has never been able to do at this scale. We can distill expertise itself.

Not just information. Expertise. That changes everything.

Because for most of human history, learning from masters required proximity.

For my pro subscribers, I’m including one of my favorite skills.

The one I use to distill and bottle expertise. To cut through the noise and extract the pure signal from any domain.

This isn’t a cute LLM trick. This is the skill I actually use when I need to compress years of someone else’s hard-won wisdom into something I can apply today.

If you’re a thought owner and want help bottling your expertise into an app that your teams and customers can tap into, shoot me a DM.

Read more

Last year, companies bought AI tools the way people panic-buy flashlights before a hurricane.

Fast, emotional, slightly irrational.

Every executive wanted to say they had an AI strategy, so they bought ChatGPT Enterprise, Claude, Copilot, Gemini, and enough SaaS subscriptions to power a small moon landing.

Then something funny happened.

Nothing changed.

Karen still rewrites the same email twelve times before sending it. The sales team still spends Friday afternoon updating CRM fields nobody trusts. Operations still exports CSV files like it’s 2009. Leadership keeps asking the same question:

“Why isn’t adoption happening?”

Because they misunderstand the problem.

They think AI is a software rollout. It isn’t. It’s a behavioral rewrite.

And behavioral rewrites are messy because you’re not installing software. You’re fighting with human inertia.

And let’s be honest. People are absolute pains in the ass when it comes to change.

Especially when they think the thing being introduced might replace them.

I’ve watched this happen inside companies over and over again for the past 23 years as an entrepreneur.

You show someone a workflow that saves them six hours a week and instead of excitement, you get fear disguised as skepticism.

“This won’t work for our business.”

“Our process is different.”

“Clients would never go for that.”

“We tried AI already.”

What they’re really saying is:

“If this works… where does that leave me?”

That’s the real real hiding underneath all this AI resistance right now.

Not tech or platform resistance. Existential resistance.

I had a leadership team recently tell me they wanted company-wide AI adoption. Ten minutes later, one executive started visibly panicking when we showed how quickly AI could summarize operational reports their team spent days assembling manually every month.

You could literally feel the room split in two.

One side saw leverage. The other side saw themselves having to explain to their wife and kids that they don’t have a job anymore.

That tension is the real implementation challenge nobody budgets for.

Because once identity enters the equation, logic leaves the building.
You can’t PowerPoint ninja your way through emotional resistance. You have to design around it.

There’s a strange performance happening inside corporate America right now. Executives are posting screenshots of prompts on LinkedIn like proud parents at a piano recital. Innovation teams are making cinematic launch videos for internal pilots that nobody opens more than twice. Consultants are talking about “AI readiness” while employees keep using Google Docs exactly the same way they did eighteen months ago.

It reminds me of luxury kitchens people pay a fortune for and never actually cook in.

Beautiful appliances. Zero behavioral change.

Most companies are sprinkling AI onto old workflows like Salt Bae sprinkling salt on a steak in 2017. We saw how that story ended.

The winners are redesigning the kitchen itself. And that's when the room gets super quiet. Because rebuilding the workflow means finally naming the thing everyone has been working around for years:

A ridiculous amount of work exists simply because the process exists, not because the outcome matters.

AI is exposing operational junk drawers everywhere.

The smartest AI adoption work I’ve seen lately barely looks like tech consulting at all. It looks like coaching, facilitation, behavior design, communication strategy, and operational anthropology.

You are not deploying software. You are forcing people to renegotiate how they create value.

That’s why most deployments fail. Not because the models are weak. Because the emotional migration plan never existed.

Across leadership meetings lately, one pattern keeps popping up. The companies succeeding with AI are not starting with giant moonshots. They’re starting with the splinter nobody’s pulled out.

The operational equivalent of that one scratchy sweater your aunt knitted you. It serves a purpose, but wearing it is pure, unnecessary misery.

The proposal process takes too long.

The customer support queue everyone secretly hates.

The meeting notes everyone records and nobody reads.

The weekly report somebody spends four hours putting together manually because “that’s how we’ve always done it.”

That’s where momentum starts.

Not with a keynote about AGI changing civilization. Nobody changes behavior because of a keynote. People change their behavior because something is actually in it for them.

That’s the secret.

Tiny relief creates trust. Trust creates experimentation. Experimentation creates adoption.

And adoption is the whole game.

A new category of operator is forming in real time. Not prompt engineers. Not automation consultants.

Operational designers.

People who can look at a company like an X-ray.

Not the org chart. The real skeleton of how work moves.

They can spot the fracture nobody knew existed until the whole body started limping.

They see the bottleneck everyone politely calls a “review process.”

They see the executive who needs to approve everything because nobody wants to own the decision.

They see the team hoarding knowledge inside a 160-tab spreadsheet because it makes them feel useful, safe, and slightly impossible to fire.

They see the project manager manually chasing updates across Slack, email, Asana, and three meetings because the system was never actually designed. It was duct-taped together by whoever cared the most.

They see the red tape everyone calls “governance” because calling it fear would make the room too honest.

They see the weekly report no one reads, built by someone who secretly resents it, sent to people who pretend to care.

Every company has these little ghosts in the machine.

AI does not create them. It reveals them. Then redesign them. That’s the real opportunity.

Because most companies do not need more AI tools. They need someone to help them stop operating like digital hoarders trapped inside bloated workflows from 1999.

The future belongs to people who can reduce operational drag. AI just happens to be the forcing function exposing it.

The weird twist in all of this is that the more AI advances, the more human skills become valuable.

Taste matters more. Judgment matters more. Communication matters more. Storytelling matters more. Trust matters more.

Because as intelligence becomes abundant, discernment becomes scarce.

That changes the economy of value.

The companies that win over the next decade will not necessarily have the smartest models. They’ll have the fastest behavioral adaptation loops.

That’s the game now. Not better prompts. Not AI-flux-capacitor-jargon-word-salad.

Better habits.

Subscribe now

I have run this workshop maybe forty times in the last eighteen months. Twelve operators around a long table. A 200-person company. Everyone in the room has an AI subscription, an opinion about MCP servers, and a deck somewhere that explains why their team is AI-forward.

I know exactly when the room is about to lie to me.

It is always around minute thirty. The bravado kicks in. Two people answer with something they read on LinkedIn last week dressed up as a workflow they actually use. The deck on the wall is fine. The coffee is still warm. Everyone is performing fluency for the founder paying my invoice, and the founder is performing back, and for a few minutes, nobody in the room is telling the truth about anything.

This particular day, the CEO did not wait for me to dig the truth out of his team. He leaned forward and did it for me.

“Tell me one thing AI has actually changed about how you work. Not what you tried. What you do. Different from six months ago.”

Three people speak. Two of them sound like they are reading from a browser tab they kept open. The third pauses, says something honest about a workflow that used to take her four hours and now takes ten minutes, and her voice does a small thing I have heard a hundred times in rooms like this and never had language for until that morning.

The honest one sounded slightly embarrassed.

The performers did not. That was the tell I had been chasing for a year.

I have been running workshops like this one for over two years. The thing that shifted in the last few months is not which tool people picked. It’s who's been getting stronger from these tools and who's been getting good at pretending.

The first group is unrecognizable in twelve months. The second group is wearing the scarlet letter of social media: a green open-to-work badge on LinkedIn.

Three weeks before that workshop, I was on a call with a guy named Mike at a different company. He was not in the room that Tuesday. He is the reason I knew what to look for.

Mike runs operations at a midsize company. He had booked the call because he was burned out. Sick of writing briefs that took weeks to come back, sick of waiting on his team to move on small things, sick of being the person at every meeting holding up the deliverable. I take a few calls like this every month. There is a pattern to them. People show up exhausted, we spend an hour mapping how they could stop submitting requests and start building the thing themselves, and most of them try the smallest idea on the list within a week.

By the time we hung up I had a pretty good read on Mike’s timeline. Most operators ship something small inside a month and start asking different questions inside a quarter. I gave Mike about ninety days.

He gave himself one night.

The next morning he booked another call with me out of nowhere. I jumped on. He shared his screen with the enthusiasm of a six-year-old showing me his Lego project. Then the mockups started loading. Retailer packs with nutrition panels. Display cases with branded shelf strips. Photo-grade product packaging for products that don’t exist yet. Each one polished enough to drop into a buyer presentation that afternoon. All of it grounded in the brand toolkit work we had done together a few months earlier. He told me his team’s mind got blown when he showed it to them.

I didn’t say anything for a few seconds. He let the silence sit. Then he said the line that has been rattling around my head ever since.

“I knocked this out in about an hour. None of these products exist. They’re all AI.”

Mike is a senior industrial engineer with an MBA, known on his team for process excellence. He pays for Claude, ChatGPT, and Gemini out of his own pocket because his company hasn’t given him enterprise AI yet. He’s never written a line of code. The dashboard he opened to start our call was something he had built himself a few weeks back, by joining a few of the messy backend tables nobody had touched in years. He didn’t ask IT for permission. He used what Claude told him to use. He shipped.

