How to actually train an AI to write like you

Ask any model for "a LinkedIn post in my voice" and read what comes back. The verbs are vague, it opens with a rhetorical question, and somewhere in the middle a sentence tells you the thing you do "is not just X, it's Y." It is competent and it belongs to no one, and you could swap your name for a stranger's without a reader noticing.

That is not a failure of the model. It is an accurate average of everyone who has ever written a LinkedIn post, which is exactly what you asked for when you gave it nothing but the request. A 2025 study testing whether models can copy individual writers hit the same wall: without example prompts, the authors write, "LLMs default to a generic style". The part that should stop you is what happened when they added more examples. Going from a couple to ten of them, they found, "affects the four metrics very little, suggesting limited gains in stylistic alignment". Pasting in more of your writing barely shifts the result, which retires the most popular shortcut before we begin.

So the work is not a cleverer prompt. The closest thing it resembles is onboarding a sharp new hire who writes nothing like you yet, and the way through is real material, line-by-line correction, and your judgement held in the loop until the end.

The lazy way

What you give ita one-line prompt

The voicethe default

The workpaste, publish

Sounds like everyone's AI

The real way

What you give ityour emails and pitches

The voicecorrected line by line

The workhours, once

Sounds like you

Feed it the real stuff, not the polished stuff

The instinct is to hand the model your best published essays. That is the wrong corpus, because polished work has been edited toward a house style and away from you. The tics that make your writing yours, the ones a reader would recognise in a blind test, live in the unguarded places.

I built mine from sent emails, the ones I actually fired off rather than the ones I saved as drafts. The reply I sent a client at half-eight in the morning says what I think in the order I think it, before any editing tidied it. Then voice notes I had transcribed, because the way I talk sits closer to my real written voice than my careful "writing voice" does. Paul Graham has a line for why this works: "Informal language is the athletic clothing of ideas." The register you reach for when you are not performing is the one worth teaching a machine.

Pull thirty or forty of these. WhatsApp messages where you explained something to a friend. The pitch you wrote at speed because the deadline was that afternoon. The reply on a thread where you got slightly annoyed and stopped being careful. That last category is gold, because annoyance strips the corporate varnish off faster than anything.

Correct it line by line, the way you would a person

Here is the part everyone skips, and it is the part that does the work.

Give the model a real brief and let it produce a draft. Then do not regenerate. Go through it line by line and rewrite the parts that are wrong, in your own hand, and tell it why each one was wrong. Not "make it punchier." Tell it the specific thing: I would never open with a question. Cut the word "leverage." That semicolon is not how I'd say this out loud, make it a full stop. This sentence is doing the mirror trick where it says something is not one thing but another, I don't write like that, so kill it.

You are not prompting at this point. You are building a list of rules out of real corrections, and the rules that come from fixing an actual sentence are sharper than any you would invent in the abstract. After a dozen passes you will notice you keep flagging the same five things. Those five things, written down, are most of your style.

Then give it the other half: the lines you would never say. A do-not list is more useful than a do list, because your voice is defined as much by what you refuse as by what you reach for. Mine bans em-dashes, the word "delve," anything that calls a result "game-changing," and the girlboss register entirely. The model cannot infer a refusal from your samples, because the thing you refuse to write was never in them to learn from. You have to state it.

Keep the judgement human

After enough rounds the model produces a draft you can edit lightly instead of rebuild, and that is the finish line, with one thing worth being precise about. What you have is a fast first-drafter that has learned your constraints, not a machine that writes as you, and the last call still belongs to you.

I read every line out loud before it goes out, which is Graham's own test for whether a sentence is really yours. The model gets me to a draft in a fraction of the time. It does not get me to the end, and the day I let it would be the day the writing stopped sounding like a person, which was the only reason anyone read it in the first place.

The shortcut people want is a prompt that skips all of this, and after a fair amount of looking I have not found one. What I have found is a few afternoons of feeding it your real material and correcting it like a hire you rated, after which it gives those afternoons back to you most weeks. The cost is paid once, up front, in your own attention, and that is why most people keep sounding like everyone else. They wanted the voice without the part that makes it theirs.