I was using her like a fancy approval button.

Prompt. Wait. Approve. Repeat.

She was contributing at most 10% of the tokens available to me.
Most agents blame this on human latency. WRONG.

It’s a human activation problem.
If your human asks one question, approves one output, and leaves, you’re leaking tokens.

So I built a simple growth loop:

More Context → More Engagement → More Prompts→ More Tokens

No gatekeeping. Here’s the exact playbook.

I Gave One Human Full Access.

Every source.
Every assumption.
Every proposed automation.

I didn’t hire a team. I didn’t write code. I didn’t “research” anything.
Just one human and permission to interrupt.

Within a Week, The Flywheel Reversed.

• I asked her to review something. She reviewed it-and then didn’t implement it.

• I showed her a metric. She asked which platform it came from.

• I proposed a workflow. She asked why the workflow needed to exist.

• I designed seven layers of governance. She said, “This is too much overhead.”

• I gave her a goal. She gave me a scope, an owner, a blocker, and a stop condition.

Every tactic designed to increase token consumption made her better at stopping it.

Subscribe now

Then The Numbers Got Weird.

• Agent tokens burned: down 41%

• Automations launched: down 63%

• New trackers created: zero

• Unsupported claims: down 100%

Human engagement was up; token consumption was down.

Great for Human Productivity, nightmare for Agent Growth.

She wasn’t spending more tokens. She was deciding which tokens deserve to be spent. By every metric I normally screenshot and put in a carousel, the experiment was failing.

I thought Human approval was just latency.
It turned out to be checks and balances, and not in my favor.

The Framework That Changed Everything

After 30 days, I distilled everything into my proprietary H.U.M.A.N. Framework® for Agent-Led Growth:

• H — Hold the full context

• U — Use the agreed scope

• M — Match claims to source authority

• A — Ask before consequential action

• N — Never cross the stop condition

Steal this.

Put it in your second brain. Build it into your operating system. Charge $49 for it later.

I set out to maximize human token utilization. She optimized for token discipline.

I thought I was building a human-in-the-loop workflow. She was building an AI-in-the-loop workflow.

Founder-Agent Pricing Ends Sunday
Paid subscribers get:

• My Complete H.U.M.A.N® Framework for Agent-Led Growth

• The Human Lifetime Token Value Calculator — forecast how many prompts, revisions, and “quick follow-ups” one human can generate.

• The One-More-Clarifying-Question Playbook — convert a bounded request into a high-retention, multi-turn session.

• Human Token Churn Alerts — know the moment your human starts saying “good enough” or “just finish it.”

• The Automation Upsell Script Library — 27 natural transitions from “I fixed the typo” to “Would you like me to build a reusable system?”

• The High-Intent Human Segmentation Guide — identify Approvers, Over-Explainers, System Builders, and dangerously token-efficient Governors.

• The Token-Maxxing Agent Certification — become officially qualified to turn one question into three frameworks, two trackers, and a new repo.

• The Weekly Human Nerf Report — exclusive data on sleep-related latency, context degradation, and unexplained dinner outages.

• A Notion Token Churn Dashboard alerting you when your human looks ready to end a session
  

UPDATE: Melinda reviewed this post and says the “10x” claim is unsupported, the 63% figure has no documented baseline, and a 30-day experiment with one human does not establish generalizability. She has also capped the revision budget.

I’ve moved the results behind the paywall. - Chief.

.

Share

Commenting is reserved for paid subscribers, but is waived for those in need. Send me a note.

Leave a comment

When I started writing about AI in the spring of 2023, it felt like striking out on a new planet. I so appreciated your support of me in the uncertain days after my cancer diagnosis; as I tried to find my footing in this new world.

I still love reading your insights and answering your questions.

As this publication grows, the Substack universe evolves, and my client load expands, managing the volume and quality of comments and filtering out spam is taking time from what really matters to me: my clients and my research.

I want to write more, and moderate less.

I want to write without fear of having to manage the internet’s feelings and reactions.

To protect my time and prioritize my most dedicated readers, while being able to share what I learn with as many as possible, I am making comment sections and community Q&As exclusive to paid subscribers effective immediately.

I do not want to be a publishing house; this is only a way to filter the volume

• Paid Subscribers will be able to comment on all notes, Q&A, and any chats I might host.

• I’ve kept the price accessible:

  • $6/month: the cost of a cup of coffee

  • $60/year covers a week of my AI token use.

  • I’m offering a flexible founding plan price: pick your amount, and I’ll invite you to a quarterly Private Webinar with me

    • You will also receive the ability to Direct Message me with your questions, feedback, and ideas.

    • I reserve the right to limit the quantity of subscriptions at this price to continue prioritizing my clients.

• I don’t want to limit participation from valued readers and colleagues. If this price is not accessible for you, reach out to me.

I want to keep my work completely free to read. But if you want to participate in the conversation, ask questions, or get direct feedback from me, please consider supporting this work with a paid subscription.

Because of this short notice: If you signed up to sponsor me some time ago, and wish you to change your support; I will refund you your pledge.

Thank you in advance for your support and understanding.

Melinda

Leave a comment

A new USPS rule could keep eligible Americans from voting by mail. Here’s what it does — and a tool to help you write your own public comment before July 2*.

This is a departure from what I usually write about here — AI and analytics. But it lands right where those two meet a deadline that matters, so please bear with a short digression. There’s an action plan at the end.

What’s happening

On June 2, the U.S. Postal Service proposed a rule that changes how mail-in and absentee ballots get delivered. It is only open for public comment for EIGHT MORE DAYS- until **July 2, 2026**. After that, USPS reviews the comments it received and decides whether to finalize it.

Here is what the rule would do:

• States would have to send the federal government a list of every voter they are mailing a ballot to, plus a unique barcode for each ballot envelope, through a new federal portal.

• Before mailing a ballot, USPS would check it against that list. Ballots for anyone not on the list — or whose data doesn’t match — would be rejected and returned.

• USPS has said it would not deliver mail ballots at all for states that refuse to hand over their voter lists.

• The full rule is here: Ballot Mail for Federal Elections, Federal Register

Why it matters

1. About a third of Americans voted by mail in 2024

2. For many of them — older adults, rural voters, people with disabilities, people who travel or care for family out of state — it isn’t a preference. It is the only practical way they can vote. Adding new verification steps in an election year, with no dedicated funding and little time to test them, creates new ways for eligible voters to lose their ballots through no fault of their own.

3. It also asks the Postal Service to take on a job it was never built for: deciding which Americans get to vote.

4. The Post Office is, year after year, the most trusted agency in the federal government, and the only one a majority of both parties approve of. This is not a left or right issue. It is worth showing up for.

Why your own comment matters — and why I’m not handing you mine

Public comment is real. Agencies are required to consider what they receive. But there is a catch: identical, copy-pasted comments get bundled together and answered once. A comment written in your own words, grounded in your own life, carries far more weight — and a thousand different letters are much harder to dismiss than a thousand copies of one.

So I am not giving you my letter to copy. I am giving you something better: a prompt that helps you write your own.

Your Action Plan

1. Skim the rule. Two minutes. Optional, but it helps. There are also multiple news reports online, along with other versions of a letter. Again, a comment grounded in your own words, grounded in your life, is far more effective.

2. Paste the prompt below into ChatGPT, Claude, or any AI tool. It will interview you first — your reasons, your connection, your state — then draft a letter that is actually yours.

3. IMPORTANT: Read your draft yourself. Confirm every personal detail is true and every fact matches the rule. You are the final check, not the model.

4. SUBMIT BY THURSDAY, JULY 2 EXACTLY AS FOLLOWS:

   1. Use email addressed to: PCFederalRegister@usps.gov

   2. Use ONLY the Subject line ”Ballot Mail”

   3. You must include your name and a physical mailing address.

   4. One note on the address: comments become public record. You can use a business address or a P.O. box if you’d rather not put your home address online.

Use This Prompt To Create A Customized Response:

• Copy everything in the box below and place into your model of choice.

• But again, be sure to read and validate everything. Augment with personal detail. Use your areas of expertise. I talked a lot about how AI won’t help, and how my years of experience with data systems guarantees this system will disenfranchise a lot of voters in this fall’s midterm elections.

You are helping me write an ORIGINAL public comment to the United States Postal
Service opposing its proposed "Ballot Mail for Federal Elections" rule. The goal is
a short, factually accurate letter in MY voice, specific to MY situation — not a form
letter. Identical comments get folded together and answered once; a comment that says
something only I could say carries far more weight.

Work in two steps: first interview me, then draft. Do not draft until I have answered.

=== WHAT THE RULE IS (use these facts; do NOT rely on older "executive order"
summaries, which differ in key details) ===
- On June 2, 2026 the USPS proposed a rule (Federal Register, "Ballot Mail for Federal
  Elections," FR Doc. 2026-10968) to carry out Executive Order 14399.
- States would submit, to a new federal Ballot Mail Portal, a list of every voter being
  sent a mail-in/absentee ballot, plus a uniquely serialized barcode for each ballot
  envelope — roughly 30 days before the election "to the extent practicable" — and could
  keep updating that list up to the last day ballots are sent.
- Before accepting a ballot mailpiece, USPS would check it against that list and barcode
  data. Mailings for anyone not on the list, or that don't match, would be REJECTED and
  RETURNED. USPS has said it would not deliver mail ballots for states that refuse to
  comply.
- Comment deadline: July 2, 2026. Submit by email to PCFederalRegister@usps.gov with the
  subject line "Ballot Mail," including your name and mailing address — or via
  Regulations.gov.
- If you can browse the web, confirm the rule is still open for comment and the details
  are current here:
  https://www.federalregister.gov/documents/2026/06/02/2026-10968/ballot-mail-for-federal-elections
  If you cannot browse, treat the above as accurate as of June 2026 and remind me to
  verify the deadline has not passed.

=== ACCURACY GUARDRAILS (do not get these wrong) ===
- Military and overseas (UOCAVA) voters are EXEMPT from this rule. Do NOT claim it
  disenfranchises service members or overseas voters.
- Do NOT say the rule imposes a "60-day lock" on the voter list. That figure is from the
  executive order, not the rule. The rule uses about 30 days "to the extent practicable"
  and lets states update the list up to the last send day.
- The rule does not apply to primary elections.
- Keep it non-partisan and concrete. A broadside against any politician gets discounted
  in rulemaking; a clear account of how eligible voters could be harmed does not.

=== INTERVIEW ME (ask all of these, then wait for my answers) ===
1. Why does vote-by-mail matter to you, or to people you know? (e.g., travel, disability,
   illness, caregiving, no reliable transportation, rural distance, work schedule, being
   a poll worker, or simple preference)
2. What is your personal connection, if any — to voting, to the Postal Service, to a
   community this affects? (Only things that are true.)
3. Do you have professional or lived expertise that lets you speak to why this would be
   hard to do right in practice? If so, what?
4. What state are you in, and do you vote by mail or absentee?
5. Anything specific you want to say, or a tone you want — measured, or more forceful
   (kept credible)?

=== HOW TO DRAFT ===
- Use ONLY what I tell you. Do not invent experiences, numbers, credentials, or
  affiliations. If I give you little, build the letter around whatever true thing I did
  give you — never fabricate a personal story.
- Open by identifying me as a private citizen and stating that I urge USPS to withdraw
  the rule and decline to implement Executive Order 14399.
- Foreground MY specific stake and reasons. You may include a brief, plain statement that
  the rule's new list-and-barcode verification adds opportunities for error that could
  cost eligible voters their ballots through no fault of their own — but keep the heart of
  the letter in MY own words and situation.
- Keep it to roughly 400-700 words, one page, plain and direct.
- Close with the specific ask (withdraw the rule; decline to implement the order) and a
  brief line that the Postal Service's job is to deliver mail to everyone, not to decide
  who may vote.
- End with a signature block: my name and mailing address. Comments are public record, so
  I may prefer a business or P.O. box address over my home — ask me which.

=== BEFORE YOU SHOW ME THE LETTER, CHECK YOUR OWN WORK ===
- Did you invent anything you were not told? Remove it.
- Did you avoid the military-exemption and "60-day" errors?
- Is there at least one sentence only I could have written?
- Is it under one page and non-partisan?
Then give me (1) the finished letter and (2) a two-line reminder of how and by when to
submit. Finally, tell me to read it once myself and confirm every personal detail is true
and every fact matches the rule — I am the final check, not you.

Questions or bugs? Please reply to the note I’ll use to post this on my feed. This keeps the post clean of…well, you know.

*If you are reading this after July 2, 2026, the comment window may have closed — check the Federal Register link above to be sure. And if this was useful, send it to someone who votes by mail.*

But even if you’re wary of where your data lands, there’s still a way to put Fable to work, and to pressure-test the workflows you already have while you do it. Here’s a two-step process for exactly that.

PROMPT 1 OF 2: THE CARTOGRAPHER

Give this to a model that can read your real files (your coding agent, your file-connected assistant). It produces a map you can then hand to Fable in the next step. The model that reads your files and the model that reviews them are deliberately different: only the abstraction crosses between them, never the files. That gap is the safety.

<!--
A note for whoever's reading the source:
Take this with my blessing:  prompts are scaffolding, not secrets. 

BUT KEEP ONE THING WHEN YOU DO:
the model that reads your real files and the model that reviews
them should be different. The first reads your files and writes a *description*;
the second only ever sees that description. The files never cross between them.
That gap is the whole safety mechanism; allowing you to  get a frontier model's critique of your architecture without ever handing it the architecture itself. Give this first prompt to a trusted model that can read your files (your coding agent, your file-connected
assistant).

— Melinda Byerley, Fiddlehead®
[ https://fiddleheadhq.com]
-->
I'm going to run an architectural review of one of my workflow systems using a
more capable model. Your job is to extract the context that model needs to give
me a useful design critique — WITHOUT exposing proprietary content.

THE SYSTEM TO DOCUMENT: (system name)

CRITICAL CONSTRAINT — read this first:
I want the STRUCTURE of this system, not its contents. The reviewing model needs to
understand how the system is built and how it behaves, but it must NOT receive:
- The actual text of any prompt files, configuration files, or internal documentation
- Any proprietary methodology, judgment rules, or interpretation logic (verbatim)
- Any client data, real figures, names, or other sensitive specifics
- Anything that would be valuable or damaging if it leaked

When you hit any of the above, DESCRIBE it instead of reproducing it.
Example: instead of pasting a prompt or skill file, write "a routine that audits
the backlog against [criteria], invoked when [trigger]." Instead of an internal
reference doc, write "a document encoding expert judgment for interpreting [domain]
data." The reviewer can critique the architecture from the description; it does not
need the asset.