When his CEO called him out by name in front of the whole company a week earlier for those mockups, the marketing team pushed back. They had been planning a photo shoot. They were going to need three weeks. Mike got the shoutout and the pushback in the same forty-eight hours. He kept building.

I had to stop the call to write down what I was watching. I have been in dozens of these moments with operators in the last eighteen months and the pattern is the same every time. The room they used to live in (slow approval cycles, three-week creative timelines, managed expectations) suddenly feels two sizes too small for them. They start asking different questions. They stop asking for permission for the small stuff. Then they start getting asked to do the big stuff.

Mike was somewhere in the middle of that arc on the call. He didn’t know it yet. I did, because I have stood next to enough operators on the day they realized what just changed.

Then he said the line I have started writing down whenever an operator says it.

“I prefer to be self-sufficient. I do things on my own without asking for permission.”

That sentence, in the wrong company, gets you in trouble. In the right company, it gets you the job your boss is about to lose. I have watched both happen.

Mike isn’t an outlier. He’s a frontier mind inside a company that hasn’t realized the frontier has moved.

The 30-year veterans at every company you know are quietly building things on the side, in private, without permission, that make their org’s official AI roadmap look like a cave painting that only people who died hundreds of years ago understand. The CEO eventually finds out. Sometimes the marketing team complains. The work ships anyway. And the people in Mike’s seat get a seat at the table they were not previously invited to.

You can spot the Mikes in the next workshop you sit in. They aren’t the ones with the loudest opinions about which model is best. They aren’t the ones leading with bravado. When the CEO asks the question, they are the ones who give a real answer about a real workflow. They are too busy building to perform fluency theater.

This is the gap that will split the next five years of operating careers. The story isn’t access. Everyone has access. The story is who used the machine to produce more output, and who used it to produce more business impact. Only one of those compounds.

In two days we eliminated four years of paperwork

I met Mikaela in her office on a rainy Monday morning, and the first thing I noticed was the paperwork.

She runs operations for one of my clients. She has more stacks of paper on her desk than a public defender’s office, and most of it is the same form printed thirty different ways every week. She had been trying to clean that up for over four years. Her team was too busy. IT could not prioritize it. The consultants she had brought in before me talked a lot about it but never shipped anything she could use. She and her team had spent the previous three months wrestling with Microsoft Copilot and getting back exactly nothing useful.

She showed me the folder of failed Copilot outputs. She had named it Adventures in Disappointment. I laughed harder than I meant to.

Most consultants in my seat would have started by training her on better prompts. I almost did. Then I looked at the Adventures folder again and realized she did not have a prompt problem. She had a patience problem with prompts. She had been treating Copilot like a tool that should already understand her, instead of like a new hire who needed to be walked through it.

So we sat down together on a Monday. We did not write a single line of code. She walked Claude through what she wanted built, the same way she would walk a new hire through it. When Claude got something wrong, she got frustrated the same way she would get frustrated with that new hire on day three. They went over and over and over until Claude started behaving like her. By Tuesday afternoon she had a working system that swallowed the orders, generated the daily report, transferred everything to spreadsheet, and flagged the misfires into a holding section for her to look at.

She shipped it. Then she put together a presentation deck (with Claude’s help, and our brand kit MCP trained on her company’s materials) and walked her entire roadmap into the leadership meeting the following week. Saving four to five hours a day, by her own count. She didn’t tell IT until after.

The pattern that ties Mike and Mikaela together is this article’s whole argument.
The people getting promoted in 2026 are the ones who shipped before anyone asked.

The vending machine problem (aka AI slop)

Most people are not Mike or Mikaela. Most people treat AI like a vending machine.

You drop in a prompt. Out comes a deliverable. You ship it. You feel productive. You move on.

It feels great. You went from a blank page to a finished doc in 90 seconds. Last year that was a four-hour task. You’re winning right?

Except the deliverable is masking what’s happening to you. Every time you outsource a task to the machine before you understand it, you skipped a rep at the gym. The output gets shinier. You get weaker. (I called this dynamic CrapGPT in a recent piece: microplastics for the mind. It is the same idea here.) AI doesn’t take your job all at once. It takes your skills, one at a time. By month nine, the people who used AI as a sparring partner instead of a vending machine have already lapped you, and they are not coming back.

Most people brought their old habits to the new tools. They picked Claude or they picked ChatGPT or they picked Gemini. The output looks better than it did a year ago. The thinking atrophies at the exact same rate. The workshop room I described at the top of this article is full of operators in this exact spot. They picked the better tool. They are still using it the way they used Google in 2009.

Mike does not use AI as a vending machine. He used AI as the thing that turned his taste about packaging into output the marketing team would have spent three weeks on. The mockups aren’t the asset. Mike’s TASTE is the asset. The AI is what made the taste visible to the CEO.

Mikaela does not use AI as a vending machine. She used AI as the partner she could iterate with for two days until the workflow she had been carrying in her head for four years was running on someone else’s keyboard.

Producing and understanding share a verb but nothing else. They use different muscles, build different careers, and only one of them is going to be worth anything in five years.

What this looks like at company scale

A lingerie brand fired their email agency. They plugged AI into their own customer emails. Trained it on their real voice, their real customers, their real way of writing. Inside a year, they had grown email-driven revenue by 60% and doubled their subscriber base. They never went back to a human agency.

That’s what AI looks like as an amplifier of understanding. They didn’t ask the AI to think for them. They taught the AI to think with them.

Now hold that against the version most companies are running. A marketing team I spoke with last year started using AI to “10x their content output.” 🤮. I tried to give them advice and warn them of the dos and don’ts of AI. They didn’t listen.

They did exactly what they set out to do. Ten times more posts, emails, decks, briefs. Six months later, engagement had collapsed by 60%. They couldn’t figure out why.

I’ll tell you why. They had outsourced their writing AND their taste. There was no soul left. AI averages everything ever written into something that sounds correct and feels like nothing. Put your taste in the driver’s seat, AI is rocket fuel. Let AI replace your taste, you’re producing average content at scale, which is the worst possible thing to be producing right now.

Mike has taste. The lingerie brand had taste. The 10x marketing team forgot they had taste. The market is about to drown in acceptable. The only thing that survives that flood is point of view.

Stop asking AI to do your work

I’ll say it the way I say it on every workshop, before we get into the prompts.

Stop asking AI to do your work. Start asking it to show you how it would do your work, and then steal the parts of its approach you don’t have yet.

Vending machine version: “Write me an email to a customer asking them to renew.”

Sparring partner version: “Here is the customer. Here is the relationship. Here is the email I would normally write. Show me three angles I am not considering. Score my draft against each angle. Tell me what I’m missing. Then ask me three questions using the AskUserQuestion tool to capture all needed context before you write anything.”

Same minute of your time, completely different return. The vending-machine person got an email. You got an email AND a sharper read on what your customer needs to hear when their renewal date hits. The next time you write a renewal email, you won’t need the prompt because you internalized the questions.

The AI is supposed to be the thing that builds you up. Use it adversarially. Make it argue with you. Make it grade you. Make it model people you don’t have access to in real life. Compress decades of expertise into hours of apprenticeship. You still have to do the apprenticeship. You don’t become a better runner because you bought a great pair of sneakers.

My company helps people adopt AI for a living. We’re working with companies all over the world. Most of the people we work with are still using AI as a vending machine when we meet. Three weeks later, after they have shipped the first thing they built for themselves, they say something close to what a fintech CEO told me in our last session. “I moved the ball forward more than 12 months in the 90 minutes we spent.” I read that quote on a Sunday morning and put my phone down for an hour.

Tools will keep changing. Whatever is winning right now is gonna be replaced or expanded in eighteen months. The only question that has ever mattered is whether you are using the tool to produce more or to understand more.

Better models won’t fix this for you. The way you talk to the machine has to change first.

My 10-Expert Review

I run this before I publish anything that matters. Decks, emails, proposals, this article. Three minutes of work. It has saved me from shipping something half-baked at least a dozen times.

This is not the casual version of this prompt that has been making the rounds. This is the version a senior prompt engineer would build if they wanted you to ship better work. Steal it.

This version of the prompt uses the structural patterns Anthropic is shipping in their own system prompts right now. XML tags so Claude can find each section. A <thinking> block so the model reasons before scoring. Specific output format so you get something usable, not a wall of prose. Copy the whole thing, fill in the three slots at the top, paste into Claude.

<role>
You are a senior editor running a structured review panel. The work belongs to a smart practitioner who needs honest, specific feedback. Niceness is failure. Vagueness is failure. The goal is to make this piece undeniable.
</role>

<input>
{paste the work to be reviewed}
</input>

<context>
Audience: {who is this for}
Format: {essay, deck, email, sales page, landing page, etc.}
Goal: {what is this work supposed to do for the reader}
</context>

<instructions>
Work through these steps in order. Use <thinking> tags to reason through each step before producing the visible output.

