For years I assumed knowing how I write and being able to describe it were the same thing. Ask me how I write, I’d give you a reasonable answer: short sentences, plain English, no corporate jargon, commit to the opinion. That’s not wrong. It’s just not useful as a spec.
I found this out when I tried to build one.
The blog automation pipeline runs daily. A cloud routine reads my memory files, finds the compelling narrative thread, and writes a draft. I flip draft: false and it publishes. That workflow needed a reference doc: something the AI could read before drafting so the output sounded like me and not like a press release.
First I tried analyzing the existing posts. Problem was, most of them were already AI-drafted. The voice in them was Claude’s interpretation of what I sound like, not my actual voice. Circular.
Then I tried a structured questionnaire. How do you open a post? How do you end one? What’s your relationship to hedging? That failed faster. Answering questions about how you write surfaces your theory of yourself, which is not the same thing as your actual behavior. I had tidy answers for all of them. The answers were wrong in ways I couldn’t see yet.
So I just wrote something instead.
Two real assignments. Publishable formats, one sitting each, no structural rules imposed. Just: here’s a prompt, write it the way you’d write it. I wrote two posts and handed them over.
The patterns that came out were things I hadn’t named. The opener anchored to a specific moment or belief, never abstract scene-setting. The mid-post tone reset using a short phrase to shift register without breaking voice (“Don’t get me wrong, it hasn’t been all sunshine and rainbows”). The closer landing on one committed sentence rather than summarizing anything. None of that was in my questionnaire answers.
Because I hadn’t known to name them. That’s the thing.
That’s now voice.md, sitting in the project repo. Every morning at 5am Pacific, an AI reads it in full before writing the day’s draft. It reads my memory files, finds the narrative thread, writes about 500 words, and pushes to master with draft: true. I flip the flag. It publishes.
The version of me who tried answering the questionnaire would have produced a spec that was mostly accurate and mostly useless. The version of me who just wrote for an hour produced something the AI can actually operate from.
You can’t answer your way into a voice spec. Give yourself the assignment.