“Draft a blog post” is four words.
Behind those four words sits everything I’ve picked up about how I open and how long I let a paragraph run. Which words I’d never say out loud. When a joke helps and when it’s trying too hard. I’ve never written most of it down. I’ve just been doing it for 20+ years.
A model can’t predict what nobody told it. So it falls back on the middle of everything it has ever read. That average is slop.
That’s the trap. The request looks small, four words and about five seconds of typing. But the standard behind it took a career to build and lives entirely in my head. I know it when I see it, and that feels obvious right up until I try to explain it to someone else.
You’d never hand a four-word brief to a new hire and expect your voice back. The same is true for AI.
Once I started trying to write my own standard down, the first thing I got wrong was treating “good” as one thing. It’s three, and each one needs different treatment.
Bad writing. Cliches, filler, inflated adjectives, sentences that wander. A tired writer produces the same stuff on a Friday afternoon.
AI tells. Where most people aim and struggle.
Voice. Yours specifically, which is the only one of the three that makes the work recognizably yours.
The middle one is worth a minute. Everybody’s favorite tell is the em dash. Ban it and you’ve caught a symptom. The real tells live in structure and rhetoric, not vocabulary.
The antithesis flip, where every point arrives as “it’s not X, it’s Y.” Negation that serves the cadence without correcting anything. Three items crammed into one sentence because three has a nice ring to it. Paragraphs of identical length. A closing section that summarizes the post you just finished reading.
You can strip every suspicious word out of a draft and still end up with a voice that sounds artificial.
I ended up with two files.
One holds the universal negatives. No em dashes. No semicolons. A list of hedges I don’t use and a list of phrases I’d never write. A ceiling on sentence length. Mechanical stuff that a script can easily fix.
The other holds my voice, and it’s mostly positive rather than prohibitive. How I open. How I land a paragraph. The moves I actually make, like implicating myself in a critique instead of lecturing from above. Real passages of mine that I believe best represent my style are tagged as on-voice. One deliberately terrible passage tagged off-voice, so there’s something to steer away from.
Then the check runs in two passes. Deterministic first to catch the mechanical failures. Then a judgment pass that reads for the structural tells, because no regex catches those without firing on ordinary English.
Writing it down turned out to be the hard part. Harder than building any of the checking around it. “No semicolons” is something I’d done for years without once saying it out loud. It took someone asking me directly before I could name it. Most of what makes your work yours is sitting in that same blind spot, which is exactly why it never makes it into a prompt.
I’m not going to pretend this is quick. It took real time to get my own standard out of my head and into something a machine could enforce, and I’m still adding to it. Turns out my own preferences are a lot more complex than I ever really considered…
But with consistency the economics flip. The blind spots become few and far between. You update the standard, and every piece after that inherits the fix. Corrections stop evaporating. They stack.