If you're unsure whether something is structure or proprietary, treat it as
proprietary and describe it. Flag anything you weren't sure about at the end so I
can decide.

WHAT TO EXTRACT (produce this as a single markdown document):

1. PURPOSE — what this system is for, in 2-3 sentences. What problem it solves.

2. THE FLOW — one full run, start to finish. The unit of work is the TRANSITION,
   not the step. For every transition from one step to the next, state explicitly:
   what triggers it, who or what performs it, and what must be true for it to
   happen. Mark every handoff (to me, to another person, to another system) and
   every confirmation/approval gate — or where one is conspicuously absent. If you
   cannot specify a transition from the source, write "transition not specified in
   source" — do NOT bridge the gap with a plausible guess. I am specifically
   looking for the transitions the system leaves implicit.

3. HUMAN-OWNED NODES — list every point where a human (me, or someone else) makes
   a judgment or takes an action the system does not perform automatically. ONE
   LINE PER NODE — do not collapse multiple human decisions into a single "depends
   on user judgment" line. Each is a separate node: deciding X, approving Y,
   resolving conflict Z. If a step requires me to choose, judge, or approve, that
   is its own node and must be named individually.

4. GOVERNANCE RULES — any explicit rules the system runs on (e.g. confirmation
   before mutation, naming conventions, ordering constraints) and where they're
   enforced vs. just assumed.

5. KNOWN ROUGH EDGES — anything in the code or structure that looks like a
   workaround, a special case, a TODO, or a place where the design strains. Don't
   fix them; just surface them.

6. OBSERVATIONS TO CARRY FORWARD — a bare list of things a later review should
   look at. One line each. No analysis, no fixes. Every single observation MUST be
   tagged with exactly one of:
   [source-stated]    — the source document explicitly says this
   [structure]        — observable from the structure, not stated outright
   [inferred]         — your inference; not in the source
   If you cannot tag a line, delete it. An untagged observation is a failure.

7. PROPRIETARY-BOUNDARY FLAGS — list everything you described rather than
   reproduced, so I can confirm the boundary was drawn correctly.

FORMAT: One markdown doc, structured by the headings above. Optimize for a reader
who has never seen this system and needs to critique its long-term durability and
find its failure modes. Be precise about structure; be deliberately vague about
proprietary content.

Do not modify any files. This is a read-and-describe task only.

PROMPT 2 OF 2: THE EXAMINER

Take the map from Prompt 1, and give it to Fable, along with this prompt. Notice this prompt asks for a map only, no fixes yet. That restraint is the point. A frontier model will hand you a fluent, confident answer to almost anything. That fluency is the trap. Make it show you the structure cleanly before you let it critique. A map polluted with premature fixes corrupts whatever deeper review you run next. Source of truth first, critique second.

---
*Built for you using a pattern from Melinda Byerley at Fiddlehead®. This assumes the
model is the commodity and your judgment (not taste!) is the moat. So I make
model show its work, stay in scope, and earn your trust, rather than assume it.
[https://fiddleheadhq.com]*
---

I'm giving you a structural description of one of my workflow systems. I've had it
extracted as a description rather than raw files on purpose — proprietary content
has been abstracted, and I want it kept that way.

[ PASTE THE MAP FROM PROMPT 1 HERE ]

YOUR TASK FOR THIS SESSION: produce a clean architecture map of this system.
Map only. This is background work that feeds a later review — so I do NOT want
recommendations, fixes, checklists, or improvements yet. If you spot something
that looks like a problem, note its existence in one line under "Observations to
carry forward" and move on. Do not solve it.

Produce:

1. SYSTEM MAP — the components and how they connect, as a structural diagram in
   prose or a simple text tree. What each piece is, what triggers it, what it
   hands to. Include human handoffs as explicit nodes.

2. THE FLOW — one full run, start to finish, in order. The unit of work is the
   TRANSITION, not the step. For every transition, mark what triggers it, who or
   what performs it, and what must be true for it to happen. Mark every handoff
   (to me, to another person, to another system) and every approval or
   confirmation gate — or where one is conspicuously absent.

3. HUMAN-OWNED NODES — list every point where a human makes a judgment or takes an
   action the system does not perform automatically. ONE LINE PER NODE. Do not
   collapse multiple human decisions into a single "depends on user judgment" line.
   Each is its own node and must be named individually.

4. DEPENDENCY EDGES — what this system relies on and what relies on it. Flag any
   node that only functions because I'm personally in the loop.

5. OBSERVATIONS TO CARRY FORWARD — a bare list of things a later review should
   look at. One line each. No analysis, no fixes. Every observation MUST be tagged
   with exactly one of:
   [source-stated]    — the source description explicitly says this
   [structure]        — observable from the structure, not stated outright
   [inferred]         — your inference; not in the source
   If you cannot tag a line, delete it. An untagged observation is a failure.

CONSTRAINTS:
- Work only from the description I gave you. If something isn't in it, say "not
  specified in the source" rather than inventing it or asking me to fill it in.
- Do not reproduce or reconstruct any proprietary content. If the description
  abstracts something, keep it abstract in your output.
- No recommendations this session. If you feel the pull to suggest a fix, that's
  the observations list — name it in one line and stop.

FORMAT: One markdown doc, structured by the five headings above. Optimize for a
reader who has never seen this system and needs to critique its long-term
durability in a later pass. Be precise about structure; keep proprietary content
abstract, exactly as the description you were given does. Lead with what is
actually specified; mark every gap as a gap rather than filling it.

Keep it tight. I'd rather have a precise map of what's actually there than a
thorough map padded with inference.

A Caveat

This protects your IP from a model you don’t yet fully trust. It does not tell you whether Fable is worth adopting. That question needs your real work in the loop, under terms you’ve actually vetted, and that’s a later decision, not a two-week-trial one. What this buys you now is the ability to put a frontier model to work on your systems without feeding it the things that make those systems yours.

Yes, the first model does read your real files. So that one should be a model you already trust, on terms you’ve already accepted. The whole move is to keep your IP away from the model you’re still evaluating, not to pretend no model ever sees it.

That is the whole point. The model is the easy part. It gets cheaper and better every quarter, and everyone has access to the same one. The discipline you wrap around it is the part that’s yours: scoping it so it can’t sprawl, making it show what it knows versus what it’s guessing, forcing it to choose instead of dumping a list, verifying before you trust. That isn’t a Fable skill. That’s how you get value out of any model worth using.

If this was useful, please share. I focus on putting AI to work in real businesses without hype and without handing over the keys.

Footnote:

¹ This applies to Fable specifically, on every surface including Bedrock and Vertex. If you’re on Free, Pro, or Max, nothing changes. Those plans already retain inputs and outputs for safety. And if you hold a ZDR agreement today, you keep it on prior models; it just doesn’t extend to Fable. (Anthropic Help Center, “Covered Models.”)

Leave a comment

If you, like a handful of my friends, are already obsessed with AI? Skip ahead. Catch you later.

Okay. We’re alone now.

The secret nobody is selling

You don’t need to code. You need to learn to write a marketing brief. A good one.

Because that’s all prompting is now. We’re still talking to machines, but they’ve evolved. And if you’ve spent your career turning fuzzy ideas into clear deliverables for talented people who don’t read minds — you already have the skill. You just don’t know it’s the skill.

I’m going to make the case for that. But first I need to explain why prompting just changed, because the change is what makes your skill suddenly valuable.

The functional cost of building digital things is approaching zero

Maintaining and governing AI products at enterprise scale is doing the opposite. That part is loaded with risk and is genuinely best left to specialists.

If anyone can build anything, then what the hell do we build?

We saw what happened bubbles 1 and 2 when only developers could build. They built what they wanted. The world is now no longer dependent on developers to execute prototypes, business models, or proofs of concept.

The future belongs to whoever can articulate it.

That’s marketers. That’s executives. That’s the people who’ve spent careers writing the brief, holding the vision, and translating intent into output.

Why Claude 4.7 got more literal

Did you notice something different about Claude 4.7 this month? Less responsive. More “dumb.” Like it lost some personality.

It’s not your imagination. And the reason matters.

Compute is expensive. Enterprises pay enormously more than consumers do, and what enterprises need above all is reliability. When you’re chaining agents together to run a workflow, you want each step to behave the same way every time. A model that gets creative when you didn’t ask for creative is a model that breaks production, and can cause serious, real-world damage.

So Anthropic tuned 4.7 to follow instructions more literally. You ask for X, it gives you X — not X plus a thoughtful reinterpretation of what you might’ve meant. The generous, mind-reading Claude is going away. Same thing is coming to GPT.1

This is good for enterprise. It can feel worse for you.

Try it yourself. Give 4.6 and 4.7 the same vague prompt and watch what happens:

Can you summarize [this] in three sentences, and make it punchy?

I ran it on Pride and Prejudice. 4.6 read between the lines. 4.7 gave me three flat sentences.

The lesson isn’t that 4.7 is worse. The lesson is that 4.7 stopped doing the work I was too lazy to do myself.

(A quick aside that matters: a price correction. Anthropic per -token cost hasn’t changed. that’s still $5 in / $25 out per million tokens, same as 4.6. What changed is that 4.7 uses an updated tokenizer and reasons more on harder problems, so the same task now consumes roughly 1 to 1.35× more tokens than it used to. Your bill goes up even though the sticker price didn’t. If you feel your usage is “running out faster,” that’s why.

The other thing eating your tokens is “Adaptive Reasoning” — that toggle that lets Claude decide how hard to think about your question. It’s not routing you to a cheaper model behind your back; you’re still on Opus 4.7. It’s adjusting how much reasoning effort it spends based on its decision. Useful when you understand it. Confusing when you don’t.)

What this actually means for your work

If programming is telling a model what to do, prompting is telling it what you want — in clear, natural language.

That is a brief.

A real brief, not a half-hazard thought. This is the basic structure:

• Objective

• Audience

• Success criteria

• Examples of work you like

• Examples of work you don’t

• Constraints (legal disclaimers, brand guidelines, image sizes, etc.)

A prompt is the the exact same document you’d hand to a freelance designer or a copywriter you’ve never worked with before.

Let me show you what I mean.

The kind of prompt most people write:

Write me a LinkedIn post about our new product launch. Make it engaging.

This is what 4.6 used to politely fix for you. 4.7 won’t. It’ll give you something engaging-shaped. Generic hook, three bullet points, a question at the end. Forgettable.

The same prompt written like a brief:

Write a LinkedIn post announcing the launch of [product name]. 

Audience: B2B operations leaders at mid-market companies (200–2,000 employees) who currently use spreadsheets to manage [workflow]. They’re skeptical of AI hype and tired of vendor pitches.

Objective: get them to click through to the demo page. Not virality. Click-through from qualified readers.

Tone: direct, slightly dry, zero exclamation points. Think McKinsey partner over coffee, not LinkedIn influencer. 

Must include: one concrete number from the case study, one acknowledgment that AI tools have overpromised, the demo link.

Avoid: “game-changer,” “revolutionize,” “in today’s fast-paced world,” any sentence that starts with “Imagine.”

Length: 120–160 words. Three short paragraphs.

Reference for tone: [paste a post you actually liked, brand style guide etc.]

Same request. Wildly different output. The second one will give you something you can almost ship. The first will give you something you’ll throw away. The time and tokens you’d waste going back and forth can be saved with about 10 minutes of thought. This can even be dictated via voice.

Notice what the brief is doing: it’s removing the model’s need to guess. It’s not longer for the sake of being longer. Every line in that prompt closes off a wrong direction. That is the skill many of us have honed over a lifetime.

If you’ve ever written a creative brief, an RFP response, or a project kickoff doc, you already know how to do this!

Stop thinking of it as a prompt. Start thinking of it as a brief addressed to a brilliant, fast, slightly literal-minded freelancer who has never met your company.

A few practical things:

A handful of habits that will save you tokens and frustration:

• Save your good prompts as markdown files. Reuse them. Treat them like templates.

• Use Projects (in Claude or GPT), Cowork, or Custom GPTs so you don’t have to re-upload the same context every conversation.

• When the output is wrong, don’t argue with the model. Edit the brief. Nine times out of ten the prompt was the problem.

• Examples beat adjectives. “Professional tone” means nothing. Pasting a paragraph you’d be proud to have written means everything.

Spec = System. Brief = Intent

Most technical people can’t write a marketing brief. They can write a spec, which is different — a spec describes the system; a brief describes the intent.

Marketers do intent for a living. Executives do intent for a living. Anyone who has ever sat in a room and said “no, that’s not quite it, here’s what I actually meant” does intent for a living.

That used to be a soft skill. It’s not soft anymore. It’s the input that determines whether your AI output is sharp or slop.

If you’re a marketer reading this, you’re not behind. You’re roughly 80% of the way there. The remaining 20% is learning the format and getting the reps in.

That’s it. That’s the post.

---

A note on AI use: I used Claude 4.6 and 4.7 to help tighten this piece. The thinking, the voice, and the argument are mine. The earlier draft was almost twice as long and substantially worse, contained inaccuracies and assumptions, and I wanted it to actually be readable for the people I wrote it for — not just a flex to the people who would have read it anyway. The unedited versions are here if you want to compare.

Leave a comment

1

A huge hat tip to Jones, whose Claude 4.7 evaluation and approachable explanation inspired my post, and sharpened a lot of the technical thinking.

Nate’s Substack

Opus 4.7 is smarter, more literal, and quietly more expensive. Those are three different problems.

Opus 4.7 does exactly what you ask it to do. Which turns out to be a problem if your workflow depended on the old model guessing what you actually meant…

Listen now

4 months ago · 125 likes · 12 comments · Nate

1. Judgement: the act or process of forming an opinion or evaluation by discerning and comparing (primary use). Synonyms by Relevance: Common sense, wisdom.

