I could judge my own writing in two seconds. I couldn't say why.

I could judge a text of mine in two seconds. Good, not good. I couldn't say why.

That isn't a problem while I'm the one writing. It becomes one the moment I try to delegate.

I asked a generic AI to write in my place. What came out was a perfectly ordinary LinkedIn post. Correct, fluent, written by nobody.

The model wasn't the problem. It knew nothing about me.

The brain isn't the AI. It's the files.

So I stopped looking for the right prompt and built a system. The knowledge and the rules live in markdown files the project reads as context:

• Positioning: who I am, what I'm not, the filter I use to discard an idea
• Style: the rhythm, the openings, how I actually write
• Posts: everything I've published, with the performance of each one
• Forbidden: what must never come out
• Frameworks, Hooks, Ideas: the rest of the machine

None of those files existed before. I didn't start from rules, I didn't have any. I generated, I corrected, and every correction became a line.

How the work splits

I update the brain in Claude Cowork, which writes directly to the files. I give the feedback out loud, it rewrites the rule where it belongs.

I write the posts in Claude chat, from my phone. The brain is already there. I propose the idea, we argue about the angle, the text comes out.

The carousel is generated by Uramaki, starting from the finished post.

What came out isn't a text generator

It's an editorial director that knows everything I've published, knows what worked, and tells me no when the idea is weak. It doesn't write in my place. It argues with me.

The idea stays mine. So do the style and the final judgment. What disappears is the time I used to spend looking for how to say it. Same standard, a third of the time.

The value isn't in the model. I could swap the model tomorrow and lose nothing. The value is in the files: knowledge I'd carried in my head for years that nobody had ever asked me to write down.

Takeaway

Building a system that writes like you isn't a prompt problem. It's an externalization problem. The rules you apply without noticing have to become text before anything else can apply them, and you find them by correcting output, not by sitting down to list them.

If tomorrow you had to hand a system the part of the work you do best, which rule would you discover you'd never written down?