How to train AI to write like you

Paste three of your best LinkedIn posts into ChatGPT and ask it to write like you, and what comes back will not be you. It will be a competent stranger doing an impression: your sentence length, none of your judgment. That is what most people mean when they search train ai to write like me or teach ai to write like me: paste in examples, get consistent output. It rarely survives past the first post. What actually holds a voice steady at post fifty is not a longer prompt. It is treating your own writing as evidence, and correcting the rules extracted from it one at a time instead of re-explaining everything from a blank chat window.

Why does the impression fall apart by the fourth post?

A first draft usually looks close enough to be convincing. The model has three examples fresh in context, and it pattern-matches the obvious surface markers: short sentences, a specific opening move, maybe your habit of starting a paragraph with "Say" instead of "For example." That is real signal, and a language model is genuinely good at copying it for one output.

The second draft is where it slips. Ask for a different topic in a new chat, and the model has no memory of what it kept or dropped last time. It re-reads the same three examples and pattern-matches slightly differently, because nothing forced it to commit to a rule. One week your posts open with a question. The next week they open with a flat statement, and neither version got corrected, because nobody was checking for the pattern. They were only checking whether the post read fine on its own.

Why does posting more make the drift worse, not better?

The instinct is that more content should mean more signal for the model to lock onto. The opposite tends to happen once a founder is publishing three or four times a week. More posts means more topics, and more topics means more chances for the model to borrow phrasing from whichever example is topically closest, structure and all, instead of the one that actually represents how you write. A founder posting about a delay gets matched against the one example in the prompt that was also bad news, even if that example happens to be the least representative post in the whole set.

Volume also hides the problem from the person who should be catching it. Reading three posts closely is easy on a Sunday afternoon. Reading the fortieth post from that same month, on top of running the company those posts are about, is not. The drift accumulates exactly where nobody is looking closely enough to see it.

What is a prompt actually doing with those examples?

This is a documented technique, not a guess. Anthropic and OpenAI both call it few-shot or multishot prompting: give a model a handful of input and output pairs, and it steers its next output toward the pattern in those pairs. Anthropic's own prompting guidance recommends three to five examples for the technique to work reliably, diverse enough that the model "doesn't pick up unintended patterns" (Claude's prompting best practices). OpenAI's guidance lands on the same point in fewer words: show "a diverse range of possible inputs with the desired outputs" (OpenAI's prompt engineering guide).

Read closely, both vendors are describing the same target: output format, tone and structure. Neither promises judgment. A model shown three posts learns that your sentences run short and that you like a colon before a list. It has no way to learn that you would never publish a launch announcement without hedging the timeline, because that rule only shows up as an absence, in the posts you didn't write. A prompt can only show a model what you did write, never what you deliberately left out.

What can a handful of examples actually teach a model?

Say you never end a post with a question, on principle, because it reads like you're fishing for comments. That's a real rule, and it is exactly the kind of thing three pasted examples will not reliably encode. "Never" gets proven by the absence across dozens of posts, not the presence in three of them. A model working from a small sample has no way to tell the difference between a rule you hold and a coincidence.

The same problem shows up with hedging. Some founders always soften a claim about their own product, saying "we think" rather than "we know." Others never do, on principle, because hedging reads as weakness in their specific market. Get this backwards once and the post reads like someone else wrote it wearing your name, even though every individual sentence is the kind of sentence you'd write.

What does extracting a rule from evidence actually look like?

The alternative to hoping a pattern survives a handful of examples is pulling rules out of a larger body of writing deliberately, one at a time, each anchored to the sentence that proves it. Not "this founder writes short sentences," which is an impression. "This founder never opens with a statistic, checked against the last forty posts, zero exceptions," which is a rule with evidence attached.

That distinction matters because a rule with evidence can be checked. If the forty-first post opens with a number, that's either a mistake to fix or a genuine exception that updates the rule. An impression can't be checked the same way. It just gets vaguer the more content you feed it.