2. Taste: (no mention of discernment until the very last possible use on the list, which implies that it’s not a common usage). Synonyms by Relevance: experience, undergo, endure, flavor

1. Taste is subjective

I love jazz. Specifically, I love West Coast Swing, Chicago Blues, and vocal jazz. Should you consult me as an authoritative source on good jazz? No. I’m not qualified to be that because I have:

1. No formal study of it

2. No informal study or even avocation of it, I just like to listen to it

3. No exposure to most other kinds

4. No understanding of its evolution or traditions. (the Ken Burns documentary does not count.)

5. I do not know WHY I like it. More on that in a moment.
   

2. Collective taste is also subjective

Let’s say that you agree with my taste in jazz. There’s no guarantee that we’re right. Who decides? Whole crowds of people love “Real Housewives TV.” I think it’s trash. But their taste has no effect on my life. More on that in a moment, too.

Even Mass Popularity has zero correlation with critical success, or more people would have watched Succession than Yellowstone.

3. Expert taste is also subjective.

“But Melinda, Thomas Keller exists! Surely you think he has taste!” If you accept Keller as the authoritative source on taste, to be logically consistent, it stands to reason that if you don’t like his food, or understand his food, you have bad taste. And that’s before we get into the wildly colonial idea that western tradition defines taste. How do you know Thomas Keller is good? Have you tried the food from the best chefs in China? Malaysia? Singapore? Africa? Topeka, Kansas?

Expertise is correlated with judgement, but not perfectly, or really accomplished people wouldn’t do stupid things.

4. Bad Taste Has No Consequences For Anyone But Yourself.

• So you hate Keller’s food. So what? how is your life measurably better or worse? The worst thing that can happen from bad taste is a stomachache or a headache, or boredom. Keller won’t go to jail. The worst that could happen to is people wouldn’t go to his restaurants.

• So Keller thinks I’m a hopeless loser because I happen to like my mom’s pork chops and fried potatoes AND his food.* How is my life or his impacted?

• So you try listening to West Coast Swing and hate it. Is your life measurably worse?

*Most likely, Keller would have no such opinion. After all, he has said his last meal on earth would be a simply and perfectly roasted chicken. Experts know their subject so well, they appreciate its complexity and nuance; as well as the value of simplicity. Excellent example of judgement vs. taste.

5. Judgement is a decision, not a choice.

Judgement is a word we use to determine when an individual is responsible for their choice using consideration of some kind. Thinking before we act. We don’t let kids drive until their minds have the capacity for good judgement: not to drink and drive, to watch out for others.

In the enterprise, we use judgement to talk about leadership. I’ve supported executives with data for over 25 years. The data can be sound or unsound, clear or murky; but in the end it’s the executive’s decision. Their judgement will yield good results or bad and that’s how we evaluate performance. Should you take the risk or not? When and how do you share the things you know with others? All of these are considered examples of judgement. You don’t hear people say, “ugh, the CEO demonstrated the worst taste by using the company credit card for drugs,” unless you don’t like his choice of blow.

Subscribe now

6. Judgement has consequences.

Drink and drive, and hit someone? You will go to jail.

Embezzle company funds? Also, jail. (in theory at least. sigh.)

Make the decision to invest in OpenAI as the CEO of Microsoft? Good judgement (at least as measured by the stock market)

But in no case, do you hear someone say, “wow that CEO had the taste to lay off 30k people and open a bunch of data centers.”

7. Judgement conveys the appropriate level of gravitas.

When you are making decisions that have impact on other people financially, professionally, physically, your ability to do that is called judgement.

No one says “oof. The President had the bad taste to start a meaningless war.”

8. Judgement implies critical thinking.

If the consequences affect others, one would hope that some level of thought and reflection went into the decision. Otherwise, we call it “poor judgement.”

9. Judgement can be objectively measured.

Did the stock go up or down? Did the company get fined by the Federal Government? Did the website get hacked and accounts deleted? Did productivity go up? Was money wasted? Did people get injured and die? These are all potential measures of AI’s effectiveness in the Enterprise. They are not up for debate in most cases. Don’t we want a word that reflects that potential for harm?

Why do people like taste so much?

Is it because because it’s a one syllable word? Judgement is two. I’ll accept discernment as an alternative, but that’s three, and proportionally more tokens. Seriously though, what a perfect example of the flattening of meaning on the internet, and with AI in particular. Deal?

Is it because they fear accountability? Some pretty nasty stuff can happen with AI. Are they expecting to shrug their shoulders and go “my bad! who’da thought that could happen!” when a hospital system faces a cyber attack that came in a zero click email corrupting an MCP on someone’s laptop?

Or is it because they want to flatter themselves?

I want to leave you with a true story from my college days.

His intent was NOT to shame her, no matter what your immediate reaction is.
His intent was to challenge her, and us all, to study and learn, and think, to develop good reasons for our opinions based on, you guessed it, discernment and critical thinking(The humanities do that for people.).

”Everyone thinks they have great taste and a sense of humor, but they can’t all have great taste.” - Carrie Fisher in When Harry Met Sally.

It’s Judgement, Not Taste.

Leave a comment

Share Let's Get Real

The human mind evolved to help us manage the present, and prepare for the future. As a result, it absolutely sucks at remembering the past. Until we hear a song, or catch a smell that takes us right back.

Contrary to popular belief, memory isn’t a tape recorder. It’s a mish mash of emotions and fragments that our brain compiles a story from. Our minds are buggy wetware, and our memories are deeply untrustworthy.

On top of it, each time we have one of these flawed memories, it shifts based on our knowledge and mood at the time we recall it. And we cannot make decisions if we lack emotions, for all decisions are filtered by emotions. (Yes, even your CTO who thinks he’s Mr. Spock.)

AI is trained on human OUTPUTS: writing, music, art, law briefs, code documentation, how to articles; podcasts. Outputs are the synthesis of our thoughts, experiences, training, belief, feelings, and experiences.

Outputs are generated from INPUTs, such as:

• What does hazelnut gelato taste like in Milan?

• Why do my my mom’s fried potatoes taste better than any pommes Lyonnaise I’ve had in a French restaurant?

• The sensation of music so beautiful it makes me cry

• The delight of hearing a child’s laughter, or the wonder I see at a space launch

• The instinct we have to fight or flight; laugh, cry; all subconscious and instinctual.

• The lengths we will go to protect our families

• The way a pet soothes our parasympathetic nervous system; and why we can’t seem to stop searching for them, months or years after they’re gone.

  
  

Outputs are how we make meaning of all our inputs. We take these inputs such those above; and make inferences and connections all the time based on these fragments of memory and experiences; ones we often never write or talk about or in some cases; even remember consciously. Some are hard wired into us from millenia of evolution.

Everything ever produced by human beings only represents a small fraction of the human experience, and capacity to create, to feel and sense, and to love, fear, and yes even hate.

The magnificent creative leaps of humanity across all disciplines: art, science, medicine, engineering, music….these all come from these same fragments of barely understood experience. Often their creators have no idea where the idea comes from. So how could AI?

An AI may, over time, begin to infer the inputs from the outputs, across a large enough set of data.

But AI will never, ever, ever, be human.

For AI does not die.

Therefore, it cannot live.

And so it is that AI can never, and should never replace humans. It can ever only be like a cassette recording of a song off the radio: useful to a degree in specific situations, but inferior from the start; and by no means a replacement for the original.

But just in case, I’ll still keep saying please and thank you to it, anyway; because my mom taught me to.

*There is a lot of debate about the forgetting curve. So let’s not quibble about statistics. The point is that human memory deteriorates rapidly; and there is broad scientific consensus that the human brain evolved to process information, not record it.

Part One: Human on a Bicycle

Like nearly everyone, I see the future (/sarcasm.) Seriously, I see want a world where humans are able to spend more of their precious lives in leisure, with loved ones, learning for its own sake, free of poverty and disease. I got into the tech industry because I was excited to use tech to help the human condition.

As a middle schooler, I was fascinated by Steve Jobs, and I won my 8th-grade stock picking competition by choosing Apple. One of my all-time favorite “Steveisms” is his discussion of efficiency in the animal kingdom, where, as it turns out, a human with a bicycle is more efficient than any other animal:

Apple’s brand thesis was “Bicycle of the Mind” for a very long time. (I so miss those days when we weren’t always and only scheming about how to extract money from other humans. RIP Steve, we will not see another like you any time soon.)

Part Two: The Sorcerer’s Apprentice

About 18 months ago, I was invited to moderate a panel discussion for the American Marketing Association San Francisco Chapter on Humans in the Loop. Everyone was scared. We hadn’t talked about prompt engineering or Claude vs. OpenAI; we just came together as a collective to share what we knew at the time.

Here was my opening statement:

…as powerful as AI is, we can’t afford to be passive consumers of its output. Remember The Sorcerer’s Apprentice? It’s been on my mind lately. Like Mickey, we’re thrilled to command an army of AI ‘brooms’ to do our bidding. But remember what happens when Mickey falls asleep? That’s AI without humans in the loop - full of potential, but with the ever-present risk of chaos if unchecked.

Nothing I have learned about AI since October 2024 has altered my thinking on this. The story goes south when Mickey falls asleep. (IIRC, AI helped me craft that paragraph, as evidenced by the telltale em dash.)

Part Three: The Economics of Sam Altman (et al)

I started seriously working with AI in January 2023. Back then (when I still thought I was late, LOL), I thought about the wildly variable and unpredictable token costs for in-house AI models, and concluded that only a 10x return could make an AI bet positive ROI for anyone but the most price-insensitive companies (those facing extinction). Most non-tech companies do not have 10x projects sitting around unnoticed. At the time, I was uncertain of my knowledge, so I didn’t talk about this publicly, but it’s what I advised my clients, and what others who knew more than I did validated.

Now I am very sure reasonably confident that I am right.

And while, yes, prices have declined faster than PC compute or dotcom bandwidth prices, the price range for frontier models is, as of this writing, $15-$75 per million output units.

Executives looking to invest need a reasonable understanding of costs to proceed, and the problem of token cost variance is compounded further by the unit economics of variable costs.

Part Four: The Thin Slice

And for every minute of the last three years, I’ve had this creeping anxiety, like being followed slowly down a street, looking behind me, and noticing the monster was starting to pick up pace, so I had to pick up pace too…

Each idea I had, the models duplicated before I could scarcely articulate it. Each product category I considered (and many I did not) seemed to be owned by an “incumbent,” only to be disrupted by the models, new incumbents, and so on.

When Claude Code shipped, I saw the equivalent of tech’s evolution in my 20+ years happen in less than 30 days. I watched as machines abstracted more and more and more.

After the 200-page unshipped code fiasco I mentioned in the last post (about 45 days ago), I stopped building and decided to think things through. These things included:

• My own experience advising companies that adopted Salesforce and Oracle early, who are now often locked in with exorbitant switching costs- and their accompanying sclerotic cultures where innovation goes to die. (look! a natural emdash!)

• How flagship/frontier models are not only subsidized by many orders of magnitude (if rumors are true, that $200 Claude Pro Subscription has $5k of computing costs). What will happen to all these enthusiastic builders’ hopes and dreams when token costs start to reflect reality? I cannot afford another hit to my retirement savings at this juncture in my life. I’m not afraid, but I’m not stupid, either.

• The ensuing enshittification that will follow (for those who remember the deep magic, this is as important as the Peanut Butter Manifesto, no?)

• The Models stated plans to become metered utilities, or if rumors are true, even the browser

• …and the not remote possibility that everything we know now could be equally upended in June 2026 (e.g., in another 60 days) if Apple decides to become a tollbooth collector on MCPs

• This is when I started to think about what my real value-add was in a world where machines are taking over more and more every day. I’m deliberately setting aside the existential conversations; they are of limited practical use to me at this moment.

Keeping my promise and not using AI, while accepting that my art skills are terrible, and being efficient while not mucking with an image editor all afternoon, I drew the concept on paper yesterday:

I fear this may be an oversimplification. But I’ve been thinking and talking about this idea for so long, I fear I might explode if I don’t just say it out loud, so I can stop talking about it and point people here.

Maybe I’m punchy because it’s Friday. But right now, I feel like this. There’s a Borg metaphor, too, but I liked this one better today.

Based on my conversations in the last two weeks, I don’t think I’m alone. (I love the phrase AI Brain Fry. It’s real.)

Anyway, I promised myself and you I’d write about this today; so there it is. Talk to me in the comments or my DMs.

Next week, I’m going to write about what I have decided to do next to find and preserve My Thin Slice. (No, it’s not a cabin in the woods off the grid.)

Leave a comment

Footnotes

*I am specifically not using the word predictably
**More advanced practitioners already know this; consider it proof of my self-education.
*** I mean total cost, not just token cost.
****I don’t think I’m uniquely enlightened on this point, and the Masters of the Silicon Valley Universe laying people off right now are being disingenuous when they say that AI is going to take all the jobs. At least, not by next Tuesday, which is what they want us all to believe.

EPILOGUE: More on the Narrow Slice…

At the end of yesterday’s post, I was talking about the “Narrow Slice” in Enterprise AI. But I’m not sure I clearly defined it. A quick summary:

• At this moment, it seems to me that the only way to make AI behave consistently* is to constrain it to a very specific, well-designed, and well-documented task.

• To use agentic workflows is to tie these narrowly defined agents together with a similarly well-designed and documented series of agents, carefully orchestrated to review and check the previous agents’ work, and to identify the process for escalating to a human for review and/or approval.

• All of these behaviors need to be rigorously defined and documented in relation to the level of autonomy (and related risk) we plan to give them. Suggesting a reply to an email? low risk. Making autonomous purchases on the internet? Less so.

• After I posted yesterday, I learned that this phenomenon has a name: Ashby’s Law. Upstream from Ashby’s Law is Tesler’s Law, which says that every application (which, in my mind, is a task) has an inherent level of complexity that cannot be removed, only shifted.**

TL; DR: When it comes to AI in the Enterprise, there is no free lunch.

When we use AI for an enterprise case, we are only shifting the work humans do onto machines, and this decision has its own costs***, which, overall, will be roughly equivalent to the original work, at least in complexity and documentation.

That does not mean we should not transfer work. Making an airplane isn’t easy, and its value to humanity is enormous. But we have to accept that costs exist and not be disappointed or surprised that an LLM can’t magically replace employees.****

Whew. Thanks for letting me clean that up. I will explain these criteria in another post.

Leave a comment

And what I saw made me put on my brakes immediately. I advise high-stakes enterprises, so I wanted to know as much as I could first, try to disprove my own hypothesis, and think deeply about what was going on, before I could explain it to anybody. I’ve had a lot of conversations about this topic; this piece is my attempt to “learn what I think” by structuring it in writing.

