Make It Sound Like You: Taste Belongs in the Setup

I kept thinking about my last post.

I wrote that you don’t love what you didn’t make — that AI output arrives finished and confident, and because it never passed through your hands, the reviewer in you clocks out. You skim, you nod, you ship.

I stand by that. But I missed something. There’s a lever I skipped right over.

The part I got wrong

I framed it as a process problem. Rewrite it, argue with it, do the last mile by hand. All true. But all of it happens after the output shows up. I was treating style as something you fix at the end, with your own hands, one document at a time.

That’s backwards. The place to fix it is the setup.

Because here’s what actually happens when the output sounds nothing like me: I don’t engage with it. Not “I engage less carefully” — I don’t engage at all. It reads like a stranger wrote it, so I treat it like a stranger’s work. I file it and move on. The output being fine was the trap in the last post. The output being not mine in voice is the trap before that one.

Taste is a config, not a mood

You know what’s rewarding? When the model hands you something and it already sounds like you’d have said it.

Suddenly you’re in it. You catch the weak sentence, because it’s a sentence you’d never have written. You feel the wrong word the way you feel a typo in your own name. Voice is what turns the reviewer back on. If it sounds like me, I defend it like it’s mine.

And voice isn’t something you should be re-supplying by hand on every single task. That’s the mistake. You’d never re-type your git config for every commit. You set it once, in one place, and everything downstream inherits it.

Style deserves the same treatment. One central place in your AI setup where you write down, explicitly and in detail, how you want the output to read. Not “be concise.” That’s not a preference, that’s a fortune cookie. I mean the specifics:

• no em-dashes as insets, use commas or parentheses
• no “not X, but Y” antithesis constructions
• short paragraphs, one-sentence lines for rhythm
• no marketing fluff, no emoji headings
• first person, direct, dry

Mine lives in a file the assistant reads on every session. It’s boring and it’s long and it’s the highest-leverage thing in the whole setup. Every response passes through it. I wrote it once. It pays out forever.

Why this is the real fix

The last post ended on caring. That the danger isn’t the model being wrong, it’s you not caring whether it’s wrong.

Caring doesn’t switch on by willpower. I tried. “Review the AI output more carefully” is a New Year’s resolution, not a system. It fails the same way every resolution fails, because it asks the tired version of you to do the work the setup should have done.

But caring does switch on by recognition. When the thing in front of me sounds like me, I can’t not engage with it. The ownership the last post was missing? A style file manufactures it. It gives you back the “this is mine to defend” reflex, before the output even lands, without a rewrite.

So the sequence flips. Old way: get output, then work to make it yours. New way: define what yours sounds like, once, so the output arrives already half-owned.

If you can’t grep your own taste, you don’t own your output. Write it down. Put it where the machine reads it every time.

The output was never the point. But it turns out the voice is how you get yourself to care about the point.