Getting AI to write in your voice (and why "write like me" fails)

You paste three of your best posts into ChatGPT, add "write in this voice", and get back something that is unmistakably not you. It has your topic and none of your speech. Add more examples and it gets more average, not more accurate.

This is not a prompting skill issue. The instruction "write like me" asks the model to infer a voice from evidence and then apply it, and the inference step is where it fails. What the model extracts from three posts is not your voice, it is the genre those posts belong to.

What the model actually learns from your examples

Give a model five LinkedIn posts and it will reliably pick up:

• The format: line breaks, list use, length, whether you open with a one-liner.
• The register: casual or formal, first person or third.
• The topic domain: SaaS, recruiting, design.

Those are the visible layer, and copying them produces something that looks like your posts from a distance. What it will not pick up from five examples:

• Which words you never use.
• Which claims you would refuse to make.
• What you find embarrassing.
• The specific things you have actually done, which are the only source of material that could not have come from anyone else.

So the output lands in a predictable place: correct shape, generic substance. It reads like a competent stranger doing an impression, and readers detect it immediately even when they cannot name what is wrong. That gap is why AI-written posts get ignored rather than disliked. Nothing about them is bad. There is just nobody in them.

A voice is three separate things

Treating "voice" as one setting is what makes it unfixable. It is at least three, and they need different inputs.

Style. Sentence length, punctuation habits, whether you use questions, how you open and close. This is the layer examples genuinely teach. It is also the least important, because it is the layer readers notice last.

Vocabulary boundaries. The words that are yours and the words that are disqualifying. Most people cannot list these from memory but recognise them instantly: "leverage", "unlock", "in today's landscape", "game-changer". A negative list works better than a positive one, and it is much easier to write.

Substance. The facts, numbers, opinions and stories you are allowed to draw on. This is the layer that makes a post identifiably yours, and no amount of style transfer produces it. If the model does not have your material, it will invent something plausible in your shape, which is the worst of both.

Almost everyone tries to solve a substance problem with style instructions. That is why the loop feels endless.

What to give a model instead

If you want output you would actually publish, give it these four things and stop tuning adjectives.

Ten to twenty of your own posts, not three. Three is enough to average, not enough to characterise. And use ordinary posts, not only your best: your best posts are your most edited, which means they are the least like how you write.

A short list of banned words and constructions. Ten to fifteen entries is plenty. Include the ones that are technically fine but not you. This is the single highest-return input and takes fifteen minutes; if you want a starting list, the common AI tells are most of it.

A file of facts you are allowed to cite. Numbers with the date they were true, customer results with permission status, opinions you hold and would defend. This is boring to assemble and it is the entire difference between a post about your company and a post that could be about any company.

Two or three examples of what you rejected and why. "Too salesy", "I would never say hustle", "this claims something I cannot prove." Rejections carry more signal per word than approvals, because approval only says the output was acceptable while a rejection names a boundary.

The test that tells you whether it worked

Take the draft and remove your name, your company name and the topic. Show it to someone who knows you.

If they can tell it is yours, the voice landed. If they say "this could be anyone in your industry", it did not — regardless of how well it matched your line breaks.

A cheaper version you can run alone: read it out loud. The places you stumble or feel faintly embarrassed are almost always sentences you would not have written. Voice mismatches are much easier to hear than to see, which is why they survive so many silent re-reads.

For the mechanical layer, the constructions that mark text as machine-written regardless of whose voice it was aiming at, you can run a draft through our AI writing check and see the specific spans flagged rather than a score.

Why a chat window keeps losing your voice

Even when you get a good result in ChatGPT, it does not persist. Next week you open a new conversation, the model has none of it, and you re-explain from scratch or paste a prompt block you now maintain by hand.

This is the actual daily cost, and it is not about output quality at all. It is that the voice lives in a conversation rather than in the account, so every session starts at zero and every improvement has to be manually carried forward. Anyone who has used ChatGPT or Claude for social content for more than a month has a document somewhere that exists only to re-teach the model things it already learned.

Making the voice a stored profile rather than a prompt is most of what SelfSM does: the samples, the banned list, the facts and the rejections accumulate in one place, and every draft starts from them instead of from a blank context. The interesting part is not that drafts get better. It is that a correction you make once stops being a correction you make weekly.

What will still not work

Two honest limits, because the promise gets oversold.

Nothing gives you material. If you have not done anything worth writing about this month, no voice profile produces a post worth reading. The model can shape what you have; it cannot supply it. This is the same reason hiring a ghostwriter does not remove the founder from the loop.

Voice is not a substitute for judgment. A well-tuned model will happily write a sentence in your voice that you should not publish. The last read is still yours, and the useful version of that read is checking claims rather than admiring phrasing.

FAQ

Why doesn't ChatGPT sound like me even when I give it examples? Because from a handful of examples it extracts format and register, the visible layer, not your vocabulary boundaries or your material. You get the right shape with generic substance, which reads as an impression. Fix it by supplying more samples, an explicit banned-word list and a file of facts you can cite, rather than more adjectives.

How many writing samples does an AI need to learn a voice? Ten to twenty is the useful range for social posts. Three averages rather than characterises. Include ordinary posts, not only your best ones — heavily edited work is the least representative of how you actually write.

Is it better to describe my voice or show examples? Both, for different layers. Examples teach rhythm and structure, which people describe badly. Explicit rules work better for boundaries: what you never say, what you refuse to claim. Descriptions like "professional but approachable" carry almost no information and are the most common wasted instruction.

Can AI-written posts be detected? Detectors measure statistical patterns and are unreliable on short text, so a confident verdict on a 200-word post is worth little either way. Readers are the real detector, and they respond to the absence of specifics rather than to phrasing. See what AI detectors actually catch.

What is the fastest thing I can do to make AI drafts sound more like me? Write a list of ten words and phrases you would never use, and put it in the prompt. It takes fifteen minutes, needs no tooling, and removes most of what makes a draft read as machine-written.