Why “write like me” never works

You have done this. You were drafting something with an AI, felt the beige creeping in, and typed the obvious correction. Write in my voice. Make it warmer, less like a press release. The model agreed at once, thanked you for the helpful note, and returned a paragraph identical to the first one with two extra contractions and an exclamation point it seemed rather proud of. The request was reasonable. It was also hopeless, for reasons worth knowing, because they are the same reasons a real fix has to look nothing like a request.

The failure is not the model being lazy. It is built into what you asked and how the thing works, and it stacks up in a few specific places. Walk through them with me, because the shape of the problem is the shape of the answer.

01An adjective is a destination, not a map

Start with the words you reached for. Warmer. More human. Each one feels exact in your head and lands as mush on the page. The model has read a staggering number of things people have called “warm,” from birthday cards to passive-aggressive HR emails, and when you ask for warm it quietly averages all of them. You handed it a destination and withheld the route. What comes back is the statistical middle of warmth, which is a temperature nobody actually writes at.

This is the trouble with every style adjective. Punchy could mean Hemingway or it could mean a LinkedIn motivational post, and the model has no way to know which dialect of punchy is yours. Vague instructions do not narrow its choices; they hand the choice back to the average. You wanted to sound more like one specific person and accidentally asked to sound like everyone who has ever been described the same way.

“Sound like me” is a wish. It is not an instruction.

02You are the worst witness to your own voice

Suppose adjectives did work. You would still pick the wrong ones. Your real style lives in machinery you cannot feel: the throwaway function words you lean on, how far you let a sentence run before you cut it short, whether you reach for a story or a statistic when you need to prove something. Stylometry has known this for sixty years — the signal is in the parts you stopped noticing in grade school.

Ask people to describe their writing and they reach, understandably, for who they would like to be. They say they are witty and direct. The writing says they hedge on every other line and have never met a semicolon they didn’t want to keep. That gap is not dishonesty. It is that your fingerprint sits on the inside of your own hand, where you of all people cannot get a good look at it. So the description you give the model is sincere, flattering, and about a person who does not quite exist.

03The model has gravity

Now hand the model a perfect description anyway. You still have to win a fight against everything it has ever read. Its defaults were laid down by training on a serious fraction of everything ever published, then tuned by people paid to prefer the safe and the agreeable. That produced a confident, frictionless house style with real mass. Your instruction is one sentence. The house style is the whole library leaning the other way.

And the little weight you did add will not stay put. Models pay closest attention to the beginning and the end of whatever sits in front of them, so as a conversation grows your careful note drifts into the soft middle where attention goes to die. Ten exchanges later the model is not defying your voice on purpose. It genuinely cannot find it anymore. That is why the first reply sometimes almost lands and the fourth has quietly reverted to corporate. The instruction did not fail all at once; it evaporated.

04What actually moves it

Look at the shape of the failure. You were vague when the model needed something exact. You described the writer you wish you were instead of the one the page reveals. And the single sentence you offered was never going to outweigh the library that set the defaults in the first place. A real fix has to turn each of those around.

So you stop describing and start measuring. You let a system watch which way you actually lean rather than taking your word for it. Then you write the result down as a standing instruction, with hard rules and named exceptions, so it keeps working after the second message instead of dissolving. That last part matters more than it sounds: the difference between a prompt and a profile is mostly whether it survives the conversation.

None of that is a plea for the model to be more like you. It is a set of constraints the model can follow without guessing, because there is nothing left to interpret. That is the entire idea behind a Timbrel voiceprint: the builder shows you real passages, watches which way you lean, and turns the answer into exactly this kind of file. It is more work than typing “sound like me.” It is also the only version that works.

The chatbot, in the end, was never being difficult. It did precisely what you asked, which was the problem. Sound like me is not something it can do. Use these words, ban those, swing your sentence lengths, and keep this one phrase is. The second one is just tedious to write by hand, every time, forever, which is the whole reason to measure it once and keep the file.

Written the hard way

This used the same Timbrel voiceprint as the rest of these notes — witty, relaxed, measured. It was told, among other things, never to ask a rhetorical question and never to bundle ideas into tidy threes. A post arguing against vague instructions had to take its own seriously.

Build your own →