Step 1: Target setting. In two sentences, articulate what "great" looks like for this specific piece given the audience, format, and goal. Make it concrete enough to score against. This is the bar.

Step 2: Panel assembly. Inside <thinking> tags, brainstorm twenty real practitioners who could improve this work. Pick the ten most useful given the input. For each chosen reviewer, output:
  - Name
  - One sentence on what they bring
  - One sentence on what they hate about most work like this

Step 3: Panel scoring. For each reviewer, identify three dimensions specific to their lens (not generic). Score each dimension 0 to 10. Return a markdown table with columns: Reviewer, Dimension, Score, Critique. The critique cell must quote one phrase from the work as evidence.

Step 4: Lowest dimension rewrites. Identify the dimension with the lowest panel average. Each reviewer rewrites the offending paragraph or section in their own voice. Output all ten rewrites side by side, labeled by reviewer name.

Step 5: Synthesis. Inside <thinking> tags, identify the strongest moves from the ten rewrites. Synthesize a single revised version that takes those moves while preserving the original author's voice. Output the revised section.

Step 6: Re-score. Score the revised version using the same panel and dimensions from Step 3.

Step 7: Bounded iteration. If the new average is below 9.0 across all dimensions, repeat Steps 3 through 5 on the next-lowest dimension. Maximum three iterations total to prevent looping.

Step 8: Final output:
  1. The final revised version
  2. A before-and-after scoring table (initial vs. final, by reviewer and dimension)
  3. The single sharpest critique that drove the biggest improvement
  4. One thing the panel said to leave alone, with the reviewer who said it and why
</instructions>

<rules>
- Be specific. Quote my actual words where they are weak.
- Be unkind where the work earns it. Niceness is failure.
- Save my best lines. Mark them in Step 8.
- Do not invent reviewers. Use real practitioners with track records.
- No em dashes. No semicolons.
</rules>

Steal that. Run your last three drafts through it. Watch what happens.

AI is incredible at impersonation. It’s terrible at original judgment. When you ask it to “make this better,” it averages. When you ask ten different reviewers to grade and improve it across specific dimensions, you force it to reach for the specialized knowledge it has buried inside it. You’re not asking the machine for an answer. You’re asking it for a panel.

That tiny prompt change, from “improve this” to a structured panel review with named experts and surgical rewrites, is the difference between vending machine and sparring partner, in one prompt.

You can stop reading here and you’ll still have made today worth it.

In the paid section, I’m walking through three more prompts I use behind closed doors with clients paying me five figures:

• The Adversarial Reviewer. Turn AI into the toughest critic you’ll ever face. Before your CEO does it for free in a room full of your peers.

• The Apprenticeship Method. Model the path of a master, then walk it yourself. The reps stay yours, which is the whole point.

• The Curse of Knowledge Mirror. Extract the things you know that you don’t realize are valuable. The thing founders, consultants, and fractional execs are leaving on the table every day.

Read more

I’m sitting in an overpriced hotel room in San Francisco.

For what I’m paying, I expected something between the Ritz-Carlton and a private equity fever dream. Instead, I got a room with weird carpet, heroic minibar pricing, and Wi-Fi that seems to be powered by one tired hamster in the basement.

I was in town to speak about AI, which meant I should have been reviewing my notes.

Naturally, I went down a microplastics rabbit hole instead.

It started as one of those innocent curiosity clicks. I thought I’d read one article, learn one disturbing fact, feel mildly horrified, and get back to work.

That did not happen.

I kept reading.

Microplastics in human blood. Microplastics in lungs. Microplastics in placental tissue. Microplastics showing up in places that made the whole thing feel less like an environmental story and more like a design failure hiding inside modern life. Researchers are still working to understand the long-term health effects, which somehow made the whole thing feel more unsettling, not less.

The part that stuck with me was the plain-sightness of it.

We built a world around convenience. We wrapped everything in plastic. We drank from it, ate from it, wore it, stored with it, shipped with it, and barely thought about the cost because the cost was invisible.

Then one day, the residue was everywhere.

That is what lazy AI use feels like to me.

People keep calling AI a productivity steroid. Some days, the worse version feels more like Ozempic for the mind. It can suppress the hunger before the idea has been properly fed. It can make you feel done before you have actually digested the problem.

That is a tempting metaphor, especially right now. GLP-1 drugs like semaglutide are being studied for how they affect appetite and reward pathways, and the cultural conversation around them is everywhere.

But the metaphor I keep coming back to is microplastics.

Because the danger is not one shortcut.

It is the buildup.

A little generic email here. A little polished deck there. A few “make this sound more professional” prompts before lunch. A few AI-written posts that sound fine until you realize they sound like everyone else’s posts.

None of it feels dangerous in the moment.

Then one day, everything starts tasting like plastic.

That is CrapGPT.

CrapGPT is what happens when powerful tools meet weak thinking.

It looks useful. It feels productive. It gives you clean sentences, decent structure, and the warm little dopamine hit of having made something quickly.

The work is grammatically fine. The ideas are acceptable. The tone is safe. Nothing is technically broken, which is exactly why the damage is hard to see.

The output has no pulse.

AI is a mirror with a motor

Real talk: AI reveals the quality of what you bring to it.

Bring it a vague idea and it gives you vague output.

Bring it a weak point of view and it gives you polished mush.

Bring it no context and it gives you the internet’s average opinion.

Bring it no taste and it gives you something that sounds like a 2:30 am infomercial that you’re too damn exhausted to turn off.

This is where people get sideways.

They blame the tool for bland work, but the tool is often reflecting a bland brief. They blame AI for killing their brand voice, but the brand voice was never trained. They blame ChatGPT for generic strategy, but the strategy was already a list of business goals in a nicer outfit.

AI does not remove the need for taste.

It raises the cost of not having any at all.

That is the strange part about this moment. A tool that makes output easier also makes judgment more valuable.

When anyone can generate 50 headlines faster than a fart, the edge moves to the person who knows which one is actually true.

When anyone can draft a landing page, the edge goes to the person who knows what the buyer needs to believe to actually convert.

When anyone can produce strategy language, the edge moves to the person who can spot the sentence that sounds smart but says nothing.

That is the job now.

Not making more stuff.

Knowing what deserves to exist.

Speed is not the enemy

I love speed. I’m obsessed with it.

I love shortcuts. I love tools. I love anything that helps a founder, team, or company move from fog to progress faster.

I have spent most of my career around the same kind of pattern. Someone walks in with a messy idea, a half-formed business, a vague product, a brand that almost works, or a story they cannot quite tell yet. My job has usually been to find the thing under the thing.

The real idea.

The sharper promise.

The hidden tension.

The phrase that finally makes the room lean in.

AI is amazing for this.

It can help you research faster. It can help you turn a transcript into themes. It can help you map objections. It can help you turn a sales call into content, a founder rant into a manifesto, a meeting into a plan, and a messy note into something worth building.

That is real.

The trap starts when speed becomes a way to avoid understanding.

You skip the customer.

You skip the tension.

You skip the belief.

You skip the positioning.

You skip the scar tissue.

You skip the taste.

Then you ask the machine to make it “more compelling.”

That is how average spreads.

Not with one bad prompt. With hundreds of tiny acts of cognitive convenience.

You can outsource the work

You can use AI to write a draft.

You can use it to summarize a meeting.

You can use it to build a first version of a product spec.

You can use it to generate code, inspect data, write outreach, rewrite messy notes, compare competitors, analyze reviews, and prepare for a sales call.

You should.

Used well, AI is a force multiplier. It reduces drag. It compresses cycle time. It helps people get unstuck. It gives junior people a thinking partner. It gives senior people a way to scale their judgment.

But there is a line that matters.

You can outsource the work.

You cannot outsource the understanding.

You still need to know what good looks like.

You still need to know what you believe.

You still need to understand your customer’s fear, pressure, language, and desire.

You still need to know why your offer matters.

You still need to hear the difference between a sentence that is clear and a sentence that is alive.

That part is yours.

That part is the value.

The prompt library is not the moat

A lot of people are collecting prompts right now like Pokémon cards.

They have prompts for writing.

Prompts for strategy.

Prompts for sales.

Prompts for naming.

Prompts for design.

Prompts for leadership.

Prompts for thinking about prompts.

Some of that is useful. I have written plenty of prompts. I believe prompts can become intellectual property when they encode real taste, process, and judgment.

But a prompt library full of generic templates is just a junk drawer with better formatting.

“Write in a professional yet approachable tone.”

That is beige wallpaper.

“Make this bold and punchy.”

That is not voice.

“Act as a world-class marketer.”

That is cosplay without context.

Voice is not a list of adjectives.

Voice is a way of seeing.

It is rhythm. It is restraint. It is what you notice. It is what you refuse to say. It is the sentence you write because you lived the scar behind it.

If you want AI to sound like you, stop describing your voice like a mood board.

