Stay in the Room
13 Jul 2026The post you’re reading took multiple rounds with an LLM before I was willing to put my name on it. Not one prompt and a paste: pass after pass of me handing the draft back. Cut that sentence. That transition is carrying nothing. I don’t talk like that. The punctuation tics are a tell, take them out. The point’s buried in the middle, bring it up top.
What came out is something I couldn’t have written from a blank page, and it sounds like me. Getting the first part is easy now. Getting it without losing the second is the whole problem.
Why it all reads the same
Most AI writing is bad, and you can feel it a sentence in. The frictionless sameness. The competent, weightless prose that could have come from anywhere. A voice with no fingerprints on it.
We know how it gets made, because the workflow has quietly collapsed to three steps: drop the idea into the model, copy the output, ship it. One pass, no friction, done in ninety seconds. The speed is the whole appeal, and the speed is the problem. The draft shipped the instant it existed, before anyone stayed around to make it good.
The pull toward the average
The writing isn’t weightless because the model can’t write. It’s weightless because of how it writes. A language model is trained on an enormous pile of text and learns to produce what’s most probable across all of it. Left to its own defaults, its output is the statistical average of everything it read, and the average of everything sounds like everything, which is to say like nothing in particular. That’s the sameness you’re feeling. It isn’t a flaw the next version quietly fixes. It’s the center of gravity the model always pulls toward, which is exactly why staying in the room is the whole job. On its own the draft settles onto that average; the one force that drags it off, toward an actual person, is someone reading it and saying what’s wrong.
There’s an irony in that. The tool that gets blamed for flattening everything is genuinely good at helping you find your own voice, on one condition: you treat it as a collaborator you push back on, not a vending machine you take output from. It will draft all day. It will not tell you what’s wrong with the draft. That part is yours, and it was always the part that mattered.
Taste doesn’t get automated. It gets amplified. If you have a clear sense of what you sound like and what you’re trying to say, the model gives you more surface to apply it to, faster. If you don’t, it amplifies that too, at the same speed, and ships it.
What staying in the room actually looks like
Written out, my half of the exchange looks like a fight. Cut this. No. Wrong. That’s not it. But it doesn’t feel like one, and this is where “arguing with the AI” gets the picture wrong. It’s closer to working with an editor who drafts fast, never gets defensive, and doesn’t need me to soften a single note. I can say “this whole section is just warm-up, cut it” and get a better paragraph back in seconds, then tell it that one’s worse and go again.
The loop is small and repetitive. Read the draft as a reader, not as the writer. Find the specific thing that’s off: a sentence that’s fake, a transition doing nothing, a rhythm that’s too even, a word I’d never say out loud. Name it directly. Get the next version. Repeat until the tells are gone.
The two posts already on this blog came out of exactly this. Neither was a first draft. Each was multiple rounds of the same back-and-forth, and the reason they read like me is that I kept rejecting the versions that didn’t.
There’s a quieter payoff in working this way. Naming what’s wrong, out loud, round after round, sharpens your own sense of your voice. I edit my own unassisted writing better now, because I’ve spent so many passes articulating what “sounds like me” actually means. The model made me describe my taste, and describing it made it sharper.
The price is speed
This is slow. The post you’re reading was a couple of hours of work, most of it me saying no. The entire promise of AI writing, the reason the copy-and-ship workflow exists at all, is that it’s fast. Staying in the room throws that away and asks for your afternoon.
Here’s why I’ll pay it. The fast version produces the exact prose we all recognize and skip: smooth, surface-level, and identical to every other draft that came out the same way. If what you shipped could have come from anywhere, the speed bought you nothing worth keeping. It gets read once, forgotten, and never tied back to you. The slow version costs an afternoon and produces something a reader connects to a person. For writing that goes out under my name and is meant to sound like me, that trade isn’t close.
The people getting genuinely good work out of these tools aren’t the ones with the cleverest prompts. They’re the ones willing to go twenty rounds when the workflow told them to stop at one.
The job that didn’t move
The model is not the writer, nor is it the editor. It drafts. Editing is the act of knowing what’s wrong and refusing to keep it, and there is still exactly one person in the room who can do that. The models keep getting better at drafting. Caring how the writing sounds is not on that curve. That job didn’t move, and it’s the only one that was ever really yours.