Tuesday, September 15th, 2026

Somebody sent me a message last week that I've now gotten some version of about forty times this year.

"I've been using AI for months and everything it writes sounds like nobody."

That's the exact phrasing. Sounds like nobody. Not bad, not wrong, not broken. Just weightless. The kind of prose that could have come from any business in any industry in any year, which is a strange thing to complain about until you realize it's the whole problem. You didn't want generic output. You wanted your output, faster.

And then you blamed the model.

I want to make a case today that the model is almost never the thing holding you back, and that the fix has nothing to do with prompts, better tools, or whatever got launched this month. The fix is boring and structural and takes about a week. I'm going to give you the whole thing.

The reason I'm writing it now rather than in January is that we're in the stretch where operators actually change how they work. Everybody's back, Q4 is close enough to see, and the calendar is not yet a wall. If you're going to rebuild how you use these tools, this is the four week window where it's still possible.

Let's go.

There's a piece from IBM's team earlier this year that names it better than I'd been naming it. They call it the context gap: the distance between what a system can access and what it needs to understand in order to be useful.

Their framing is about enterprise deployments, but it maps perfectly onto a one person shop. The models perform beautifully in controlled conditions. Then you put them into a real business with real policies, real exceptions, real history, and they fall over, because they don't know any of it. They can't tell the difference between data that's technically correct and data that's usable in your actual situation. The Gartner number they cite is that through 2026, organizations will walk away from sixty percent of AI projects that weren't supported by data the AI could actually work with.

Sixty percent. Abandoned. Not because the intelligence wasn't there. Because nobody fed it the environment.

Now shrink that to your desk. When you open a chat window and type "write me a follow up email to a prospect," what have you actually given it? A verb and a noun. It doesn't know what you sell, what you charge, who this prospect is, what you already promised them on the call, what you refuse to do, how you talk, or what happened the last three times you sent an email like this.

So it does the only thing available. It writes the average of every follow up email on the internet.

Which is exactly what you asked for. You just didn't know that's what you asked for.

Here's where it gets interesting, and it's the part I think most operators are sleeping on.

There's a piece in California Management Review from March by Teresa Tung and Philippe Roussiere out of Accenture, and their argument is that in a world where data is abundant and models are open to everybody, the actual moat is tacit knowledge. Their definition is the good one: the reasoning patterns, informal heuristics, situational awareness and interpretive skill that experts build over years. The stuff that never shows up in a manual or a dashboard but determines how you read an ambiguous situation.

They're writing for companies losing it to retirement. The version that applies to you and me is smaller and more urgent. Your tacit knowledge isn't walking out the door. It's just never left your head, which means every tool you use is operating with none of it.

One of their examples stuck with me. A cosmetics company had regulatory decisions living inside a handful of specialists who understood ingredient families, edge case exceptions, and the contextual nuance regulators actually apply. They built a structure that encoded that judgment, not a checklist of rules but the reasoning behind the rules. Evaluations went from hundreds a month to over forty thousand. Expert workload dropped by roughly eighty percent. Anything the system couldn't read confidently got routed to a human.

The line I keep coming back to: they didn't codify a checklist, they built a living model of institutional judgment.

You have institutional judgment. You've got twenty years of it. It's just not written down anywhere, and so the most powerful tool anybody has ever handed you is running on the same information as a stranger.

The reframe that fixed this for me is embarrassingly simple.

You are not writing prompts. You are briefing a contractor.

Think about what happens when you actually hire a good freelancer. You don't send them a one line text and expect brilliance. You send them the brand guide, the last three things you shipped, the pricing, the client background, and a note about the two things you never do. Then they come back with something usable on the first pass.

Nobody would call that a prompt. It's a brief. And the reason your AI output sounds like nobody is that you've been sending one line texts to a contractor with an eidetic memory and no files.

So build the files.

This is what I run. Six documents. Plain text, nothing fancy, living in one folder. I load the relevant ones at the start of any real piece of work and the difference is not incremental.

File one, the Operating Brief. One page. What the business actually is, who it serves, what problem it solves, what it sells, and critically, what it refuses to do. That last section matters more than the rest combined, because refusal is where your judgment lives. "We don't take clients under a certain size." "We don't do one off projects." "We don't build anything we can't hand off." Write ten of those. Nobody ever writes the refusals down and they're the highest signal thing in the whole stack.

File two, the Voice File. Not adjectives. Nobody's voice file should say "professional yet approachable," that means nothing and produces nothing. Paste in three things you wrote that sounded exactly like you. Then paste in one thing that sounded wrong and write a sentence about why. Then list your actual rules. Mine include no em dashes, contractions always, short sentences, no corporate language, never open with a rhetorical question. Specific and checkable beats descriptive every time.

File three, the Customer File. Your segments, their real objections, and the words they actually use. Not your words for their problem. Theirs. If your clients say "I'm drowning" and you've been writing "capacity constraints," you've been translating your customers out of their own language for years, and now you're teaching a machine to do it at scale.

File four, the Offer File. Everything you sell, what it costs, what's included, what's explicitly not, the guarantee, the terms, the delivery timeline. This one takes twenty minutes and eliminates the single most common failure, which is output that confidently describes an offer you don't have.