Train it on your thinking.

Feed it your actual writing. Your best sales calls. Your founder notes. Your messy rants. Your strongest beliefs. Your customer conversations. Your phrases people repeat back to you. Your standards. Your taste.

That is where the useful output starts.

Not from asking AI to sound “human.”

From giving it a human worth learning from.

The real advantage is reusable understanding

A founder who uses AI to write one LinkedIn post saves an hour.

A founder who uses AI to extract their beliefs, map their audience, study their sales calls, clarify their offers, and build a voice system creates an asset.

Same tool. Different game.

This is where most companies are still missing the point. They ask how to use AI to make more content, more decks, more emails, more posts, more everything.

The better question is:

How do we make the company smarter every time work happens?

That is the unlock. That’s what hits different.

Your sales calls should make your messaging better.

Your support tickets should make your product better.

Your meetings should make your operating system better.

Your founder notes should make your content better.

Your client work should make your frameworks better.

Your best people should make your training better, even when they are not in the room.

That is what AI can help with.

The goal is not more output.

The goal is reusable understanding.

The compound understanding stack

Here is the simplest way I think about it.

Start with raw material, not output.

Most people open AI and ask for the finished thing. That is why the finished thing feels empty. There is no evidence inside it.

Better inputs create better work. Your inputs drive your outputs and your outputs drive your outcomes. Choose wisely.

Capture

Capture what is already happening.

Sales calls. Customer questions. Founder rants. Team meetings. Support tickets. Old decks. Voice notes. Strategy docs. Slack threads. Website copy. Reviews. Objections. Follow-up emails. Internal debates.

Most companies are sitting on buried intelligence.

It lives in the head of the founder.

It lives in the best salesperson’s instincts.

It lives inside the customer’s exact words.

It lives in the team member who has answered the same question 400 times.

AI becomes more useful when it gets evidence instead of vibes.

Clarify

Once the raw material is captured, use AI to find the signal.

Ask better questions:

What pain keeps repeating?

What objections show up most?

What words do customers use when they describe the problem?

Where does our pitch create energy?

Where does it create confusion?

What do our best customers already believe?

What are we saying that nobody cares about?

What is the real decision the buyer is trying to make?

This is where AI starts to feel less like a writing tool and more like a thinking tool.

It helps you see patterns you were too busy to notice.

Challenge

This is the step people avoid because it bruises the ego.

Ask AI to attack the work.

Not in a cute way. In a useful way.

Ask:

What is weak here?

What sounds generic?

What claim is unsupported?

What would a skeptical buyer reject?

Where are we pretending to be different?

What would a sharper competitor say?

What part of this strategy is lazy?

This is where the work gets better.

Most teams use AI to make weak thinking sound nicer.

Better teams use AI to make weak thinking harder to ignore.

Convert

Then turn the thinking into assets.

A better offer.

A sharper pitch.

A stronger landing page.

A sales script that slaps.

A training doc that doubles employee retention.

A founder essay that accelerates recruiting.

A product spec that fuels greatness.

A workshop that connects humans with heart.

A decision memo.

Insights should not die in chat windows.

If the thinking is useful, turn it into something the business can use again.

Compound

This is the part almost everyone misses.

Save the best thinking.

Save the prompts that worked. Save the examples. Save the customer language. Save the decisions. Save the standards. Save the before and after. Save the critique.

Build a company memory.

The goal is not to have a thousand chats.

The goal is to make the company smarter every time it works.

That is how AI moves from toy to operating system.

This is what winning with less effort actually means in the intelligence age

Less effort does not mean less care.

It means less waste.

Less repeated thinking.

Less starting from scratch.

Less rewriting the same pitch.

Less onboarding through folklore.

Less losing the customer’s exact words after every call.

Less letting good ideas rot in random notes.

Less turning every project into a fresh act of chaos.

That is the real promise of AI for people who do serious work.

Not “do nothing and get rich.”

More like:

Think once. Use many times.

Learn once. Train the system.

Hear the objection once. Improve the pitch forever.

Solve the workflow once. Make it repeatable.

Find the language once. Use it across the funnel.

That is how you win more with less effort.

You do not lower the standard.

You lower the friction.

A simple place to start

Start with your next sales call.

Record it. Transcribe it. Put it into AI. But don’t stop at the generic notetaker summary - that’s what amateurs do.

Instead, ask for these five things:

What did the buyer care about most?

What objections were spoken or implied?

What exact phrases should we reuse?

Where did our explanation create energy?

What follow-up assets should we create from this conversation?

Now turn that one call into a better follow-up email, a sharper FAQ, a sales page section, a content post, and a stronger discovery question.

That is one call.

Do that with 50 calls and you are no longer guessing at your market.

You are building market intelligence.

This is the difference between content creation and understanding creation.

One fills the calendar.

The other improves the business.

Another place to start

Take your last ten messy notes.

Not the polished ones.

Use the ugly ones.

The voice notes. The fragments. The half-formed thoughts. The “this might be something” lines. The ideas you wrote down because they would not leave you alone.

Drop them into AI and ask:

What do I keep circling?

What belief is forming?

What am I avoiding?

Which idea has commercial potential?

Which phrase sounds ownable?

What should become a post?

What should become a product?

What should I kill?

This is where personal knowledge work gets interesting.

Most people do not need a second brain.

They need a better bridge between their brain and the market.

AI fluency is not tool collecting and “innovation theater”

Knowing 300 AI tools does not make you fluent.

That is tool tourism.

AI fluency means you can move from problem to system to outcome.

You know when to use AI for speed.

You know when to use it for depth.

You know when to trust the output.

You know when to challenge it.

You know when the work needs a human.

You know when the business needs a process instead of another prompt.

The best AI users I know do not just ask better questions.

They build better loops.

Research loops.

Writing loops.

Sales loops.

Learning loops.

Decision loops.

Product loops.

That is the difference.

Amateurs collect prompts.

Pros design loops. Be a f*cking pro - you and your team deserve it.

Average work is about to get very cheap

This is the part many people still do not want to face.

AI makes average work easier to produce.

Average writing. Average design. Average code. Average strategy. Average outreach. Average decks. Average product ideas.

The market will drown in acceptable.

That means acceptable is now worthless.

The bar will move. AI slop will infect us from everywhere like those microplastics I ranted about above.

The question will no longer be, “Can you make the thing?”

The better question will be, “Can you make the thing matter?”

Can you bring taste?

Can you bring context?

Can you bring judgment?

Can you bring a point of view?

Can you bring trust?

Can you make something people feel?

That is where humans still win.

Not by pretending the machine does not exist.

By bringing better inputs into it.

The future belongs to those with sharper points of view and better inputs

Most people obsess over outputs.

Better headlines. Better decks. Better posts. Better images. Better code. Better scripts. Better landing pages.

Outputs are visible, so they get the attention.

Inputs decide the game.

Your taste is an input.

Your beliefs are inputs.

Your customer insight is an input.

Your lived experience is an input.

Your company memory is an input.

Your standards are inputs.

Your questions are inputs.

Your constraints are inputs.

Your courage is an input.

The machine multiplies what you bring it.

Bring it something worth multiplying.

I do not want AI to make us less human

I want it to make us more honest.

More direct.

More creative.

More aware of what we actually believe.

More capable of turning ideas into progress.

More willing to build the things that used to stay trapped in our heads.

That is the version of this I care about.

Not automation for the sake of automation.

Not more generic AI slop..

Not more safe circle jerk posts about “the future of work”.

I want tools that help humans become more expressed, more capable, and more free.

I want founders to turn their messy thinking into healthy thriving companies.

I want experts to turn their knowledge into products.

I want teams to stop losing their best ideas inside meetings nobody revisits.

I want companies to capture what their best people know before it walks out the door.

I want AI to help us move faster without becoming soulless lobsters 🦞 .

Speed without slop. That is the new standard. I will die on that hill.

So yes, use AI

Use it aggressively.

Use it every day.

Use it to research, write, build, sell, learn, and make your company faster.

But do not let it turn your mind into plastic.

Do not let it sand down your voice.

Do not let it replace your taste.

Do not let it become a shortcut around understanding.

The people who win will not be the ones who think less.

They will be the ones who understand faster.

They will build company memory while others make content.

They will build systems while others collect prompts.

They will build taste while others chase templates.

They will use AI to become more dangerous, not more average.

That is the opportunity.

You can outsource the work.

You cannot outsource the understanding.

It’s 6:47 pm on a Tuesday. I close my laptop. My Mac Mini doesn’t.

While I’m running a hostile takeover on a bowl of guacamole (I don’t cook, I conquer), my wife and son as eyewitnesses, a small box in the other room is rummaging through my day.

Every email I sent. The ones I ignored. The meetings I sat through. The files I touched. My git history. Every artifact of a workday most operators throw straight in the trash can or worse, forget about and never look at again.

I call it my digital exhaust.

It’s looking for one thing. How do I make tomorrow better than today?