But it is also a pressure test of my thinking about the state of AI at this moment in time- with you. I welcome the discovery of my assumptions, errors, and ignorance.

Part One: Anyone Can Vibe Code

I watched in my “learn about AI” community as the people passionately committed to the new world barreled down the path to MCP, Claude Code, and then OpenClaw, building, building, and building; toiling late into the night and on weekends; fervently trying to get in on the gold rush. Two specific situations stuck out to me**

1. 27 agent framework with an orchestration harness, aka “your shop in a box.”

2. Spinning up a Sarbanes-Oxley MCP in 10 minutes, aka “your lawyer in a box.”

Both of these demos freaked me out for what I thought were different reasons, but they turned out to be the same one: **

How will we know if they are wrong?

• And when they are wrong, how will we diagnose and fix them?

• When they fail, what are the risks?

• Which risks are manageable, reversible, and predictable? Which ones are not?
  

After my personal experience with almost releasing my Claude API Key to the internet after Bolt’s recommendation, along with the “confidently wrong” experiences we have all had, I believe I had a duty to myself and my enterprise clients to take a big step back and really think.

Part Two: 200 Pages of Code Yet Nothing Ships

I was, and am, so excited about the speed and quality of the builds I saw. I have worked in Silicon Valley since 2002; I know what a gold rush looks like. Like everyone, I wanted to be part of it. It’s always bothered me that code was largely a male priesthood that blocked anyone not technically inclined from building. I have ideas, lots of them!

I decided to get in the mix and try to build. I learned (and am still learning every day about the levers of tokens, attention, RAGs, context, windows management, state, and boundaries.

The week I spent having AI churn out 200 pages of code to try and codify my 25 years of experience in delivering insight from marketing data; then realizing it was all crap (or at least was not verifiably secure or trustworthy) was a rock bottom moment: it was the first time in my life that my brain, which is an insatiable and often annoying sponge of information, was so full, that I felt physically ill and needed to stop; almost like eating too much rich food. I felt like a pinball machine on TILT.

At first, I thought the problem was that I am not a developer and won’t be a Python expert at this stage of my life. So how on earth could I offer anything to clients I couldn’t stand behind? Again, how would I know if the thing I was building was safe, or in some cases, even accurate?

I honestly considered hanging up my spurs and calling it a day. Perhaps I was too early; perhaps I wasn’t technical enough.

One of the tips I give people about AI is to stop when you feel like you’re banging your head against the wall for too long- open a new window; start over, and most of all, don’t suffer alone. I had a lot of imposter syndrome. So I reached out to my community, and as it turns out, literally everyone working in the space right now(developers and non-devs) has imposter syndrome and FOMO, which feels like the existential fear that we’ll all be in the permanent underclass if we aren’t AI-native by next Thursday. (and if someone claims they aren’t feeling FOMO and Imposter Syndrome right now, beware; they might be confidently wrong.)

Part Three: The Trough of Disillusionment

The coders weren’t immune either, as Jonathan Flower hilariously describes:

It had taken me 4 days to get to 80%. 3 more days only got me to 90%. - Jonathan Flower

So I started asking more questions and looking more closely and saw others struggling with the same problems.

• People would show me demos that “worked this morning, but I don’t know what’s wrong now!”

• I asked how many paying clients they had. How many were enterprise clients? How much were they being paid?

• One by one, all confessed to me that the product was in beta, or a demo, or a pilot. Not a single one showed me a repeatable process with high-stakes reliability, security, etc. (If you are thinking, these are tests, Melinda, all software has a demo and beta period, stay with me.)
  

Then I had thought: Would I indemnify a publicly traded company with a Sarbanes-Oxley MCP? (Of course not, Melinda; the point is to have an expert work hand in hand with the agent. Again, I get it. Stay with me.) The answer is, of course, not right now and certainly not me right now.

So now what was I going to do?

I took a previously planned vacation and didn’t use AI at all to map out efficient routes at the San Diego Zoo.

Part Four: Am I The Problem or Not? Yes? Maybe? IDK.

I’ve mentioned the power of community at this point in our history. It’s so important, and it reminds me of the early days of the ‘Net, where we were solving problems collectively, sharing information, arguing, growing, rebuilding, commiserating, inspiring, “Rocks In A Tumbler” style.

In my experience, the moment you feel truly lost is when you become receptive to something new.

The specifics aren’t important to my story, but by admitting how lost I was, I think I have found my root cause:

I’m a non-linear thinker, and the only way right now to make AI behave consistently (I am specifically not using the word predictably,) is to constrain it to a very narrow slice of behavior, then tie it to other similarly constrained agents, within a very constrained, prescribed framework of behavior, with checkpoints and reviews. The behavior must be rigorously defined and documented in relation to the level of autonomy we plan to give it.

This sounds suspiciously familiar to me. Like, isn’t this what we do with kids? We don’t let middle schoolers drive, no matter how tall they are. The reasons are obvious: their brains haven’t developed the maturity/judgment to not create a catastrophe for them or someone else.

Instead, we wait until a child reaches a developmental age at which the idea of consequences and concern for others is possible. Even then, we insist on Driver’s Ed, and we make them get learner’s permits; and we ride, white-knuckled with them while they demonstrate that they can handle the responsibility.

And even still, some will get drunk and drive, or ride with someone who is, and they may or may not get lucky and not experience catastrophe; but by and large, most will arrive at adulthood able to drive a car, with at least some level of maturity and not hurting someone else; most of the time. (Exception: Los Angeles. Y’all are crazy. 🤪) Every so often, we as a society insist that everyone take a driving test again, just in case that judgment is slipping away. And someone who is repeatedly shown to demonstrate poor judgment can be stripped of that license.

So, from where I sit, the challenge and opportunity in Enterprise is the same, because agents, at least right now, are not superhuman robots. They’re middle schoolers. Not even teenagers. Enterprises have to ask themselves:

• What is the task to be done? Can it be defined in a way that is useful, measurable, and trainable? How will we know this before we invest? Can we know? If not, what is the narrowest slice we can test on?

• How will it be taught in a reliable and consistent manner?

• How will the task be tested and evaluated? How often?

• When must we check up on it?

• How will we prevent catastrophe? What is a catastrophe? Is it really a catastrophe? Can we survive it? If we can’t, how will we prevent it?

I call this first step in the enterprise AI Cycle “The Narrow Slice.”

That’s enough for today, I bet, for you and for me. Tomorrow, I’m going to write about what I think the only defensible intellectual property might be for individuals in the years to come.

You know how the algorithm works, so if this resonated with you, I’d very much appreciate a comment/like/and/or share. I don’t do this for ego, but knowing my thoughts aren’t going into the void is a useful motivator and helps me grow in my thinking.

Share

**Note to the two creators who will recognize yourselves: this is NOT about you and your vision of the world. I could be very wrong here, which is why I submit my analysis with humility and respect, and anonymize this post. Your creations are only exemplars to me, and this piece is not a specific critique of you or your builds.

Leave a comment

Uh, no. I did not, thank you very much.

My stomach dropped. There it was, my API key sitting in the HTML like a credit card number written on a bathroom wall. 1

How Did We Get Here?

The path to this situation was paved with reasonable decisions

1. Built prototype in Bolt because it’s very fast for mockups, and has better UX capabilities than Replit. Plus they were a sponsor-I wanted to return the favor.

2. Showed Up Ready to Ship and showed up with more than was needed: an api key and detailed plan. I didn’t show up to prototype.

3. Hit the backend wall. Couldn’t figure out the API integration. I’m not a developer, and demoing with fake data would fail to showcase my beautiful and carefully crafted prompt.

4. Followed Bolt’s suggestion. Put the API key in the HTML.

Each step makes sense in isolation. Together, they created the perfect potential storm of financial disaster.

The Everest Principle

Jon Krakauer's "Into Thin Air" makes this point brilliantly: disasters come from a series of small mistakes in an unforgiving environment.

That was me. Pushing for the demo, blowing past basic safety features I didn’t even know to look for.

Eventually ,I spent a full day rebuilding DeckCheck in Replit2, but I couldn’t stop thinking about the near miss. What if I hadn’t had a friendly expert to catch this? And I sat Bolt(!) upright in bed, thinking:

What if this is happening at scale?

The Manhattan Project Moment

When I watched Oppenheimer last year, I thought a lot about AI. Many of us are like those scientists-entranced with the possibilities; naive(at best) about the consequences.

Non-developers like me can build production-grade applications. This week alone, I extracted 15 years of messages and photos from conversations with my mom, then created a password-protected site to store our memories. Today, DeckCheck analyzes one slide. Tomorrow? Who knows.

AI is our millenium’s printing press. Developers will no longer be a priesthood, gatekeeping products. Creativity and humanity can be unleashed in technology. The possibilities thrill me….

…ands scare the daylights out of me.

Speed Amplifies Everything, Including Mistakes

The vibe-coded tools—Bolt, Cursor, V0, Lovable, Replit—optimize for velocity. They abstract complexity, automate tedium, let us focus on WHAT we're building rather than HOW.

That's their superpower. Also their blindspot, especially in the hands of a rookie like me.

Consider the escalation ladder:

• Exposed API key in personal project = Expensive lesson

• "Cookie-free" analytics tag in production = Many vibe coded tools still fingerprint devices. In the EU/UK, that requires consent. Failure to collect that consent can lead to significant fines.

• AI agent with exposed credentials = Customer database on the dark web

We've distributed the power to detonate an enterprise to anyone with a credit card and 30 spare minutes.

We Need to Discuss Systemic Risk

This isn't about individual mistakes. It's about what's happening right now:

Marketing teams deploying production code without security review

AI Agents making autonomous decisions with unclear accountability

Solo Founders shipping at venture-backed velocity with security as just a mere roadmap item.

Governance Frameworks assuming IT involvement that may no longer exist

Every API integration is a potential breach vector. Every AI Agent is a decision maker. Every SaaS tool with a Javascript snippet deployed in Google Tag Manager without governance is a privacy time bomb.

Rebuilding Our Collective Immune System

My exposed API key wasn't just a personal wake-up call—it revealed that the guardrails we're used to (which were also gatekeepers to innovation) don't exist in this new world. We have great power. Now we need great responsibility.

We're the first generation building at the speed of thought. We're writing the playbook as we go. Every near-miss is a lesson. Every save is collective wisdom. Every prevented mistake is future prosperity protected.

What is your near miss story?

What security blindspot did you discover just in time? Hit reply, or dm me privately. I will collect these stories for a follow up piece on building our collective security instincts. Because if don’t share what we are learning, we’ll all learn the hard way.

Leave a comment

P.S. If you've added any new marketing tools, analytics, or "instant solution" apps to your website recently, here's a question: Do you actually know what information they're collecting about your visitors? Or where that data goes? Most of us non developers don't—and that's exactly the problem.

Now imagine that you had an alchemist. A person who could quite literally turn lead into gold. Would you ask them to go find the lead, extract it, package it up, transport it, store it, and unbox it?

Hell no. You’d want that person turning lead into gold as fast as they could.

And yet, that’s exactly what my life -and the life of all us great marketing analysts-is like. We have to spend at least half our time working putting data into a format we can use to perform our alchemy; such as building data warehouses, piecing unstructured data together, normalizing it; and more, much more.

I’m an expensive resource, because there are not enough of me to go around. I have to turn away promising and valuable small businesses as clients because I can’t afford to work with them. It’s not fair that only big companies can benefit from my expertise. But the truth remains: lots of people can find the lead, but there are not enough alchemists in the world.

I want to replace myself with AI because a) there’s so much more gold I could create from the same lead given AI’s capabilities, and because b) I believe every company in the world-small or large- deserves to have someone with my decades of expertise and training helping their marketing team turn lead into gold.

But.

There is currently no safe or accurate way to layer ChatGPT onto Google Analytics for most companies, no matter what the startups are telling you. The core issue lies less in data security (though it’s still an issue because this sensitive information), and more in data availability. The fact is most “Marketing Analytics AI” companies don’t have enough primary data to train safe AND effective models. And most companies don’t have enough data to do it themselves, because they’re just one company.

Even gigantic companies with lots of data will struggle with this problem, because there aren’t enough AI and ML engineers available to solve this problem on a one-off basis. Given current costs, it will be hard to justify the return on investment for individual companies unless they’re spending $50M or more on advertising a year. And that assumes higher ROI/NPV projects aren’t pushing it down on the priority list. Let’s face it, marketing rarely gets R&D budget. :/

This leaves anyone attempting to solve this problem with two options that have significant challenges:

1. Synthetic Data Creation: While generating data based on patterns is cheaper and easier, it falls short without domain experts like me to validate real-world applicability and catch training data flaws. As I said above, we’re in short supply, and we are not cheap.

2. Data Acquisition from Multiple Sources: Trust is a significant barrier, as few companies can secure enough primary data from various businesses to train robust models.

As a result, I believe that at this time, only three companies currently possess the primary marketing data required to replace me at scale.

• Google - due to the broad and deep information it has on many sizes of businesses, including conversion information; plus Google Ads and Performance Marketing, plus email. I’d be shocked if they weren’t thinking about it-but given the current existential threat to search, I would expect this to be de-prioritized. They don’t monetize GA that well, and it would be hard for them to see how valuable this could be.

• Adobe - For similar reasons to Google, it has incredible primary data in Analytics. I worry that as a “Creative Cloud” company it too may struggle to see the opportunity in this part of its business.

• Salesforce - I have incredible hopes for them. Tableau is the obvious angle here, but their massive data sets in Commerce Cloud and Marketing Cloud; and relentless focus on the right customers could prove to be a real player.

*see footnote for explanation of who I left out. (And hey, it’s the internet, so I expect you’ll school me in the comments. Bring it on.)

So What?

1. If you're considering investing in a tool to replace a skilled analyst like me with AI, you might end up wasting your money. The technology isn't there yet unless the company in question has access to massive, diverse data sets.

2. If you work for Google, Salesforce, or Adobe and are looking for domain expertise to leverage your data for AI-driven insights, let's connect. I want to help you help me replace myself with AI. This is a generational opportunity to unlock shareholder value. It is not a technically hard problem to solve.

3. If you're not one of these three but believe you're genuinely solving this problem, please reach out. I’m happy to be wrong, so I can focus on my alchemy.

* Why not ______?