File five, the Decision File. This is the one nobody builds and it's the most valuable. It's a running list of decisions you've made and the reasoning underneath. We said no to that partnership because the incentives pointed opposite directions. We moved off that platform because the export was hostile. We took a smaller client because the work would teach us something. Ten entries is enough to start. This is where your tacit knowledge stops being tacit.

File six, the Constraint File. Your real capacity, your tool stack, what you can't do, what you won't. Without this you'll get recommendations that assume a team you don't have and a budget you didn't approve.

Six files. Call it four hours total if you're honest and about nine if you keep rewriting the voice file, which you will.

Here's the part that makes this tractable rather than a project you'll abandon on Thursday.

You're not writing these from scratch. You're harvesting.

Most of what goes in the Customer File and the Decision File was said out loud, by you, on a call, in the last ninety days. I run Fathom on everything, and when I built my stack the first time I did it almost entirely by reading back through transcripts and lifting the sentences where I explained why we do something a certain way. I found language in there I'd forgotten I used. I also found three places where I'd told two different clients contradictory things, which was its own useful afternoon.

The Voice File comes out of your sent folder and your published work. Pull the three pieces people actually replied to.

The Offer File is already in your proposals. It's just scattered across nine of them with slightly different numbers, which is a problem you were going to have to deal with eventually anyway.

The Operating Brief is the only one you write cold, and it's one page.

Building the stack does nothing if you keep typing one line requests.

The pattern that works is load, then ask. Before any real piece of work, you paste in the relevant files. Operating Brief and Voice File for anything customer facing. Add the Customer File and Offer File for sales work. Add the Decision File when you want it to reason rather than write.

Then, and this is the part I'd underline, use it to critique before you use it to create.

My own rule hasn't changed in two years. I create, it critiques, I refine. I write the thing badly and fast in my own words, then I load the stack and ask it to tell me where I hedged, where I drifted off voice, and what objection a reader would raise that I didn't answer. I do that in Galaxy.ai because I can run the same brief across a few different models and the disagreements between them are usually where the real problem is.

The second it starts doing the deciding, I've handed over the exact muscle I was trying to build. The stack makes the critique sharper. It does not make the thinking optional.

The maintenance piece is small but it's where this dies. A stack that's nine months stale is worse than no stack, because you'll trust it. I keep one Make.com scenario that drops a task on the first Monday of the month to review two of the six files on a rotation. Ten minutes. That's the whole discipline.

Writing adjectives instead of artifacts. "Confident and warm" teaches nothing. Three real samples teach everything. If a file can't be checked against reality, it's decoration.

Building it and never loading it. The most common failure by a mile. The files sit in a folder and you keep typing one liners. Put them somewhere that takes one click.

Making it too long. Six files, a page or two each. If your Operating Brief is eleven pages, you've written a business plan, and you'll never load it, and neither will anything else.

Letting it write the Decision File. Your decisions and your reasoning, in your words, or the file is worthless. This is the one place where doing it the slow way is the entire point.

Here's how you know it worked, and it's a clean test.

Ask for something you'd normally write yourself. A follow up email, a proposal section, a post. Load the stack first.

If the output comes back and your reaction is "this is basically it, I'd change four words," the stack is live. If your reaction is "this is close but it's not how I'd say it," your Voice File is too abstract, go add samples. If your reaction is "this describes a business that isn't mine," your Operating Brief is missing the refusals.

The diagnostic points at the file. That's the thing people miss. When the output is wrong, you now know which document is thin, instead of vaguely concluding that AI isn't good at your industry.

I built the templates for this because I needed them and it seemed dumb to leave them in a private folder.

It's called The Context Stack. Six fill in templates, one for each file, with the prompts that pull the raw material out of your head, plus the harvest checklist for mining transcripts and sent mail, plus the load pattern and the critique prompt written out in full.

It's a working document. No cost, no sequence waiting behind it, nothing to sit through. Copy it and use it.

Get The Context Stack here

If you build it and the output still sounds like nobody, hit reply and tell me which file you think is thin. I read those. And if it turns out the problem isn't context but that the business itself has never been written down anywhere, which is more common than you'd think, that's a different conversation and it's the one Pinnacle Masters exists for.

Don't build all six. You won't.

Build the Voice File. Just that one. Open your sent folder, find three things you wrote that sounded like you, paste them into a document, and write five rules underneath about how you actually write.

Twenty minutes.

Then take the next thing you were going to ask AI to write, paste that file in first, and ask the same question you were going to ask anyway.

The difference will annoy you. It annoyed me. Two years of mediocre output and the fix was a document I could have written on a Tuesday afternoon in 2024.

That's usually how it goes. The leverage isn't in the tool. It's in the twenty minutes nobody wants to spend.

Go spend it.

One step, one day. Grace over guilt. — Dan Kaufman

Tools I use and trust. If you sign up through these links I may earn a small commission at no extra cost to you. I only recommend things I actually use.

Make.com, automate your workflows without writing code

Fathom.video, AI meeting notes that actually capture what matters

Galaxy.ai, all your AI tools in one place

Rize.io, time tracking that makes you honest about your hours

Beehiiv, where serious newsletter operators publish