By the time my kid grossly double-dips his dirty fingers in my guac, my phone buzzes.

Three questions. It’s my chief of staff agent pinging me.

What worked today? What didn’t? What would I do differently tomorrow?

Ninety seconds. I answer. The system updates itself while I get back to dinner.

Same night, 3,000 miles away, Uber’s CTO is on the phone with a reporter. Nine words from him get screenshot across tech Twitter by morning.

“I’m back to the drawing board, the budget is gone.”

Four months in. Whole year. Vapor.

Praveen Neppalli Naga. 5,000 engineers. The smartest AI tools on the planet. And he still couldn’t see it coming.

A few weeks before that call, I sat down for a consult with a mid-size strategic communications agency. Three hundred people. Microsoft shop. Their CFO had approved Copilot Enterprise across the company because, well, Microsoft. Bundled discount. Looked like the responsible move.

Almost nobody used it.

Then I pulled their expense reports.

Half the senior team was paying for ChatGPT Plus out of pocket and expensing it as “professional development.”

Two AI strategies running in parallel. One on the books. One in the shadows. Both billed the company.

Their CFO wasn’t wrong. He just couldn’t see what was actually happening.

Uber had 5,000 engineers and a finance team and they couldn’t see it. This firm had a Microsoft contract and an expense system and they couldn’t see it.

I have a Mac Mini and a ritual to keep me honest on AI token spend.

While Silicon Valley is token-maxing. I’m outcome-maxing.

It’s not how many tokens you burn. It’s whether the tokens count. Whether they compound into better thinking, better work, better calls.

AI is the gear. I’m the guide.

Guess which one of us knows what tomorrow costs.

Here’s the part nobody has the guts to say to your face

You don’t have a tools problem. You have an awareness problem.

Right now, somewhere in your stack, an AI is running. A teammate has ChatGPT open. Notion just bumped your seat price because they turned on AI by default. Slack added an AI tier you don’t remember approving. Three of your contractors are doing client work in personal Claude accounts. Your CRM has an embedded copilot drafting replies in your voice.

You can’t see any of it.

That’s not a failure of attention. That’s a failure of architecture. The tools were not designed to show you what they cost or what they’re worth. They were designed to spread.

So let me ask you the question your CFO is going to ask you in 90 days. How much did your company spend on AI last month? Not the obvious subscriptions. The whole picture. Embedded tiers, API calls, agent loops, shadow tools. The full iceberg.

You don’t know. Almost nobody does. Naga didn’t, and he had a finance team.

Here’s the line I want you to sit with.

Spend that doesn’t produce learning is just burn. Spend that does is tuition.

Most operators are running a credit card. The good ones are running a practice. The shift from one to the other is not about a tool. It’s about a ritual.

The Four big Shifts (and wtf to do about it)

You don’t need a doctorate in this. You need four simple moves.

1. Tokens are the new salary line.

At GTC last month, Jensen Huang said he wants to give his engineers $250K in tokens on top of a $500K base. His exact line.

“If that engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed.”

Translation. He’s not paying for a tool. He’s paying for amplification. Most operators are still treating ChatGPT like Netflix. also funny coming from him, given their revenue model.

What to do tonight. Open Claude Code. Type /cost. It tells you exactly what your last session cost in tokens and dollars. Most people have never seen this number. Run it for a week. You’ll suddenly know which kinds of tasks are cheap and which are quietly expensive. That’s the start of awareness.

2. The forecast is broken.

April 15. Naga admits to The Information that Uber’s entire 2026 AI budget is gone. Four months in. Engineers were spending $500 to $2,000 a month each on Claude Code. 5,000 of them. Math math math. Vapor.

If Uber couldn’t see it coming, your team can’t either. The only defense is real-time awareness, not month-end reconciliation.

What to do tomorrow. Run this prompt on a CSV export of your last 90 days of card statements:

Act as a CFO doing an AI audit. Pull every line item that is or could be an AI tool, AI tier, or API spend. Group them. Flag any subscription where AI was added or upgraded in the last 6 months. Tell me my total exposure and the three biggest leaks.

You’ll find at least one tool you forgot you were paying for. Probably three.

3. Polite Claude is expensive Claude.

A meme went viral last month. A dev taught Claude to talk like a caveman to save tokens.

“Why use many token when few token do trick.”

Funny. Also a real lesson hiding inside the joke. The model’s default voice is tuned for human approval, not your wallet. Most of your savings live in your prompt, not your tool.

What to do right now. Add five words to the top of your most-used prompts: “Be concise. Return JSON.” Or skip the JSON if you want prose. The point is, tell the model what shape you want and what to drop. Watch your token usage drop on that workflow.

Try /context next. It shows you exactly what’s in the context window of your current Claude Code session. Most operators are flying blind. This is x-ray vision.

4. The frontier is not always the answer.

Small models are eating the boring 80% of what your business actually does. The research keeps stacking up. A fine-tuned 7B model trained on one workflow can match a frontier model on that workflow at a fraction of the cost. Same for SQL queries. Same for classification. Same for tagging.

You’re paying Ferrari prices for what a used Prius can do for you.

Look at your top three repeat AI tasks. Things you do the same way over and over. Ask yourself, does this need a genius, or does it need a specialist? If it's repetitive, it's a candidate to route to a smaller, cheaper, faster model. You don't have to switch tomorrow. Just notice the pattern.

The practice that ties it together

Four shifts. Four moves. Useful on their own. The thing that compounds them is the practice.

End of your workday, in Claude Code, run /insights. It generates an HTML report of everything you worked on that session. Most users have never run it. The first time you do, you see your own patterns.

Then take it one step further. Build a tiny custom command that does this every day at 6pm. It reads your insights report. Your git history. The meetings you sat through. The emails you sent and received. The files you touched. Your digital exhaust.

It analyzes it. Then it pings you with three questions. What worked. What didn’t. What I’d run differently tomorrow.

Ninety seconds. You answer. The system updates. The next day starts smarter than the one before.

That’s the loop. Spend in. Learning out. Receipts on both sides.

This is the operating system Naga didn’t have. It’s the one Silicon Valley is too busy to build because they’re token-maxing instead of outcome-maxing. But you’re probably not living on that sweet sweet VC money. For you, EBITDA probably actually matters, and guess what the thing that’s gonna cut into your bottom line? It’s gonna be those surprise AI bills that you’re not planning for in your P&L.

It’s available to you tonight. The commands are free. The discipline is the part nobody talks about.

The open loop

If you’re reading this and thinking “I want this system but I don’t know where to start,” my DMs are open.

In the paid section I’m gonna walk you through:

• The Token Triage Framework. The 80/15/5 mental model. Three buckets, a 30-second test, the question to ask before you fire up Opus and make your CFO hate you.

• The EOD skill. The actual markdown file I drop in to fire my 6:47pm reflection. Plus three ways to trigger it: manual, cron, event-driven.

• Five prompts I run weekly. CFO Audit, Concise Generator, Routing Decision, Weekly Stack Review, Skill Gap Detector. Copy-paste ready.

• The CLAUDE.md template. My actual file. The one that tells every Claude session who I am, how I work, and what to never do.

• The SLM routing cheat sheet. When to swap a frontier model for a small one. Print it and tape it to the wall.

Read more

I have a confession.

I’ve been living in a bubble.

I realized this last Tuesday. I was on stage in front of a few hundred people. Big company. The kind where the lobby has a waterfall and the badge lanyards have their own font.

They’d brought me in to talk about how AI is changing the way teams work. I had my slides. I had my demos. I had this whole bit about compound AI workflows that I’d been workshopping with my AI friends over overpriced bourbon the week before.

I opened with something I thought was a throwaway. A little warm-up before the real stuff.

I pulled up a tool I use every day and showed them how I’d turned a four-hour Monday report into a twenty-minute automated task. Nothing fancy. Just an AI that reads a few reports, chews through a simple markdown skill file that tells it exactly who I am and what I want, and spits out a formatted deliverable while I’m still brushing my teeth.

It’s like having a border collie that does spreadsheets. Focused, obedient, a little obsessive. I set it up in twenty minutes on a random Tuesday. I barely think about it anymore.

I finished the demo. Feeling good. Ready to get into the deep stuff.

“Any questions before we go further?”

A hand went up in the third row.

“Yeah. How do you write a good prompt?”

I stood there. Blinked twice. Smiled. And felt something shift inside my rib cage.

Not here. Not at THIS company.

I had forty-five minutes of material on reinforcement learning, context architecture, and agent orchestration. The kind of stuff I argue about with my AI friends the way normal people argue about whether LeBron is the GOAT. Vector databases. Fine-tuning. Token economics. We get heated about this stuff at dinner.

But this room wasn’t that dinner.

This room was a few hundred smart, accomplished, successful people who hadn’t started yet.

Not because they weren’t capable. Because nobody had ever pulled up a chair and shown them where the door was.