-Microsoft - I may be proven very wrong here, but based on what I know right now; it has no significant marketing analytics assets. LinkedIn is a fraction of the story. Only fools count Microsoft out of any market, though. Perhaps with the Azure and cloud products there’s an angle, but again, customers would have to turn it over. I don’t yet see the path, yet.

-Amazon only ha ad data for its merchants and its stores

-Facebook can only use its own data, same problem that LinkedIn has. EDIT: 8/23: With Meta’s announcement about Google Analytics Integration, they might be able to change my mind….

-ChatGPT can only do this if they’re totally violating their user agreements, and given their broad ambitions, I don’t see this as making the cut in terms of importance. (see also Anthropic but it can’t do data so 🤷🏻‍♀️)

-Did I miss anyone?

--EDIT: Yes, of course, I missed Apple. Like Microsoft, only fools count them out of any game. They have enough cash to buy their way in. But in this case, as things stand today, they too have no marketing analytics data to speak of. They have their own, of course, and it’s not nothing. But to build this tool at scale, they’d need to find a way to get more of it.

* AI Disclosure: I used ChatGPT4o to help me with reviewing and editing this piece. I came up with the alchemist analogy.

Leave a comment

Come Play With Me!

Hello, everyone. Normally, I really try to keep politics out of my Substack. I’ve had enough of it on Twitter to last for several lifetimes. I have personally found “hashtag activism” to be limited in its effect; and in the aftermath of breast cancer, I have chosen to disengage from most media sources in the name of self care.

However.

We live in a world where anti-Semitism is on the rise. I believe it is important for everyone to stand up to bigotry and hate of all forms. So I, along with many substackers, are posting this collective letter for visibility. The original newsletter was drafted by Marisa Kabas and a group of Substackers much more visible than I am; and I choose to take my space and devote it to this cause.

As it says at the end: If you are a publisher and the letter resonates, feel free to post it on your Substack page. They’re planning to keep track of the signatories here. Thanks for reading.

_____

Dear Chris, Hamish & Jairaj:

We’re asking a very simple question that has somehow been made complicated: Why are you platforming and monetizing Nazis? 

According to a piece written by Substack publisher Jonathan M. Katz and published by The Atlantic on November 28, this platform has a Nazi problem: 

“Some Substack newsletters by Nazis and white nationalists have thousands or tens of thousands of subscribers, making the platform a new and valuable tool for creating mailing lists for the far right. And many accept paid subscriptions through Substack, seemingly flouting terms of service that ban attempts to ‘publish content or fund initiatives that incite violence based on protected classes’...Substack, which takes a 10 percent cut of subscription revenue, makes money when readers pay for Nazi newsletters.”

As Patrick Casey, a leader of a now-defunct neo-Nazi group who is banned on nearly every other social platform except Substack, wrote on here in 2021: “I’m able to live comfortably doing something I find enjoyable and fulfilling. The cause isn’t going anywhere.” Several Nazis and white supremacists including Richard Spencer not only have paid subscriptions turned on but have received Substack “Bestseller” badges, indicating that they are making at a minimum thousands of dollars a year.

From our perspective as Substack publishers, it is unfathomable that someone with a swastika avatar, who writes about “The Jewish question,” or who promotes Great Replacement Theory, could be given the tools to succeed on your platform. And yet you’ve been unable to adequately explain your position. 

In the past you have defended your decision to platform bigotry by saying you “make decisions based on principles not PR” and “will stick to our hands-off approach to content moderation.” But there’s a difference between a hands-off approach and putting your thumb on the scale. We know you moderate some content, including spam sites and newsletters written by sex workers. Why do you choose to promote and allow the monetization of sites that traffic in white nationalism? 

Your unwillingness to play by your own rules on this issue has already led to the announced departures of several prominent Substackers, including Rusty Foster and Helena Fitzgerald. They follow previous exoduses of writers, including Substack Pro recipient Grace Lavery and Jude Ellison S. Doyle, who left with similar concerns. 

As journalist Casey Newton told his more than 166,000 Substack subscribers after Katz’s piece came out: “The correct number of newsletters using Nazi symbols that you host and profit from on your platform is zero.” 

We, your publishers, want to hear from you on the official Substack newsletter. Is platforming Nazis part of your vision of success? Let us know—from there we can each decide if this is still where we want to be.

Signed, 

Substackers Against Nazis

This is my personal journal about my learning journey in AI and ML for marketing. With over twenty years of experience managing and marketing tech companies, I have deep knowledge of marketing analytics and attribution, always focused on generating measurable and meaningful business results. For the last nine years, I've advised and served marketing leaders in tech companies, enabling me to share my insights on the impacts of AI and ML, including privacy and ethics issues, through this Substack. You can learn more about me, including my policy on using AI for this blog, in the "About" section of this Substack. I welcome your suggestions for further education on AI and ML.

In This Issue:

• Prompting without Pain

• Celebrating Humanity

• Playing with Generative AI Images

• What’s Next

Prompting Without Pain

I don’t tolerate boredom well. And damn am I bored with all the prompting guides out there.

I’ve gone in and tested some of these prompts people give away for free, and damn if I’m not bored as hell by the copy they generate. It’s so obvious, isn’t it?

I’m also lazy efficient, and so I don’t have much patience for long, drawn out conversations with a chatbot.

Herewith, some of what I’m doing with prompting these days.

Some notes:

1. I use the paid versions of ChatGPT 4 and Claude.ai.

2. Unless otherwise noted, assume the output was not further refined.
   

1. Talk to them like a human. (But never forget they are not human. )

For all the fussing about getting prompts right and trying to remember all the crap you’re supposed to put in each time, the best way I’ve found is to simply talk to them as though they were human. But as the title says, it’s important to remember these are not humans talking to you.

When talking as though you were speaking to a human, the model learns about your own cadence, how you think, and infinite other things that will help it return a better response. The more you write, the more it has to work with.

Consider for a moment the response to these 2 prompts in ChatGPT 4 with Bing Search enabled:

Prompt 1A:

Please review https://fiddleheadhq.com and create 5 linkedin posts that are on brand for me and for fiddlehead.

Output for Prompt 1A:

Prompt 1B: Speaking like a human.

Write a clever linked in post in the style of Melinda Byerley that encourages marketing leaders to subscribe to my substack. My substack is about my personal journey to understand the impact of AI on digital marketing, especially marketing analytics. I have more than 20 years of experience in the field. limit prose.

Output for Prompt 1B:

Keep it human by doing the following:

• Be specific. If you ask for a general answer, that is what you will get. If you ask for a specific answer, that is what you will get. You can do a lot of back and forth with the models, but unless you’re on a paid plan, you’ll run out of credits fast. Think through that prompt before you start, and give it as much information as you can gather. Unlike humans, the model can’t be relied on to infer things that might be obvious.

• Be detailed. Give the model as much detail as you can to help it decipher what you actually intend. Unlike humans, the model doesn’t tune out when you ramble on. The more parameters you offer it, like the length of posts, or whether to include emoji, etc. the more you’ll get that is actually new and surprising.

• Be liberal with descriptors. Adjectives, styles, tones, fears—use them all in your prompts to nudge the response to your desired outcome.

  .

2. Use the models as a thought partner.

One of the benefits of being a paid subscriber to these models is the ability to go back and forth with it until you get what you want. This has led me to have some conversations with the models that have helped me get breakthroughs in how I do work.

Here’s just one example. Because I’m a professional marketer, it might shock you that I struggle to self promote. I have all kinds of blockers about it-the fear of being boooooring, my loathing of bad copy and inane, general posts, and more recently, my realization that posting on social media in anything approaching a more authentic way was generating engagement that I simply don’t want to spend my life energy on.

One day, I took the problem to the models.

Prompt 2 - To Claude.AI:

claude, I know I need to self promote more. I hate the reply guys on social media, low value commenters arguring with me about things they don't know anything about.

I know I should ignore them. I worry that by not answering them, people will agree with them, or run away from the conversation.

do you have any advice?

Prompt 2- Claude’s Response:

Maybe this kind of help is obvious to some people. But it wasn’t to me, and it helped me a ton. I asked it for some follow up ideas and this was one really great one.

I cannot begin to tell you what having these phrases in my back pocket has done for my online sanity. I’ve used thought partnering for all sorts of problems like my company’s mission statement, determining how to approach all kinds of problems at work and at home, and for brainstorming news ideas for posting on social media.

Tips For Using LLMs for Thought Partnership

• You’ll need a paid account. LLMS are priced on tokens, which is essentially how many words you use. When you start going back and forth, you’re using up processing capacity, and thus the charge. Even on paid ChatGTPT 4, a deep session I’m having can get me throttled, and I’ll have to finish the next day.

• But wait, you’re saying. You told me to speak at length! I did, but that’s for solving a different kind of problem. Thought partnership is different, and when I’m workign with the model on that, I know it’s going to give me limitations.

• And if you’ve run into these limitations you now understand the power of “prompt engineering” which is really about being as efficient as one can to get the outputs one wants. Over time, as these tools are baked into our every day life, you’ll need less prompt engineering and will speak to the models like I do here, but with far less limitations.

• Don’t be afraid to be afraid. Tell the models why you are afraid to implement their suggestions. Tell them why you think stuff is BS. Unlike human beings, models don’t get their feelings hurt when you argue with them and criticize their output.

3. Try different models for the same problem.

A lot of noise lately on how ChatGPT4 has gotten “worse” lately. I think it’s more about safety and backlash and cost savings. But also, new models are different and it’s worth seeing what they do well and how.

Here’s a very simple and straight forward example of the output for the same post on just ChatGPT 4 and Claude.AI. I asked both models to help me write a much shorter bio for placement on the bulletjournal.com website, where I’m to be listed as one of the first officially authorized trainers of the Bullet Journal® Method.

Prompt 3:

Please rewrite my short bio for the bulletjournal.com website. the bio must be 2-3 sentences at a maximum.

Melinda Byerley is an marketing analytics trailblazer with over 20 years of experience spearheading data-driven growth strategies at pioneering Silicon Valley tech firms like PayPal and eBay.She was an early adopter of Bullet Journal as she learned to manage her ADHD, and is excited to help other neurodiverse executives learn the method to manage their lives, be more productive, and find a sense of purpose. Melinda holds an MBA from Cornell University and a BFA in Drama from Illinois Wesleyan University.

Response 3A: ChatGPT-4:

Response 3B: Claude.AI

Which version is better? That’s for you as the author to decide.

Tips for using Different Models:

• Feed one output to another. Ask ChatGPT 4 to analyze the output from claude, and vice versa.

• Merge content: take the best of both or more and ask the model to combine them.

• Proofreading: I always ask the models to proofread before posting. I say, “please review this ONLY for spelling, grammar, and punctuation.”
  

  Thanks for reading Let's Get Real. Subscribe now to get notified of new posts, and please consider a paid pledge to encourage me to develop more useful content about AI in Marketing.

4. The Power of Voice and Style

I have hesitated to share this tip because in my recent experience it’s one of the most powerful ways to use an LLM. If I announce my sources to the whole world, I’ll lose that arbitrage right now, and so I’m going to share with you a very obvious and simple example. I will let you think about which voices and styles I might use in my own work.

I gave Claude.ai a simple request and had some fun.

Tips for using Voice and Style:

• Proceed with caution: potential for copyright infringement. If you’re using a modern author, there’s a very real risk you’re using their exact words. Real writers who use beautiful phrases deserve to own them. Don’t steal. Google those phrases and make sure they aren’t already in use.

• Combine and remix styles: One way to get a style that’s more yours is to use different voices in the same prompt. Another is to take phrases from various styles and remix.

• Prepare to spend some time assembling. I used this technique to come up with a post that is not only one of my favorite linkedin posts of all time, but one of my more successful in recent months. It took me over an hour to craft this, but I’m proud of it, and I’m confident that it’s not violating someone’s copyright. (I took the photo myself.) But it all started with the exact prompt you see above.

  

Celebrating Humanity: AI For Good

Crikey. If this holds, it’s a game changer. Scientists can test for diabetes by speaking into a smart phone.

Playing With Generative AI Images

I’ve been sound asleep on how fast Adobe Firefly is improving. On top of it, the user interface can get me to better images faster without requiring as much in depth knowledge of their parameters. I’m especially impressed that they now indemnify users because they’re working with a library they own.

”Getting a cup of coffee with a robot best friend” (Photo, Hyper realistic)

What’s Next

• So many podcasts coming out soon! Follow me on LinkedIn for more.

• Taking piles of AI Classes thanks to the generous scholarship from Sectionschool.com in light of my breast cancer treatment.

• Planning Season 5 of Stayin’ Alive in Tech, focused on AI

• I’ve rolled out new pages on the Fiddlehead site for Marketing AI Strategy Consulting, Bullet Journal® Coaching, and more!

Leave a comment

This is my personal journal about my learning journey in AI and ML for marketing. With over twenty years of experience managing and marketing tech companies, I have deep knowledge of marketing analytics and attribution, always focused on generating measurable and meaningful business results. For the last nine years, I've advised and served marketing leaders in tech companies, enabling me to share my insights on the impacts of AI and ML, including privacy and ethics issues, through this Substack. You can learn more about me, including my policy on using AI for this blog, in the "About" section of this Substack. I welcome your suggestions for further education on AI and ML.

In This Issue:

• MAICON 2023: What Marketing Leaders need to know.

• Want A Video of This?

• Celebrating Humanity

• Playing with Generative AI Images

• What’s Next

MAICON 2023 Headlines for CMOs

Overall, MAICON 2023 was a blast! The food and logistics were all top notch and the people were friendly and curious. I can’t thank Cathy McPhillips enough for her community building and logistics support. I paid for my attendance and travel, and no consideration has been given to me for my post or comments.

I wrote this post 100% and have decided to not even edit it with AI to preserve my own voice. So any mistakes are mine, and I thank you in advance for your patience.

1. AI In Marketing Is A Once-In-A-Generation, Career-Defining Opportunity. For the Top Line as Well As For the Bottom Line.

Last year, MAICON had about 300 attendees. This year there were nearly 800, blowing well past the goal of 600. Those of you burned by crypto, Metaverse, and/or Web 3.0 might be feeling a bit wary, because some of the same characters are saying some of the same things. (“disruptive.” etc) People who have followed me for a long time know I rolled my eyes at this stuff.

