You have done this. You paste in three of your own emails, you write "match my voice, here are samples," and you get back something that is almost right. The words are close. The facts are yours. And it still reads like a press release from a company that does not exist.
So you try harder. More samples. A style guide in the system prompt. A list of words to avoid. The first paragraph comes back better. By the third one it has drifted back to the same place.
That is not you failing at prompting. It is the model doing what it was built to do.
Why samples do not fix it
Three things are working against you.
The first is what training did to the model. A base model will happily write like anybody. Then it gets tuned on human preference, which means it learns to produce the text that raters liked, on average, across millions of comparisons. Averages have a shape. Balanced. Hedged. Thesis up front, three examples, a summary that restates the thesis. That shape is not a bug the lab forgot to fix. It is the thing they optimized for, and it is very hard to talk a model out of the behavior it was rewarded for.
The second is that a prompt is a soft constraint and sampling is probabilistic. Every token is a draw. Your instruction shifts the odds for a while, and then the training distribution reasserts itself, because it is the stronger prior. This is why style prompts work best in the first hundred words and worst in the last hundred. You are not being ignored. You are being outvoted, token by token.
The third is the one nobody talks about. Few-shot samples move surface features. Word choice, maybe an opener, maybe a sign-off. The tells that actually give a draft away are structural, and they live below the level a sample can reach. Where the point sits in the paragraph. How often a clause hangs off the front of a sentence instead of being joined to it. The variance in sentence length. The rate of em-dashes per thousand words. You can hand a model twenty of your emails and it will pick up your vocabulary and none of your rhythm.
Then there is the small human problem underneath all of it. Most people cannot describe their own writing. Ask a good writer what makes their prose theirs and you get adjectives. Direct. Conversational. Warm. Feed that to a model and you get a caricature of direct, conversational, and warm, which is worse than what you started with.
And you have no way to check. You read your own draft, it sounds fine, because you just read it and you know what you meant.
The marks
Six kinds, and you already know all of them.
Vocabulary. Delve, leverage, robust, seamless, tapestry, myriad, testament, navigate as a verb about anything that is not a boat. Words that arrived with instruction tuning and never left.
The antithesis reflex. It's not just X, it's Y. The question isn't whether, it's how. This is the shape readers spot first and the one almost nobody names out loud.
Throat clearing. It's important to note that. Let me be clear. In today's fast-moving. A sentence that announces itself instead of starting.
Signposting. However, moreover, furthermore, ultimately. Transitions that label a structure the reader can already see on the page.
Rhythm. Every sentence twenty-two words long. This is the strongest signal there is and it is invisible until you measure it, because no individual sentence is wrong.
Punctuation rate. The em-dash is a fine mark. Nobody uses eleven of them per thousand words except a language model.
Naming one is most of the work. Once a phrase has a name you stop writing it, and you stop letting it through.
What rewr.it does instead
Three pieces, and the order matters.
A voice profile that is written down. Not a vibe, not five samples, not "be more conversational." An ordered set of rules about the things a sample cannot carry: the distribution of sentence lengths, joining clauses with and, so, but, then instead of subordinating them, where the point is allowed to land, what a concession sounds like, what a disagreement sounds like, which family of metaphors is allowed and how often. It was built from recordings of a person talking, and then the speech artifacts were deliberately thrown away, because writing that reproduces spoken filler is its own kind of fake. What survives is disposition, rendered as clean prose.
The rules are ordered by weight. The first three do most of the work, and it turns out most of what makes prose sound like a person is structural, not lexical.
A detector that does not use a model at all. Plain code, running in your browser on every keystroke, before anything is sent anywhere. It scores 97 terms across 7 categories, 8 antithesis shapes, 12 stock openers, plus statistics the eye cannot do on its own: em-dashes per thousand words, standard deviation of sentence length, share of sentences under ten words, semicolon and exclamation counts, tricolon candidates.
The rewriter is told about exactly the list the detector matches on. Same source, one file, so the thing writing your draft and the thing grading it cannot drift apart. You get a number before and a number after, and both are measurements rather than promises. The demo on the front page is not a mockup. It is a real draft at 17 tells becoming a real rewrite at 0, scored by the same code that will score yours.
Fact preservation, in front of everything else. This is the part most rewriting tools get wrong, and it is the reason people stop using them. Paraphrase drifts. A date moves. A hedge quietly becomes a commitment. Ninety days becomes six weeks because six weeks scanned better.
Every claim in the output has to trace back to a specific sentence in the input. No new assertion, not even a true one. If a sentence is too tangled to preserve safely, it stays exactly as you wrote it and gets flagged instead of being replaced with something cleaner and untrue. On top of that there is an audit that pulls every number, date, name, URL, quote, and email out of both versions and diffs them, so if something went missing you see it rather than discover it later in a reply from your boss.
A rewrite that reads better and says something you did not say is not a rewrite. It is a liability with good rhythm.
Same voice, different room
A Slack message is not a self-assessment. The voice stays constant and the formality moves, which is the opposite of what happens when a tool "humanizes" a serious document by making it chatty. Four registers ship: Slack, email, skip-level, LinkedIn. On Pro you write your own document types, so a release note or a customer apology or a standup update shows up in the same menu and behaves the way you said it should.
What it will not tell you
It will not promise to beat an AI detector. Those change weekly, any text can be edited past one, and anybody selling you that number is guessing.
It will not claim to be a clone of you. It is one considered voice, held to rules that are published and auditable, and the honest version of the promise is that your draft stops carrying the marks while every fact in it survives. That is the part a reader actually notices.
Try it on something you already wrote
Take the last thing a model drafted for you. Paste it in. Before you run anything, look at what gets marked, because that alone will change how you write next week.
Three rewrites a day without an account. Twenty a month if you sign in, no card. There is a Chrome extension for Gmail and LinkedIn, and an API if you want it in your own pipeline.
Say the thing you meant. Just not the way a model would say it.