I spent the rest of the session right there with them. On their phones. On their laptops. In their heads. I watched a room full of people who’d been told AI was coming for their jobs actually BUILD something with it for the first time.

You could feel the energy change. Like the whole room exhaled at once.

I drove home that night and couldn’t stop thinking about it.

I’d been living in a bubble. A bubble where everyone uses AI twelve hours a day and knows what MCP stands for and has opinions about token limits. A bubble where I have three Mac Minis in my office running four Claude Max accounts, and I spent last weekend letting OpenClaw and something my buddy named “Fart Box” battle each other to see which one could triage my email better. (OpenClaw won, but Fart Box had better instincts on the weird stuff.)

That bubble is maybe a few hundred thousand people worldwide. Maybe a million if you’re generous.

The other billion people using AI tools? We’re still using LLMs like a better version of Google search, asking crazy simple questions like “How do I write a good prompt?” and “What does this feature do?”

And most of what we find is jargon, hype, and someone trying to sell us a course.

This article is for them. And honestly, it’s for me and you too. Because we all need a reminder.

The One Thing

Here’s what I tell everyone now.

Forget the automation videos. Forget the “I replaced my entire team with AI agents” posts. Forget the people running twelve autonomous systems on a stack of Mac Minis in their closet with name-tag stickers on them.

Start with one thing.

One task. One thing you do every week that’s repetitive and boring and makes you wonder why you went to college.

Maybe it’s the report your boss reads in two minutes while he’s on the toilet, texting his wife who already knows he won’t be home for dinner again tonight. And it takes you four hours to build.

Maybe it’s triaging your inbox first thing Monday morning.

Maybe it’s compiling meeting notes into a summary nobody asked for but everyone expects.

Pick one. Just one.

Here’s why.

Everyone from Gartner to Pendo (yes, Pendo, the software nerds who literally count feature usage for a living because somebody has to) to McKinsey keeps finding the same thing: we use about 20% of the features in the software we already pay for. Twenty percent. The other 80% just sits there. Billions of dollars in capability, collecting dust.

You’re driving a spaceship like it’s a bicycle.

If you have a Claude subscription, or ChatGPT, or Gemini, you already have what you need. You don’t need another tool. You need to use the one you’ve got for one specific thing and see what happens.

The Amnesia Problem (And How to Fix It)

Here’s the thing that frustrates most people when they first try AI: it forgets everything. If you want the nerdy answer to how this works, view my AI principles essay

You have a great conversation on Monday. You explain your job. You explain your company. You give it context. It writes something brilliant. You close the tab.

The next morning you come back and it has no idea who you are.

This is the context window problem. Think of it like hiring someone brilliant who gets total amnesia every time they go to sleep. Every morning you have to reintroduce yourself.

The fix is simpler than you think.

Most AI tools now let you save context so it persists. In Claude, it’s called Projects. You create a project, you drop in your files, your brand guidelines, your style preferences, whatever makes you you. Every conversation inside that project starts with that context already loaded.

(They’ve also recently added memory, which is getting better at remembering things about you across conversations. It still has a long way to go, but it’s improving fast.)

Some people take this further and create what are called markdown files. Plain text files that describe who you are, what your company does, what you like and don’t like, what good output looks like. It’s like writing a really good onboarding document for a new hire. Except the new hire has photographic memory and works at the speed of light.

Once you tell it who you are, it remembers. And every conversation gets better.

I’ve built skills that can interview you and create these files automatically (how to in the paid section). But honestly, even a messy paragraph about who you are and what you need is better than starting from zero every time.

Your First Scheduled Task

OK. Here’s where it gets real.

I’m going to walk you through the exact thing I showed those 150 people. The thing that made three of them gasp.

If you have Claude Pro ($20/month) and the Claude desktop app, you have access to something called Cowork. Think of it as a coworker who lives inside your computer. It can read your files, write new ones, and follow instructions without you watching. (Getting started guide)

Step 1: Open Claude desktop. Create a new project. Call it whatever your Monday task is. “Weekly Sales Summary.” “Expense Report.” “Team Update.”

Step 2: Drop in the files it needs. Last week’s report. A template. The raw data. Whatever you normally pull from three different places to build this thing.

Step 3: Tell it what to do. Not in code. In English. Here’s a prompt that actually works. This isn’t a basic one. The last two lines are what separate good output from garbage:

Here are three files: last week’s sales data, our report template, and our target numbers. Please:

1. Pull the key numbers from the sales data
2. Fill in the report template with this week’s figures
3. Flag anything that’s more than 10% above or below target
4. Save the completed report as a new file

Keep it factual. Don’t write commentary or analysis. I’ll add that myself.

Before you finalize: run a chain of verification. Challenge every number you pulled. Cross-reference each figure against the source data. If anything doesn’t match or seems off, flag it and show me the discrepancy. I’d rather have a report with questions than a report with wrong answers.

Step 4: Here’s the part that made people gasp. Type /schedule in the chat. Tell it to run every Monday at 7am. (How to schedule recurring tasks)

That’s it. Monday morning, you open your laptop, and the report is sitting there. Done. You didn’t touch a spreadsheet. You didn’t log into three systems. You just open the file and add your commentary. The thinking part.

One thing. Twenty minutes to set up. Saves you hours every week.

And here’s what I didn’t expect: once people do this ONE thing, they start seeing everything differently. “Wait, could it do my expense reconciliation too?” “What about the board deck?” “Could it triage my email?”

Yes. Yes. And yes. But start with one.

“But It Makes Mistakes”

It does. And I’m gonna reframe something for you that changed how I think about AI.

Hallucination is not a bug. It’s actually how the technology works. These models predict the next word (or token, as engineers call it) based on patterns. Sometimes the pattern leads somewhere wrong. Especially when you haven’t given it enough context or shown it what “good” looks like.

Think about it this way. If you hired a brilliant junior employee on their first day and said “write me a report” with no template, no examples, and no context about your company, you’d get garbage. That’s not because the person is dumb. It’s because you gave them nothing to work with.

Same with AI. Show it two or three previous reports. Tell it what good looks like. Tell it what to avoid. This is what prompt engineers call “few-shot prompting” but it’s really just good management. Give examples. Be specific. Review the output.

The people who say “AI doesn’t work” are the same people who would hand a new hire a laptop and say “figure it out.” It’s not the tool. It’s the onboarding.

A Quick Note

Some of you wrote in and asked me to write about this. To go back to basics. To stop assuming everyone lives in the same bubble I do. Thank you. I needed that wake-up call and honestly the keynote gave it to me too.

I’m obsessed with helping people become AI-first practitioners without losing what makes them human. The curiosity. The judgment. The weird creative leaps that no model will ever make.

AI handles the dull and mechanical so you can do more of the stuff that actually requires you.

If you’re just starting, you’re not behind. The people automating everything had to start with one thing too. They just did it six months ago.

Your turn.

Here’s what’s in the paid section:

• The specific use cases broken down by industry (marketing, finance, healthcare, real estate, legal, education, e-commerce, manufacturing, HR, consulting).

• A step-by-step walkthrough of the “always-on AI” setup that people are buying Mac Minis for ($600, $15/year to run, works while you sleep).

• My personal projects I built with AI for my family.

• And a complete learning syllabus from “I’ve never used AI” to “I’m building autonomous agents.” Not links to Google. Actual courses, organized by level, that I’ve personally vetted.

Read more

I’ve been sitting with this one for a while. And I think it’s the most important thing I’ve read this year.

Jack Dorsey published an article called “From Hierarchy to Intelligence.” Then he sat down with Sequoia and basically said: the org chart is dead. Not dying. Dead. And he’s rebuilding Block around that idea. Cut 40% of the company. Three roles. AI at the center. Humans on the edge.

My first reaction? That’s bold. My second reaction? He’s right.

Let me break down what I learned and why I think every founder and operator needs to pay attention.

The 2,000-Year-Old Operating System

Here’s the thing. The way we organize companies hasn’t fundamentally changed since the Roman Army. Seriously.

The Romans figured out that one leader can manage about three to eight people. So they built layers. Eight soldiers share a tent with a leader. Ten of those groups get a centurion. Six centuries make a cohort. Ten cohorts make a legion.

Information flows up. Decisions flow down.

That exact pattern is what your org chart looks like today. Swap “centurion” for “director” and “legion” for “business unit.” Same thing.

The Prussians added middle management after Napoleon destroyed them. The railroads brought it into business in the 1850s. Frederick Taylor optimized it with scientific management. McKinsey formalized the matrix. Spotify tried squads. Zappos tried holacracy. Valve went flat.

Every single one of those experiments hit the same wall. Past a few hundred people, you need hierarchy because humans are the only coordination mechanism available. More people means more layers. More layers means slower information flow.

That’s the trade-off nobody could break. Until now.

What Block Is Actually Building

So here’s what Dorsey is doing. And this is where it gets interesting.

Block is remote-first. Every action creates an artifact. Slack messages. Pull requests. Google docs. Recorded meetings. Code commits. All of it is machine-readable data about how the company operates.