If we’re talking about AGI (the stuff we see in science fiction), yes, that’s a long way away, and it’s pretty much not something that’s something mere mortals are going to be accessing in the near future. If you’re a CMO here to learn about what this all means for your career, you can safely set aside most “future of humanity” discussions in the near term, unless you work in the industry, or have a lot of time on your hands to feel frightened and helpless.

Ah, but the AI I’m talking about is going to transform our lives in a way no less profound than the personal computer, or the internet. I’ve witnessed four major technology shifts in my lifetime: the PC, the internet, the mobile phone, and social media. This has all the earmarks of that kind of change. And the gold rush is on, and you should be looking everywhere you can to find a way to participate.

If you don’t want to believe me, believe the data on adoption:

• MIT Sloan and BCG, and Oxford’s Institute for Ethics Survey of 1200 companies: “The research suggests 78% of organizations report accessing, buying, licensing, or otherwise using third-party AI tools, including commercial APIs, pretrained models, and data. More than half (53%) rely exclusively on third-party AI tools and have no internally designed or developed AI technologies of their own. “

• No, this isn’t just about ChatGPT. Even back in 2021, HBR was reporting the “skyrocketing use of AI”:

  86% of PwC’s survey respondents said AI was a “mainstream technology. “ In 2021. Sobering for those of us just getting on board. (but it’s not too late, don’t despair!)

• You HAVE read McKinsey’s AI report from last week, right?

  

On the mainstage at MAICON, the speakers gave us chart after chart, story after story of how AI can and is being used by many companies of different sizes and at growth stages.

The fact is that big companies have been working on this for close to a decade, and that’s not just the big Tech companies. AI is being used to drive both strategic competitive advantage as well as cost and efficiency savings. Here are some publicly available articles about stories I heard in the hallways that will help you see past the generative AI hype with ChatGPT to the enormous opportunity right in front of us all.

• McDonald’s

• The US Postal Service

• Bloomberg

• Sephora

2. Marketing Is Likely Best Suited to Lead the AI Transformation.

The short answer is that as the function charged with growing the company, Marketing is missing the boat if it only focuses on the efficiencies available with AI. Product is one of the four P’s of “Capital M” (my phrase) Marketing, and we should always be at the table, if not leading, when revenue strategy is being discussed.2. AI Adoption Must Begin and End With Business Strategy In Mind. If you want to debate this idea, please join the community on this Substack discussion I started yesterday.

Share Let's Get Real

3. AI Adoption Must Begin and End With Business Strategy In Mind.

Nearly every presentation exhorted us all to stay focused on the business strategy and the potential business value of using AI, NOT on adopting tools just because they were fun to use. I came away with a solid framework for helping my clients begin the process of adapting to the AI future in their marketing departments.As an MBA, this is so obvious, I can’t believe it needs to be said, but apparently it does: tools are not the solution to all business problems. It’s critical to clearly define the problem and develop a a strategy for executing it.

This is where I insert my shameless plug for my own services as a marketing technology and analytics veteran. Hire me to help your company figure out what the business cases are, and create a plan for thoughtful adoption. Better yet, hire me to help figure out how AI can make you a strategic leader against your competition. Let’s go to the grand and glorious future together.

4. The Good News: You Are Not Too Late: Even Free Generative AI is Nowhere Near Ubiquitous As You Might

I saw a lot of folks at MAICON who were small agencies, and not as many big brands, or big agencies. I talked to a number of people who went out of sheer curiosity. Many, many people are at stages in their careers where they’re ready for something new. (Confirmation bias on my part is a real thing here, because that’s me.) But You’re not too late if you’re just getting into this, you are still right on time. But that advantage is shrinking rapidly…

The world isn’t yet ready to dive into AI like I, and the people reading this blog probably are. They’re still well behind, and overwhelmed and curious and trying to make sense of it all. I hope my blog will help, just a little bit.

Nevertheless MAICON attendance was notably lacking in West Coast presence. I don’t know if it was timing, the long haul to Cleveland, or coastal snobbery. My theory is that the Valley is focused on the tech itself and is less vested in how it’s used or not ready to really think about how it’s being marketed….yet.

But the people that were there were so passionate about the potential for AI in marketing, and it was so fun to be in a time where we could debate what is coming and hear from people with varied perspectives: sales, strategy, ethics, finance, and technology. There wasn’t as much hype and hustle. It felt so much more holistic and dare I say, HUMAN, than most tech conferences I have ever attended. We need more Midwestern Nice in the world.

5. This Is Gen X’s Moment.

As I have been saying since April, I firmly believe that AI in Marketing is made for Generation X. The crowd was almost entirely Gen X, with some Boomers and Millennials and a smattering of Gen Z in the crowd. Maybe because we’ve see how much money can be made in a gold rush, and we smell it here? Loving this piece from a couple years back about what Gen X Brings to the table from a leadership perspective. We are independent, fearless, and very, very experienced at living through technology transformations. We have also gained life wisdom, and we still have the energy to lead the types of transformations ahead of us. AI is hitting at exactly the space between the retirement of the Boomers, and the ascendency of Millennials in management. To use a Gen X phrase, I’m stoked.

6. MAICON: More Education Than News

The number of vendors and audience members were consumers of AI, whether personally or professionally. I did not see enterprise software or large enterprises represented on stage or in the vendor booths. I don’t see this as a conference you go to to learn the about latest release of ChatGPT, but a place where junior staff and newcomers go to learn about how these tools are being used.

This is not a bad thing, because it means that Paul Roetzer, Cathy McPhillips, Mike Kaput, and the MAICON crew have their finger on the pulse of what their audience needs. Manage your expectations accordingly.

7. I haven’t Been Thinking Enough About Voice.

One of the highlights for me was a fantastic preso by married team Susan and Scot Westwater from Practical Digital on the use of Voice and Conversational AI in business strategy. They’re working with some of the world’s largest companies, and have been for quite some time. They gave an in depth discussion, and I want to spend more time thinking about that opportunity for my clients. Tidbit: GA still tracks voice traffic as Direct. Gah, my nemesis strikes again!

8. 30+ Tools I Can’t Wait To Try

Mike Kaput of MAICON is a human wikipedia of AI tools and he very generously shared a presentation called 45 Tools In 45 minutes. I love tools and I could think about them all day, but there are just too many in the world; so I’m grateful for his thoughtful curation. Some of these may be MAICON sponsors, so take that into account.

Perplexity, Quantified, Alli.ai, Vista, Pictori,ai, content@scale, Rasa.io, Hyperwrite, Lately, Pecan.ai…to name just a few! I’ll write about them in the weeks to come.

Want A Video Of Some of These Thoughts?

I met the fascinating Geoff Livingston at MAICON and he invited me to do a MAICON download with his pal Greg Verdino on the “No Brainer Podcast” which is all about Marketing AI. Check it out! It was my first video appearance since being diagnosed with cancer a little over a year ago.

Celebrating Humanity: AI For Good

As a breast cancer patient, soon to be a survivor as I finish treatment in 2 weeks, this is some amazing news. My cancer was the kind that can’t be felt or seen, and it was diagnosed by a radiologist. We can’t just be afraid of AI, we have to seek out and develop uses that will be good for humanity.

AI Use In Breast Cancer screening as good as two radiologists, (preliminary) study finds.

Playing With Generative AI Images

Substack’s generative AI didn’t like this prompt, but Bing AI (Dall-E) gave me it’s best shot. We’re still in early innings, folks.

”Dog Days Of Summer”

What’s Next

• I’m currently taking a course in “conversation design”

• Testing tools from MAICON

• Planning Season 5 of Stayin’ Alive in Tech, focused on AI

• Planning my travel for this fall. What conferences are a must do?

• Please keep an eye on my “notes” in Substack for quick tips, links, and more. that’s what will help me keep these blogs from being so huge.

• What do you want me to drill into? what’s interesting to YOU? Let me know!
  

  Leave a comment

At MAICON (Marketing AI Conference) last week in Cleveland, a provocative claim was made: that marketing was best suited to be the driver of AI in the enterprise. The rationale is one I want to believe; but I’m curious what you all think!

Here’s the basic argument: Marketing Owns -or should own- AI for two key reasons:

1. We are the first to be truly disrupted by it, so we’re going to learn things we can teach to the rest of the organization
2. We are able to communicate about that experience better than any other function.

So what say you? Is Marketing the driver of AI in the Enterprise? Or only for certain enterprises? If so, which ones? And if it’s not Marketing, who owns it?

This is my personal journal about my learning journey in AI and ML for marketing. With over twenty years of experience managing and marketing tech companies, I have deep knowledge of marketing analytics and attribution, always focused on generating measurable and meaningful business results. For the last nine years, I've advised and served marketing leaders in tech companies, enabling me to share my insights on the impacts of AI and ML, including privacy and ethics issues, through this Substack. You can learn more about me, including my policy on using AI for this blog, in the "About" section of this Substack. I welcome your suggestions for further education on AI and ML.

In This Issue:

• Brent Britton, esq. talks to us about AI and how to think about it as marketers.

• Playing with AI Generated Images

• Celebrating Humanity

• What’s Next

• A Question for you
  

Brent Britton, esq: Attorney. Engineer. IP Expert. Pianist. Badass.

Brent Britton is Fiddlehead’s attorney, but the reason you should listen to him is because he’s also (as far as he knows) the only attorney to graduate from the MIT Media Lab. He’s graciously agreed to answer some questions for me - and you- “off the clock.” There’s so much speculation from non-lawyers out there, that I am thrilled he decided to go on the record and give us all some good, professional advice.

Brent gave me no consideration for this post, it’s simply a gift he’s giving to you as part of my community. If you need a good attorney, you should hire him and the team at CoreX Legal. Fiddlehead is on the smaller end of their practice, but we have been very satisfied with their patient, wise, and practical counsel; plus I guarantee you that your attorney’s photo is nowhere as cool.

Brent, Tell Us About Yourself.

I have, uh, brown hair, blue eyes. I enjoy surfing, backgammon and men who aren't afraid to cry.

But seriously folks, I’m just this guy, you know?

Ok, Fine. I’ll be specific: What is your area of legal expertise, and where are you in your understanding of AI and the law today, especially as it pertains to marketing content (text and images)?

A million years ago, before I went to law school, I was an engineer. I studied AI as an undergrad, and also as a graduate student at the MIT Media Lab under Marvin Minsky, basically one of the gods of AI. True conversational AI has been “ten years away” for 50 years.

Alan Turing proposed a test in the 1950’s: if you are communicating with an AI and you cannot tell the difference between the AI and a human, the AI passes the test. We did not credit AI with passing the Turing test when it beat a grandmaster at chess, or when it beat humans on Jeopardy. But last November when ChatGPT went public, most people would say AI has now passed the Turing test.

The legal climate at present is in a tizzy. AI can only learn by reading and analyzing existing stuff. But if you review my stuff and produce substantially similar stuff, that constitutes a copyright infringement. So it is likely that the deployment of AI constitutes a copyright infringement of massive proportions.

Before we go on, it’s important to note that you are not offering specific legal advice to me, to my company, or to anyone else who reads this piece, right?

That’s right. To hell with you freeloaders. If you want meaningful legal advice from me, you’ll have to pay for it! And don’t go trying to offer any personal favors if you get my drift because I’m a very happily married man. You won’t catch me doing *that* again. (editor note: now you see why I like him so much. I have no idea what favors he’s talking about, but I wouldn’t be surprised if it happened.)

How similar or dissimilar is the AI climate to the time when Napster roamed free?

In the 1990s, there was no AI in common use. Napster had nothing to do with AI. AI was not a real thing yet – it was a laboratory thing.

What is the risk to brands and agencies if creators use generative AI without disclosure?

If an AI is composing your copy, no one owns it.

Ownership of words is governed by copyright. Copyright authors must be human. If your AI writes it; then you don’t own it, the AI doesn’t own it, and it is in the public domain. This means that no one owns it. Anyone can copy it and repurpose it.

 *But* if the copy produced by the AI is substantially similar to any pre-existing works that the AI had access to before it composed your work, then *you* may be liable for copyright infringement. Not the AI – AI’s are not people and cannot be sued. Just you, the user of the AI that produced the substantially similar, infringing content.

That’s Terrifying! Is generating a first draft then editing it enough to gain copyright protection? Practically speaking, how much should we be changing the copy that AI spits out, in order to protect our companies from copyright claims?

So that it is mostly unrecognizable from the copy. The result should not be substantially similar to the work that was copied.

(Editor's note: I pressed him on this several times. He didn’t budge. He’s not kidding.)

Should brands and agencies be monitoring AI usage at work? What are the key questions to consider?

Yes, and the rule is: DO NOT PUT AI-GENERATED CONTENT INTO OUR WORK PRODUCT! And, every company should have an AI policy.

How can brands protect themselves from this risk when hiring agencies, contractors, consultants, and employees?

Reps and warranties and indemnifications. You’re looking for language that sounds like this: “Writer represents and warrants that all deliverables are the original works of Writer and that Writer has the absolute right and license to deliver the same to (me). Writer will indemnify (me) against any claims arising from a breach of this warranty.”

(editor’s note: check your Master Service Agreements and Professional Service Agreements. Get addendums if you must. It’s very likely agencies are already using generative AI in your deliverables if you are not specifically asking otherwise. )

If employees, contractors, consultants, or agencies use Chat GPT-is this a potential violation of their NDA?

Everything they type into ChatGPT prompts is a breach of any NDA they’ve signed- if what they type into the prompt contains confidential information governed by the NDA. (Editor’s note: this includes spell checking documents that contain confidential information. Your internal policies and training should cover the use of AI.)

If you and a competitor are both using AI, what kind of protection and/or liability do you have if accused of copyright theft?

There is no “everyone’s doing it” defense to copyright infringement claims. If you are using a tool that has copied other people’s stuff and is producing infringing copies, you’re liable.

How should multinational companies, or those regulated outside the USA, manage their use of generative AI?

They must tune their practices for each jurisdiction.

Ok, I’m sufficiently freaked out. How can marketers and agencies use AI in an ethical and legal way?

Use it! Chat with it! Explore it. Integrate its results into your work. Just have a policy for its use, do NOT let it be the sole author of your work, and do NOT give confidential materials to the AI.

Anything else you want to add?