What would that look like without any tool at all?

You do not need software to test whether this beats a three-example prompt. Pull your last twenty or thirty posts and read them as evidence instead of as content you already know.

• Find one sentence that would embarrass you if an AI had written it in your voice, and work out exactly what makes it yours rather than generic.
• Write the rule down as a sentence, not a vibe: "Never opens with a stat," not "punchy openings."
• Check it against five more posts before you trust it. A rule that only holds for two examples is still a coincidence wearing a label.
• When a new draft breaks the rule, decide whether to fix the rule or fix the draft, and write down which one and why.

That is slower than pasting three posts into a chat window. It is also the only version of "training" that gets more accurate the longer you keep doing it, instead of resetting to zero every time you open a new conversation.

Why does correcting a rule work better than re-prompting?

Every fresh chat starts from zero. Paste the same three examples into a new session next month and you are teaching the model your voice again, from scratch, with no memory of the correction you made three weeks ago about how you handle a customer complaint in public. The usual fix is a longer prompt: more examples, more instructions, a style guide pasted at the top. That helps for exactly as long as the prompt fits in the model's attention and nobody forgets to paste it.

Correcting a rule is a different action from writing a better prompt. It fixes the mistake at the source, so the next draft doesn't make it again without you re-explaining anything. A prompt rewritten every session is not a memory. It is a note you keep losing and rewriting from partial recall.

What happens when two rules conflict?

They will, eventually. A founder who hedges every product claim also, on occasion, wants a flat declarative sentence for emphasis. Both are real. A style guide written once tends to flatten this into a single instruction, always hedge, because a document can't hold two rules that depend on context without turning into something nobody actually reads before drafting.

Evidence-based extraction can hold the exception, because the exception has its own proof: the specific post where the founder broke their own pattern on purpose, and why. Kept, edited or rejected, one rule at a time, is slower than pasting a paragraph of instructions. It is also the version that survives contact with fifty more posts instead of three.

What changes once the rules actually hold?

The obvious answer is that drafts land closer to right on the first pass, and that's true, but it undersells the real shift. Attention moves. Instead of reading a draft top to bottom hunting for anything that sounds off, the check gets shorter: does this break a rule that's already confirmed, or does it break new ground the rules haven't covered yet. The second kind of edit is the interesting one. It's where actual judgment still lives, and it's the only kind worth spending a Sunday afternoon on.

Posts that used to take four rounds of correction start taking one, not because the model got smarter overnight, but because it stopped guessing at something you'd already told it, explicitly, with evidence, three weeks earlier.

Does this replace judgment entirely?

No, and treating it like it does is its own failure mode. A rule extracted from forty posts tells you what you have done. It says nothing about what you should say the morning a deal falls through, or whether a sharp line about a rival reads fine today and reads as a problem once the news cycle catches up with it. That call stays yours, made in real time, with context no archive of old posts contains.

What the rules remove is the repetitive part: relearning your own habits every time a blank prompt opens. That frees the actual judgment calls for the moments that deserve them, instead of spending that attention re-deciding, for the hundredth time, whether you'd ever open a post with "excited to announce."

How do you know it's actually working?

The test is not whether a single post reads well. A generic AI draft clears that bar most weeks without needing to sound like anyone in particular, and so does a rushed post from a cheap freelance tier. The real test is whether a colleague reads the fortieth post and doesn't ask whether you actually wrote it.

Watch for drift the way you would watch a ghostwriting arrangement for the same failure. Corrections getting shorter and less specific, instead of sharper, is the sign that whatever is holding your voice has stopped learning and started guessing.

None of this requires giving up on AI drafting. It requires being honest about what three pasted examples can and cannot do, and building the extraction and correction step instead of skipping it because the first draft looked close enough. The voice is the part that doesn't scale by itself. If you're trying to get past the impression stage instead of publishing another one, Pointed's early access is open.