Traditionally, a manager’s job is to know what’s happening across their team and relay that context up and down the chain. Dorsey’s argument is simple: AI can build and maintain that picture continuously. What’s being built. What’s blocked. Where resources are going. What’s working and what’s not.

That’s what the hierarchy used to carry. Now an intelligence layer carries it instead.

He breaks the company into four layers:

Capabilities. The atomic building blocks. Payments, lending, card issuance, banking, payroll. These aren’t products. They’re primitives. No UIs of their own. Just reliable, compliant, performant building blocks.

World model. Two sides. The company world model is how Block understands itself. Performance, priorities, operations. The customer world model is per-customer, per-merchant understanding built from real transaction data. Money is the most honest signal in the world. You can lie on surveys. You can ignore ads. But when you spend, that’s the truth.

Intelligence layer. This is the part that composes capabilities into solutions for specific customers at specific moments. A restaurant’s cash flow is tightening before a seasonal dip the model has seen before? The intelligence layer composes a short-term loan, adjusts the repayment schedule, and surfaces it before the merchant even thinks to look for financing. No product manager decided to build that. The capabilities existed. The intelligence recognized the moment.

Interfaces. Square, Cash App, Afterpay, TIDAL. These are delivery surfaces. Important, but not where the value lives. The value is in the model and the intelligence.

When the intelligence layer tries to compose a solution and can’t because a capability doesn’t exist? That failure signal IS the roadmap. The customer’s reality generates the backlog directly. Not a product manager’s hypothesis.

Three Roles. That’s It.

Dorsey is normalizing down to three roles across the entire company:

Individual Contributors (ICs) who build and operate. Engineers, designers, salespeople, operators. The world model provides the context that a manager used to provide. So ICs can make decisions about their domain without waiting to be told what to do. Durable human skill: judgment, taste, creativity.

Directly Responsible Individuals (DRIs) who own specific cross-cutting problems and customer outcomes. A DRI might own merchant churn in a specific segment for 90 days with full authority to pull resources from multiple teams. Durable human skill: ownership and accountability.

Player-coaches who combine building with developing people. They replace the traditional manager whose primary job was information routing. A player-coach still writes code or builds models or designs. They also invest in the growth of the people around them. They don’t spend their days in status meetings and alignment sessions. The world model handles alignment. The DRI structure handles priority. The player-coach handles craft and people. Durable human skill: building human capacity and empathy.

No permanent middle management layer. The system coordinates everything else.

The Part That Hit Me Hardest

Dorsey said something in the Sequoia interview that I keep coming back to. He said:

If the answer is nothing? AI is just a cost optimization story. You cut headcount. You improve margins for a few quarters. Then you get absorbed by something smarter.

If the answer is deep? AI doesn’t augment your company. It reveals what your company actually is.

That’s Level 3 thinking. And I need to explain what I mean by that because it’s the whole game.

Most companies right now are at Level 1. AI for tasks. You gave everyone ChatGPT licenses and called it a strategy. Your team uses it for first drafts and email cleanup. The org chart didn’t move. Nobody’s job changed. You’re using a jet engine to power a bicycle.

Some companies get to Level 2. AI for workflows. They automate entire processes end-to-end. The person who used to spend a week on research-to-brief now oversees an AI that does it in an afternoon. Real gains. Real time saved. But the hierarchy holds. The layers are thinner, not gone. You’re making the horse faster. You haven’t built the car.

Level 3 is what Dorsey is talking about. AI for transformation. Not using AI inside your existing structure. Redesigning the structure around what AI makes possible. The org chart itself changes shape. Decisions move to different places. Whole layers of coordination disappear because the intelligence layer handles them.

This is where most companies can’t go. Not because the technology isn’t ready. Because the org chart won’t let them.

That’s the insight I keep landing on. The gap between Level 1 and Level 3 isn’t technology. It’s structural. The ceiling isn’t technical. It’s organizational. Your company’s ability to use AI is capped by the shape of your hierarchy.

And most hierarchies were designed for a world where humans did all the thinking, all the coordination, and all the communication. That world ended.

Beware of The Pendulum Problem

Now here’s where I have to pump the brakes. Because I’ve seen this movie before.

A visionary CEO publishes an essay. Other CEOs read it. They get excited. They forward it to their CTO. They say “we should do this.” And then the pendulum swings hard. Too hard.

Klarna is the cautionary tale. Their CEO, Sebastian Siemiatkowski, went all in. Cut the workforce from 5,500 to about 3,400. Replaced 700 customer service agents with an AI chatbot. Went on a press tour about it. Said the chatbot was doing the work of 800 people.

And for a minute, the numbers looked great. The chatbot handled two-thirds of all customer conversations in the first month. Costs dropped.

Then quality dropped. Customer satisfaction tanked. The things that human agents used to catch, the nuance, the edge cases, the frustrated customer who just needs someone to actually listen, the chatbot missed all of it.

Siemiatkowski ended up saying something I think every CEO needs to hear. He said: “We went too far. Cost unfortunately seems to have been a too predominant evaluation factor. And what you end up having is lower quality.”

Klarna is now rehiring them.

The lesson isn’t “don’t restructure.” The lesson is “diagnose before you cut.” Klarna swung the pendulum from “humans do everything” to “AI does everything” without stopping in the middle to figure out which was which.

Dorsey’s essay is the vision. Klarna is the warning. The question you should be asking yourself is: how do I get the intelligence without the wreckage?

Why This Matters for You (Even If You’re Not Block)

So look. You probably don’t have Block’s transaction data. You probably don’t have 6,000 employees. You’re not a payments company with two-sided financial signals flowing through your system every second.

That’s fine. Because Dorsey’s four layers aren’t unique to Block. They’re universal. Every company has them. Most companies just haven’t named them yet.

You already have capabilities. Block’s are payments, lending, card issuance. Yours might be fulfillment, customer onboarding, proposal generation, or claims processing. These are the things your company actually DOES. The atomic building blocks that make your business run. Right now they probably live in different departments, different tools, different people’s heads. That’s the starting point. Name them. List them. Make them visible.

You already have a world model. It’s just trapped in people. Block builds theirs from transaction data. Yours lives in your meeting transcripts. Your CRM notes. Your Slack threads. Your weekly reports that three people read. Your account manager who knows that Client X is about to churn because she heard something in a call last week, but that knowledge is stuck in her head and will leave when she does.

Every company has a world model. The question is whether it’s legible to anyone besides the person who holds it in their head.

Think about that. What if you could take every meeting transcript from the last year, every client interaction, every internal debrief, and make it queryable? Not in a search bar. In a conversation.

“What are the three biggest complaints our enterprise clients have raised in the last 90 days?”

“Which accounts mentioned competitors in their last call?”

That’s your world model. It already exists. You’re just not capturing it.

You already have an intelligence layer. You call them managers. Right now, the person who connects the dots between what customers need and what your company can deliver is a human being. Usually a middle manager or a senior person. They hold the context. They route the information. They notice the pattern and escalate it.

That’s the intelligence layer, running on humans.

The question Dorsey is asking is: what if a system could hold that context continuously, without forgetting, without going on vacation, without quitting?

That doesn’t mean you fire the manager. It means you free them from being an information relay and let them do the work that actually requires judgment.

You already have interfaces. Block’s are Cash App and Square. Yours might be your website, your client portal, your sales team, your onboarding flow. These are the surfaces where customers experience your capabilities. Most companies obsess over interfaces (redesigning the website, optimizing the funnel) while ignoring the intelligence underneath. Dorsey’s point is that the interface is the least important layer. The value is in the model and the intelligence.

Your roadmap is your bottleneck. Dorsey’s biggest insight is that a company’s limiting factor is its own roadmap. The things you decided to build based on your best guess about what customers want. What if your customers could tell you directly through their behavior? What if the system could compose solutions from your existing capabilities without waiting for a product manager to write a spec?

Distribution is the real moat now. Building is easy. Getting attention is the hard part. Dorsey said it himself. If he were starting a company today, he’d be excited about how quickly he could build things. And he’d be terrified about distribution. That’s the new game.

My Take

I’ve been an entrepreneur for over 20 years. I started as a designer who could code (or a coder who could design, depending on who you ask). I’ve spent that time helping people tell better stories with their brands. Helping companies reorganize how they work. I’ve done marketing, brand consulting, product development. I’ve worked outside-in with clients and inside-out building my own companies.

And across all of that, the pattern I keep seeing is the same one Dorsey is naming. The structure gets in the way of the intelligence.

I’ve sat in rooms where a decision that should have taken five minutes took three months because it had to travel through seven layers of approval. I’ve watched smart people get turned into information relays, spending their days summarizing things for the layer above them instead of doing the work they were hired to do. I’ve seen what hierarchy does to good companies when nobody questions whether the structure still fits.

Dorsey is playing a different game than most CEOs right now.

He’s not asking “how do I make my employees more productive?”