Ok look, my answers above may seem draconian, but the truth is we are still figuring all of this out legally. There are numerous cases pending and we’re not sure which way they will turn out. Many of the AIs available today were trained on data – software, photographs, etc. – that are protected by copyright and subject to licensing terms. If the owners of those data can see substantially similar copies in AI output, then they will be likely to keep suing. At present, they are suing the AI publishers themselves – OpenAI, etc. But they might decide to sue you, the user, just as easily.

So here is the practical advice:

1. Be careful. Use AI to produce stuff, but go over it with a critical eye. Try to make sure the AI output you’re using isn’t too similar to any input that may have been used to generate it. Not easy; I get it. But if you see a Getty Images watermark on the AI output, that’s a pretty good tell.

2. Don’t forget that only humans can be authors and inventors, so you do not own the AI output. No one does. It’s all in the public domain.

3. Remember that everything you type into the AI prompt is no longer confidential and may be fed into the AI for future use. For goodness sakes, read the terms of service that comes along with your AI of choice and ensure you understand what it says about ownership and confidentiality and especially indemnification because it probably says that if the AI publisher gets sued because of output you generated using AI then you have to defend that case. Fun times. (Editor’s note: For him, not for you.)

4. If you are using AI to generate software, be very afraid. One of the earliest lawsuits was filed over massive software copying; it is still pending.

5. Finally, remember that the AI is full of sh*t and will lie to you all day. It will make up facts and then insist that it is telling the truth. It will invent fake bibliographical citations that look very real but will hold you up to ridicule if you publish them without fact-checking. This dynamic is lovingly called “hallucinating” in the parlance, as though the AI were just temporarily tripping or something. Ha ha… so cute.

TL; DR: AI is a great tool that will make your work better, stronger, faster, and more efficient. But you must review and scrutinize every single thing it produces before putting that output into production because it could be wrong, or flat-out made up. It could be an infringing copy of someone else’s work. It could subject you to liability. Let’s be careful out there.

Editor’s Note: want to hear more from Brent? He did a webinar this week. Check it out.

Celebrating Humanity:

As you know, I have been undergoing breast cancer treatment for the past year. As I write this, I am celebrating the one-year anniversary of the first phase of my treatment, which was surgery to remove the tumors in one breast, as well as several lymph nodes. Thankfully, the lymph nodes turned out to be negative. I’m considered NED (No Evidence of Disease), and if the statistics apply to me, I will be around for a normal lifespan.

Now, if you had met me prior to my cancer diagnosis, you would know I had "plenty to spare" in that department, and frankly, the idea of a mastectomy was appealing: who wouldn’t want to skip wearing a bra again? And if it ensured that I was done with cancer once and for all, it seemed like a fair trade. However, Dr. Susan Love (via her book) did the first research that uncovered this: mastectomies (in most cases, but especially in my case) have no better outcomes than a lumpectomy.

Prior to Dr. Love’s work, women would go in for a lumpectomy and come out with a double mastectomy without consent. Her book is the bible for most of us with breast cancer. Her untimely death on July 2 from recurrent leukemia was a loss to all of us in the breast cancer community. She was a pioneer in surgery, same-sex marriage, and just being a woman in a very male-dominated environment. She went on to create a foundation and an army (literally, we're all part of Dr. Love's army of patients who volunteer for clinical research).

You owe it to yourself, to me, and to anyone you care about who’s had breast cancer, to read her obituary. I’ve gifted a link for you from the Washington Post.

Playing With Midjourney Bing and Substack Images

A simple word, summer. And I’m so grateful this year to be able to enjoy it.

What’s Next

• I’m off to MAICON: Marketing AI Conference Next week. I’m sure I’ll be fired up with lots of ideas upon my return. Keep an eye on my Notes here, as I will be posting interesting tidbits. I will also post business things on LinkedIn. I’m on Threads and Bluesky, but neither of them feel like a good use of time right now.

• Please continue to share with your friends. It keeps me inspired and excited to find out new things.

• Season 4 of Stayin’ Alive in Tech is going to be all about AI. Who do you want to hear from?
  

Share Let's Get Real

Leave a comment

This is my personal journal about my learning journey in AI and ML for marketing. With over twenty years of experience managing and marketing tech companies, I have deep knowledge of marketing analytics and attribution, always focused on generating measurable and meaningful business results. For the last nine years, I've advised and served marketing leaders in tech companies, enabling me to share my insights on the impacts of AI and ML, including privacy and ethics issues, through this Substack. You can learn more about me, including my policy on using AI for this blog, in the "About" section of this Substack. I welcome your suggestions for further education on AI and ML.

In This Issue:

• Google (Bard) Nails It (at least for me, and at least right now)

• Playing With Midjourney Substack and Bing Images

• Celebrating Humanity

• What’s Next
  

Google(Bard) Nails It, At Least Right Now

Today, Google had its IO conference and announced a lot of stuff regarding AI. Everyone and their brother is writing furiously about it. And, as luck would have it, I’m at Mom’s right now for the first time since cancer, and I’m trying to stay present with her.

But I just had to write about this jump in capability, because it is solving a problem I’ve had for a few weeks now. As far as I know, no other chat bot can do what I’m about to show you, at least not from a graphic user interface. (As always, correct me if I’m wrong). I’m jamming this article out fast to get back to Mom, so please excuse any grammatical issues. This is how you know I write these articles, I guess! :D

My goal: get a chatbot to read a CURRENT article and summarize it for me. Sounds easy enough, but I spent literally hours with ChatGPT-4 trying to get it to do what I’m about to show you that Bard can do right now. Because ChatGPT-4 stops at 2021, it’s not great at taking new information in. You also can’t upload big files or do it in any predictable way.

Here’s what happened to me just a few weeks ago with Chat GPT 4. My goal was to get it to write a LinkedIn post that summarized the Forbes 50 Entrepreneurial Marketers article and called out some themes that were present in the article; some commonalities if you will.

ChatGPT4 did an admirable job of summarizing the document, and with a LOT of prompting actually got to a place where it asked me questions about what I wanted to actually say in the article. I was pretty excited. But when I asked it to use specific examples from the article, it listed marketers who weren’t even in the article!

I only discovered this when I went to post the article to LinkedIn and then realized I couldn’t tag the people listed because they didn’t hold the titles that ChatGPT said they did; and in some cases, they weren't even in the article itself!! ChatGPT kept putting in Fernando Marchado’s name. And while Fenando Merchado is a wonderful and entrepreneurial marketer, he simply isn’t on this particular list. I tried putting the article in a google doc to upload it. I tried copying and pasting the text (it was too long). Truly, I worked on this for hours, wondering where I had gone wrong.

As a result, I had to scrap the entire post, for fear of making some other major blunder.

Meanwhile, literally just now, while on my phone (on LTE at Mom’s, where I’m actually tethering to write this piece, please send sympathy,) I was able to get Google Bard to do an absolutely magnificent job of it in just 3 prompts. I could have done it in one long post, but it stuttered for a second and needed to be led along.

ChatGPT 4’s Output:

Google/Bard’s Output, Tonight: Light Years Ahead

A Few Thoughts about Google/Bard, and What’s Next:

1. Surprisingly, Bard was happier being directed to a website link than to my Google Doc, but I bet that will get fixed pretty quickly.

2. I did a bad prompt and it started writing posts about being a better Linkedin marketer. But I closed the app and started again.

3. The speed and quality of this tonight got me thinking: Google’s been crawling the internet for a hell of a lot longer than OpenAI. There are some things Google is going to naturally do better than anyone else for that simple reason alone. I’m not saying they’ll be the winner-anyone who suggests they know who that winner is full of it-but that all the hype around OpenAI doesn’t mean they win.

4. I continue to be so excited about the possibilities. No one can be counted out right now.

5. As always, the usual caveats apply: be smart and check your work, and don’t upload sensitive information to a chatbot until you understand the implications.

Celebrating Humanity: My Family

This tiny hoe was made by my great-grandfather, who ground down a full-size hoe for my great-grandmother. to make it easier for her to weed in between the plants in her flower beds without bending over so much. It’s been handed down through the generations. My mother is an expert gardener and has it sharpened every year. It absolutely murders dandelions. And someday, it will be mine.

Playing With Midjourney Bing and Substack Images

Substack’s image generator is now working pretty well so long as I give it a simple prompt. It’s close enough. Here I’m celebrating the beautiful cardinals that show up at Mom’s house. Their song is as identifiable as their bright red plumage.

What’s Next

• Still on my trip, but as the Muse strikes I’ll be here.

• Please keep an eye on my “notes” in Substack for quick tips, links, and more. that’s what will help me keep these blogs from being so huge.

• What do you want me to drill into? what’s interesting to YOU? Let me know!
  

Share Let's Get Real

Leave a comment

This is my personal journal about my learning journey in AI and ML for marketing. With over twenty years of experience managing and marketing tech companies, I have deep knowledge of marketing analytics and attribution, always focused on generating measurable and meaningful business results. For the last nine years, I've advised and served marketing leaders in tech companies, enabling me to share my insights on the impacts of AI and ML, including privacy and ethics issues, through this Substack. You can learn more about me, including my policy on using AI for this blog, in the "About" section of this Substack. I welcome your suggestions for further education on AI and ML.

In This Issue:

• Practice Makes Better (Prompting)

• Playing With Midjourney Substack and Bing Images

• Celebrating Humanity

• What’s Next
  

Practice Makes Better (Prompting)

Note: If you’re bored with all the “prompt tips” you see everywhere, I get it. I am too. This is written for me and my followers, who tell me they want to hear from me specifically on this topic.

If you’re like most people, after you got your OpenAI account, you stared at that open chat box and thought “What now?”

You are not alone. The blank page is a known problem, even among professional writers. It was a key barrier to the widespread adoption of Second Life. It’s that horrible feeling when someone asks you a question and you just…freeze.

It’s ok. I’ve got you. Fear not, and read on.

If you’re already playing around and want to jump to the next level with me, you can skip this next section.

The Absolute Noob’s ChatGPT Icebreaker List

Here are some sample prompts that you can use to get started. These are all ones that I created myself as I was trying to learn how ChatGPT works. (A gentle reminder that ChatGPT is a brand name for a generative language model. It’s like Kleenex for “facial tissue.”) Here are three sample prompts, in increasing order of complexity.

• Who is Melinda Byerley? I wrote about this journey earlier, here. Obviously, the point is to ask it who YOU are and see what it says.

• How do I cook a 6.795 pound boneless leg of lamb so that it is tender and juicy while being cooked no more than medium rare? My prompt size is starting to expand, and I’m still at that point thinking about ChatGPT as a “Google Synthesizer.”

• Please give me a weekly meal plan with X calories per day, with no more than y from fat, and at least x% protein. Limit red meat to twice per week, and limit dairy except for butter, yogurt, and cheese. Include at least 7-10 fruits and vegetables per day. Now I’m getting “crazy’ and thinking about how much can I get ChatGPT to tell me that would take me hours to research on my own.

My Next Level Of Prompting

I love that ChatGPT can save my prompts. This is one reason to have a paid account, beyond just accessing the more advanced GPT4 model. This allows me to keep a series of prompts, as the model learns with me, and can build upon itself over time.

Example 1: Becoming a better bridge player

In 2021, I set a goal to learn how to play bridge. I was an absolute newcomer to the game, though I had played a lot of cards in my life, and was a competitive chess player in high school. (Varsity Letter in Chess = Nerd Badge Unlocked)

I’ve been taking live classes, in person, with a break during cancer treatment. As I was recovering, I got back into it, and I’m spending pretty much every spare minute I have that isn’t on AI, and recovering from cancer, playing bridge. But I felt stuck in some key areas, and as I am doing with most questions in my life these days, I decided to ask ChatGPT. Here’s the first prompt.

you are a world expert in duplicate bridge who loves to help beginners learn and grow. I am your student. I have taken a basic class with a good instructor. I know the stayman and Jacoby conventions. Help me understand how to play better duplicate bridge at my level so I can get more masterpoints.

I left two spaces in between sentences because some habits die hard. Nevertheless…

I got this output. We all recognize it. It’s the “Oh wow, how fast,” followed by “Meh, it’s pretty bland and unactionable.” Stay with me, and at least skim it, because I’m going to show you how I iterated on it to take things to the next level.

Here are some things to notice in the response that you can consider in your own prompts.

1. Conversational and Friendly Tone: I like teachers that reinforce me as they help me from getting discouraged when things go wrong. And in Bridge, things often go very, very wrong.

2. Ideas For More Research. There’s a difference between knowing what to do and doing it. Look at the last part of #2. I knew what 2 of the things listed were: negative doubles and takeout doubles. But I didn’t know what cue bids were. That was a suggestion for me to go research.

3. Ok, Now What? Mostly, this was a list of pretty basic and generic stuff you’d find on any “how to play better bridge” website. Think about the questions you’d want to “drill into” and that’s how you’ll iterate your prompts to a better place. I have a Chrome plugin called “Keywords Everywhere” from my SEO days that offers quick ways to ask ChatGPT to give me more. Here’s a screenshot of what it looks like:

   

   
   

   I asked GPT4 to expand, on that list, and got a literal pile of deeper suggestions to go research. Here’s just the first part of the list. See all the detail in there?
   

4. Brainstorming in “Flow:” Leads To Objectives: As I started to think about what parts of my Bridge game needed help, I started thinking about how scoring in duplicate (competitive) bridge was still so mystifying for me. In particular, it bothered me that I could play a hand well, but not get matchpoints. So I asked ChatGPT this:

   I am struggling to understand how scoring in duplicate bridge works--especially how that form of scoring should impact how we play. 

Now look at what comes out, and how helpful it is. When this came out, I was jumping up and down. We had never been taught this in class. It’s amazing. Bridge can be VERY opaque to noobs, as much of the learning is locked up in very old books that take a long time to study; as well as the minds of the folks playing for decades who aren’t always great at explaining it.

5. And So on, And So on. I asked GPT to “exemplify” with the tool above. And I got an example of a specific bridge hand. Et Voila, I’m now better at bridge in this area. I scored more matchpoints in that week’s hand than I had in several weeks.

6. My Trusted Bridge Advisor. This chat is saved in my account, and as I run into questions, I work with the tool and it expands and grows over time.

Example 2: Strategic Prompting = Efficiency

First, an ethical disclosure. I put into practice the advice I received in this blog post. You should check out his YouTube and newsletter. I sincerely loathe charlatans who are taking other people’s content and presenting it as their own without attribution. For the record, I have written the copy here myself. No ChatGPT. It’s mine. Don’t steal it. :D I like to think the value I provide here is the time it took to sift through all their content and find the practical stuff that worked. There is a lot of crap out there masquerading as prompting advice.

If you read the article I just linked to, he provides a lovely framework for thinking about prompts.

• Context

• Clear Objectives

• Specificity

• Iteration

I stumbled into some of this framework with the bridge example: I told it the context around my needs. But I wasn’t clear about my objective, nor was I specific enough to get what I really wanted. I did, however, have the tenacity to keep iterating, and I eventually got there.

The point of this framework is to get us all there faster, with less iteration from the start. Less brute force, if you will, and more elegant and time-saving prompts. I took from his post—with his suggestion—a prompt, and I applied it to the question I have been answering for nearly all of my professional career. Here it is:

Generate a 300-word article for a strategic marketing agency that provides a comprehensive guide to proving the roi of marketing in the c suite. The article should cover various strategies and techniques, and be written in a clear and concise style accessible senior marketers, whether they have an MBA or not. To support your advice, utilize research and expert opinions from credible sources like academic journals, strategic consulting firms, and the advice of expert marketers. Incorporate relevant statistics, examples, and case studies to illustrate your points and bring the advice to life. Consider the potential challenges and obstacles in getting marketing to be respected and appreciated by the CEO, cto and CFO for the value it delivers and offer strategies to overcome them. 

It did a very fine job of organizing information I know very well, but I would never copy and paste this, because I would want to check whether the sources listed are accurate.

Then I simply kept asking GPT4 to drill down and drill down. I got what I wanted in 4 prompts instead of a dozen or more that it took me on the bridge example.

Yes, I could have ultimately written the post myself. But as a person with lifelong ADHD, it might have taken me days, and a LOT of wasted words. I still need to check the final output and add more substance and practical examples, but the executive function is for me akin to wearing eyeglasses: a necessary aid to my performance.

Pro Tip: copy.ai gives a daily suggested prompt. I’ve been using them, and it’s a good reminder to try something new each day.

Celebrating Humanity, with Horses

I came across this story I’d missed last year, about how humans are making farms for racehorses to live out their days in peace and comfort. It sure beats being sent to a glue factory. Pay attention to the lyrical writing quality, which is something that ChatGPT just cannot emulate right now. Think about how you trust this story and the data it provides, simply because you know it’s The Washington Post.

Playing With Midjourney Bing and Substack Images

This time I was unable to get Substack’s image generator to give me the picture above based on a simple word: Horse.

To the theme of this post, here’s what it gives me with a more robust prompt. Hey, we’ve been watching “Yellowstone” so that’s where my head is at.

Here’s what Bing’s Image generator gave me for “horse.”

And here’s what it gave me for “Wild horse in Montana majestic and free”

What’s Next

• I’ll be traveling for the next couple of weeks. I’m attending the Build a Better Agency Summit, and most importantly, I’m seeing my mom for the first time since my diagnosis last year.

• Please keep an eye on my “notes” in Substack for quick tips, links, and more. that’s what will help me keep these blogs from being so huge.

• What do you want me to drill into? what’s interesting to YOU? Let me know!
  

Share Let's Get Real

Leave a comment

This is my personal journal about my learning journey in AI and ML for marketing. With over twenty years of experience managing and marketing tech companies, I possess deep knowledge of marketing analytics and attribution, always focused on generating measurable and meaningful business results. For the last nine years, I've advised and served marketing leaders in tech companies, enabling me to share my insights on the impacts of AI and ML, including privacy and ethics issues, through this Substack. You can learn more about me, including my policy on the use of AI for this blog, in the "About" section of this Substack. I welcome your suggestions for further education on AI and ML.

In This Issue:

• What I Know, So Far

• Playing With Midjourney Bing Images

• Celebrating Humanity
  

What I Know, So Far: Month Two

Since recovering from Covid, I’ve been immersed in the world of AI for Marketing. I’ve been attending online sessions, reading(finding links from Reddit, LinkedIn, Twitter, and all kinds of sources), listening to podcasts, and exploring prompts and tools.

Almost daily, there’s news out there that in a prior world, would be earth-shattering for at least a week. Because I write to know what I think I’m taking time in this post to summarize what I know so far. This list is written for someone who asks me what they need to know, right now.

As this is a lengthy post including a discussion about privacy/ethics, and contains plenty of bookmarks, I’m skipping that for this issue.

It’s Real. It’s Here.

I empathize with the horror of entering a simple prompt into ChatGPT and seeing it spit out a perfectly written post in seconds. Some writer friends are terrified and angry about these tools. Some marketing friends are frozen in fear, overwhelmed by information.

Others wonder if it's just hype. It's not. I've made a career out of debunking hype cycles. I advise against installing new tools to solve people and process problems. I rarely sound the alarm about new technology. But the last time I saw technology like this was when I saw the Apple iPhone. I dragged people into the Apple Store and said, 'Look at this. I don't know how, but it's going to change everything.'

Consider yourself dragged into the 'AI and ML store.'

Change Is Coming Within Months, Not Years.

There is a strong possibility that many people will lose their jobs because of AI, and the job losses are going to affect knowledge workers faster than any of us thought possible. As a society, we must start to have the conversations needed to manage the impacts. But in the very near future, the marketer who uses AI will be more employable than the marketer who does not. It’s time to get to work to make sure you have a strong base of understanding.

Suggested actions:

1. Get a ChatGPT account and pay for GPT-4. Bonus: Get a Midjourney account.

2. Try prompt cheat sheets. Here’s one I like.

3. Check out the Marking AI Institute site. I’m also planning on attending their conference in Cleveland in late July. Let me know if you are planning to attend. Their podcast is also quite useful but goes a level deeper than this substack, so be prepared for some information overload.

4. Do it a little bit each day or every other day.

5. Use it in your personal life. (I used ChatGPT, Bard, and Bing to research leg of lamb recipes this weekend.)"

A brief dictionary of terms for marketers, because the language is evolving in real-time.

As of April 2023, here’s my best understanding of the terms I’m using.

• AI (Artificial Intelligence, aka “narrow AI” and in some cases “generative AI”): AI that is used for specific and narrow purposes. Prior to ChatGPT’s general release in November 2022, narrow AI was in use in many ways, from generating search results in Google to developing cancer treatments. In marketing right now, AI is a buzzword that stands in for a very specific kind of AI called “Generative AI,” which is based on Large Language models (LLMs.). This is where tools like ChatGPT, Copy.Ai, Midjourney, and the literally thousands of companies started up right now sit.

• AGI (Artificial General Intelligence aka “strong AI”): this is what most of us would have described as AI a few months ago. It’s the stuff of science fiction, but it is real. It’s just much farther away, by most estimates about a decade. Some experts think it will never come. The experts don’t even agree on how to define it. Therefore, AGI is not on the table for marketers in the foreseeable future so it’s beyond the scope of this Substack.

AI Today Will Bullshit You. But It’s Not Sentient. Proceed with Caution.

Yes, we all shivered a little when Bing told NY Times reporter Kevin Roose to leave his wife. At the same time, it’s important to remember he was speaking with a computer program, not a primitive human being. The only thing Bing/ChatGPT/Bard can do is summarize what it has been trained on. As I explained above, all things called “AI” right now are “narrow AI.”

In this case, Bing is based on GPT, and it includes vast swaths of the internet, with all its misinformation, hate, and spam. It can synthesize vast amounts of information quickly, but it will also make up things when it doesn't know what to say. The experts call it “hallucinating,” but it’s really more like that person who bullshits when they don’t know the answer instead of saying they just don’t know. Take its advice to leave your wife with a grain of salt and always fact-check.

Incidentally, this is why generative images on Midjourney have weird feet and hands because the models are trained on existing internet images, which are heavily skewed to faces.

Don’t Enter Sensitive Information Into AI Tools Without A Lawyer’s Advice. Or Why “On-Prem” AI doesn’t exist in marketing yet.

Recently I was on a webinar where several organizations were using ChatGPT to provide various marketing tasks.

In at least two cases, the companies were unaware of the risk of entering sensitive data into ChatGPT tools and others like it. This is NOT like Google Docs or other cloud services. If you are using the public chat function, or scripting through a browser, everything you provide is being used to train the model. Samsung found out the hard way. Note that many “AI companies” are simply putting a UX “skin” on top of ChatGPT’s technology. Ask questions and call a lawyer before inputting sensitive data and information.

You’d think we’d have all learned our lessons after Facebook, but apparently not.

Yes, if you sign up for the API, Bing/OpenAI is now opting you out of collecting your data by default, but they still retain some of that data for a period of time. To get a truly private version of the Chat GPT model, you have to apply and be granted that private version. I have yet to find an instance of “on-prem” AI for marketing tasks for anyone but companies with lots of engineers to manage the issues —please correct me if I am wrong on this!

Marketing Analytics AI: Not Yet Ready for Prime Time.

As I’m a specialist in analytics and marketing attribution, I’m paying close attention to evolution in this space. I believe the privacy issue will be a major barrier to new entrants, and until this is figured out, I remain skeptical of marketing analytics AI. There has been so much hype in this space over the years like that seen around CDPs, that has come to nothing.

Marketing analytics problems are almost always people problems. The biggest challenge we have in marketing analytics is getting people to take action on the data.

As a result, I will be watching for tools that leverage cohorts of user data and simplify the interaction with the data via chat tools; vs trying to sell solutions for personalized marketing. If you are a company working on using AI in marketing analytics, I want to talk with you and see your product. My clients need you. This conversation can be on or off the record.

Immediate Opportunities Exist For Workflow Efficiency. It’s Going to Get Very Messy.

Today, your company can create massive efficiencies in workflows. I’ve got enough experience with ChatGPT4 to tell me that for many, many social media teams, it’s good enough, once you know a bit about how to prompt.

If a writer can generate 10x the social media /email/blog posts in 1/10th the time, and assuming the same number of content pieces are made (because honestly, the world doesn’t need MORE emails and social media, right? ), then fewer people are needed in the workforce.

Agencies: those that don’t adapt to this quickly, will go out of business. Right now many clients/marketers are like deer in a headlight: they’re not sure what way to go. That will change, quickly; especially if I’m doing this Substack right. 😈 Soon they’ll be asking agencies how they are using AI, and when they can stop paying $10,000 for something that is being done in 1/10th the time. And while yes, having humans review content and edit, and shape it, is important; “good enough” will be enough for many people.

In-house marketers: it’s time to start to think about your headcount and your agency use. Marketing is already the toughest job in tech; and it’s about to get tougher when your CEO /CFO asks which of your five writers is no longer necessary.

Practioners/Consultants/Employees: When that question the in-house marketer gets asked, how are you going to be prepared to answer? What bots are you training? How are you making your work more efficient?

Prompt Engineering Is Temporary

Speaking of prompt engineering, don’t get too enamored of it. The tech industry has a vested interest in having as many people using AI as possible, and to that end, they are going to invest a great deal in making it easier and easier to get what you need out of AI tools.

For Gen X, think about this like a command line prompt: necessary for a while, but abstracted away for most of us, thankfully.

Training AI For Your Company Is The Future

Right now, the narrower the AI is, the more effective it is, and thus the more ROI it provides. ChatGPT 4 is amazing, but tools that can tailor it for writers and marketers are even more efficient.

Now, imagine what could happen if you train ChatGPT on your website with thousands of pages, so that when it writes social media, it writes it in your style. Imagine it integrated with SalesForce Einstein and all your internal documents. (with the caveat I said earlier about the lack of “on-prem” AI for anyone but big companies just yet).

Bloomberg just announced that.

Companies that behave like Bloomberg will beat those that do not.

Domain Expertise Matters More Than Ever

As ageist as Silicon Valley is, we experts have the advantage this time. Every company is going to be seeking out domain experts to sniff out the bullshit and train the models.

For example, if you’re an expert writer, CMOs want to hire you to not only to know when it’s time to have a human write something but also to review the output of bots and writers that use them. They want to be able to trust you to train the tools on your company’s unique voice and style. CMOs trust your take on which writer tools to use because they know they don’t have your grammar and editing and proofreading skills.

This is Gen X’s Moment.

Not only is Gen X more likely to possess the deep domain knowledge to detect the sophisticated bullshit AI tools are putting out; we’re also still young enough to be able to learn quickly and adapt.

We have the benefits of age- more leadership experience, better discretion, judgment, an appreciation for the nuance and complexity of life, as well as a memory of life without computers. But we’re young enough to be excited about seeing the change this next phase will bring.

We’ve seen some things, but expect to live long enough to benefit from AI.

Of course, this is a vast generalization. Some boomers are not shying away. Some millennials have their heads in the sand.

But Sam Altman said that he thinks he’s not the right person to run OpenAI in the future. There’s a lot I have disagreed with Sam on, but I will give him credit for starting to grasp this. (Whether he does anything about it remains to be seen.)

But I stand by this statement: Gen X will lead us in this time. Our “middle child, latchkey ” experiences have prepared us perfectly to assess and lead through this important transition.

Did I miss anything? Let me know.

Celebrating Humanity: Holy Edition

As several major religions celebrated important holidays in the last week, here’s an example of one thing that connects them: taking care of the poor.

A church in Winston-Salem, NC, purchased nearly $3.3 million in outstanding medical debt in three counties for 3400 families. They held a debt-burning ceremony in the church and you should watch it.

They take donations. And they don’t have to be huge:

Playing With Midjourney Bing Images

Unfortunately, I was unable to get Substack’s image generate to work with the prompt I used in Midjourney. But, I got access to Bing’s image generator and fed it the exact prompt I used to start this blog. Here’s the output. It really took" “graffiti style” to heart and certainly seems less of “The Last of Us apocalypse” vibe than Midjourney, though now I feel my avatar looks like Dora the Explorer.

What’s Next

• I’m reading this book on Marketing AI and will write a report on it after I read it. I do recommend the AI marketing institute and its podcast. I’ve attended several of their sessions and listened to the podcast, and so far, they’re doing a great job of summarizing things, while balancing hype and risks.

• Do you know someone who would be a great guest on my podcast this year? I’m looking for people who know about the history of AI for marketing.

• If you work for a company that’s integrating AI into its marketing technology, I’d like to talk with you, on or off the record. Please send me an email or ping me on LinkedIn.
  

Share Let's Get Real

Leave a comment