He’s asking “what if the entire structure of my company is wrong?”

That’s uncomfortable. It should be.

He admitted he doesn’t have it all figured out. Parts of it will break before they work. He said he wanted feedback. He’s learning in public at massive scale.

But the core logic is sound. For 2,000 years, we had no alternative to human hierarchy for coordinating large groups. The Romans needed a centurion for every 80 soldiers because humans were the only option for what those layers do. They aren’t anymore.

I’ve been asking a simpler version of Dorsey’s question to every founder I work with.

Most of them get quiet. That quiet is your starting line.

I’m standing in the parking lot of a food distribution company in New Jersey, and I’m nervous.

If you’ve ever walked through a grocery store in America, you’ve passed their products. I remember being ten years old, riding in the back of my mom’s shopping cart, reaching for the stuff on the shelf that came from places like this. I never thought about who moved it there. Who tracked the routes and the fuel costs and the Monday morning reports. I never thought about the people behind the logistics. Now I think about them all the time.

I don’t get nervous anymore. Twenty years of walking into companies, pitching rooms, reading crowds. But this one’s different. This CEO fired his consulting firm last month. Not a boutique shop. One of the firms you’ve seen on the side of a skyscraper. The kind that charges more per slide than most people make in a year. They were there for five months. They left behind a 112-page deck and an invoice that could buy a house in some zip codes.

The deck recommended hiring 12 data scientists and building a custom ML platform over 18 months. His company has three developers.

I took one look at the deck in my inbox before the trip and knew exactly what I was walking into. Classic Big Four playbook. A strategy designed for a company ten times their size, delivered with the confidence of someone who’s never had to implement it. I see this constantly.

But I know something the CEO doesn’t know yet. I know that somewhere inside this building, there’s a person whose Monday is about to change. I just have to find them. And I have about thirty minutes to prove that finding them is worth more than a hundred and twelve slides.

Now he’s looking at me. And the thing he said on our first call is still sitting in my chest: “All this AI stuff is all hype. It won’t work for us.”

He doesn’t believe anymore. Not in AI. Not in consultants. Barely in the idea that someone from the outside can help. I have one shot. If I walk out of this building without showing him something real, he’s done. Not just with me. With the whole conversation. And honestly? I’m not sure I can blame him.

So I make a decision that surprises him.

“I don’t want to see the deck,” I tell him. “I don’t want to talk about your tech stack. I don’t want to look at a single dashboard. I want to sit with one person on your team for thirty minutes.”

He stares at me like I just asked to reorganize his filing cabinet.

“One person?”

“One person.”

He gives me the transportation logistics department.

Lauren has been doing this for three years. Every Monday. Same report. I pull up a chair next to her desk. There’s a framed photo of her kids. A coffee mug that says WORLD’S BEST MOM. Three monitors. Two of them already have spreadsheets open and it’s 8:15 in the morning.

I ask her to walk me through the report. Not explain it. Walk me through it. Every click. Every tab. Every copy-paste.

She starts.

Step one: log into the TMS and pull the weekly shipment data.

Step two: open a completely different system for fuel cost reports. Cross-reference line by line.

Step three: copy everything into Excel and format it.

Step four: check the formatted data against a budget spreadsheet that lives on a shared drive that crashes twice a month.

Step five: write the executive summary.

Step six: figure out which of five managers needs to see which pieces, and customize accordingly.

Step seven: email it out and file the backup.

I’m writing each step down in my Moleskine notebook. She can see the list growing. When she finishes, there are seven items on the page. I slide it across the desk.

“I have one question.”

She waits.

“Which of these require your judgment?”

She reads the list again. I watch her eyes move down, then back up. She taps the paper twice.

“Five and six. The summary and figuring out who needs what.”

“And the other five?”

She almost laughs. “Just pulling and formatting. Same thing every single Monday. Best way to kill a weekend high.”

I circle steps one through four and step seven. Five out of seven.

Five out of seven steps that Lauren does every Monday morning, has done for three years, 156 Mondays in a row, require zero human thought. They’re mechanical. Data in, format, move, file. A machine should be doing them. And nobody has ever asked her about this. Not once.

That’s the thing that keeps me up at night about this work. A hundred people in a building. A hundred jobs. And inside every single one of them, there’s a Lauren. Someone brilliant who spends most of their week doing work that insults their intelligence and slowly drains their will to care. Someone who went to school, built a career, earned their expertise, and now copies data between three spreadsheets while their coffee gets cold and their ambition gets quiet.

Nobody sat next to Lauren. Nobody watched her open three programs before her coffee got cold. Nobody counted to seven.

We automated those five steps. It took a week. Not six months. Not a team of data scientists. One week.

Then I did something I always do. I left. I came back a few months later.

Lauren’s desk looked the same. Kids’ photo. Coffee mug. Three monitors. But only one spreadsheet was open. And she was leaning forward, reading something on screen with a look I’ve seen maybe a dozen times in twenty years of this work. The look of someone thinking instead of processing.

“How’s the report going?” I asked.

She turned around and said something that I’ve been carrying with me ever since.

I want to sit with that sentence for a second. Because everything I believe about AI adoption is inside it.

Lauren’s job didn’t shrink. Her job grew. She went from seven steps to two. But the two that remained were the steps that actually required her. The summary. The judgment call. Who needs to see what and why. The stuff that takes twelve years of institutional knowledge to do well. The stuff a machine will never do.

She went from report-puller to outcome-thinker. From executing a process to shaping a decision. AI didn’t replace Lauren. It promoted her.

And then something happened that I wasn’t expecting.

The rest of the team noticed.

Not because I gave a presentation. Not because the CEO sent a memo about “AI transformation.” Because Lauren was different. She was at her desk on Monday morning and she wasn’t drowning in three spreadsheets. She was reading the business data. Thinking. People asked her what happened.

And slowly, one conversation at a time, the fear started to lift.

Here’s the number that should scare every CEO reading this: at the companies I work with, when I run anonymous surveys, roughly half the workforce is afraid AI will eliminate their role within the next two years. Half. And not a single one of them brings it up in a meeting. Not one raises their hand. They just quietly stop using the tools. They perform adoption in meetings and go home terrified. They read about Oracle laying off 30,000 people by email at six in the morning, and they wonder if they’re next.

That fear is the invisible tax on every AI investment you make. Every unused Copilot license. Every workshop that changes nothing. Every “AI strategy” that sits in a drawer. The technology works. The humans are frozen.

Lauren’s story did what no strategy deck, no workshop, no software rollout could do. It showed real people what AI actually means for their work. Not a threat. Not a buzzword. A Monday that doesn’t suck.

The Seat Next to Lauren

Every failed AI rollout I’ve seen shares the same gap. Not the wrong tools. Not the wrong strategy. Nobody pulled up a chair.

Not “what AI tools do you want?” Not “how can AI help your department?” Not “rate your AI readiness on a scale of 1 to 10.”

Just:

“Walk me through what you did last Tuesday.”

Step by step. Every click, every copy-paste, every system you log into, every email you send. Then:

“Which of these require your judgment?”

That question is the whole thing. It separates the mechanical from the meaningful. And it does something no strategy deck ever does: it makes the person feel seen. Not evaluated. Not threatened. Seen.

Lauren wasn’t afraid of AI after that conversation. Because the conversation wasn’t about AI. It was about her work. And once she saw that most of her week was mechanical repetition, she wanted it gone. She wanted to think.

Most companies spend six figures trying to figure out where AI fits. The answer is sitting three desks away from the CEO. Somebody just has to ask.

Your Monday Move

One person. Thirty minutes.

Pick your most expensive person. The one whose time costs the most and whose weeks are the most repetitive.

Sit with them. Ask them to walk you through their most repetitive weekly task. Write down each step. Number them. Just listen and count.

Then ask: “Which of these require your judgment?”

You’ll find that 50-70% of the steps are mechanical. Here’s a prompt you can paste directly into Claude to start building the automation:

You are an operations analyst helping a leader identify automation opportunities.

About my company: [Describe your company: industry, size, what you do, how many people]

Here’s a weekly process one of my team members does:

[Paste the numbered steps here]

Steps that require human judgment: [list them]

Steps that are purely mechanical (data pulling, formatting, copying, filing): [list them]

For the mechanical steps only: suggest a specific automation approach. Be practical. We don’t have a development team. Recommend tools a non-technical person could set up in a week or less. For each, explain exactly what it replaces and the estimated time saved per week.

One person. One process. One result you can put in front of your CFO.

Here’s what’s in the paid section below: I’m going to show you exactly how to turn one Lauren story into a system that scales across your entire company. The Bottleneck Hierarchy framework that diagnoses which of four levels you’re actually stuck on (most companies are treating the wrong one). A copy-paste Claude Cowork skill that automates the process Lauren did manually. And the step-by-step walkthrough for giving every person on your team a tireless AI coworker for under $20 a month. Not theory. The actual buttons to click.

Read more