AI rewriter that keeps your voice: rewrite any draft into a measured style

By Jack Stovell X · Instagram · LinkedIn · published 2026-09-21 · updated 2026-10-02

An AI rewriter only keeps your voice if the voice is measured first. So this one starts by measuring your sample: counting sentence length, contractions, commas, the lot, then writing the rewrite instructions in numbers rather than adjectives. Vague prompts produce vague voices. Numbers don't.

Paste a sample into the builder below (120 words or more does it) and watch it get measured in real time. It'll count the same features the paid product counts, then hand you a system prompt built entirely from your own numbers, ready to run through any assistant.

AI rewriter, paraphraser or humanizer: which one you need

If you typed "AI rewriter", "text rewriter" or "rewrite AI", you want a tool that reworks a draft, and the right one depends on what you want left standing at the end.

Three tools share those search words. They do different jobs.

A paraphraser changes the wording and keeps the meaning. Synonyms swapped in, clauses reordered, the same point said another way. It does that well. What it lacks is any idea whose voice the result is in. Feed it a clipped, contraction-heavy paragraph and it can come back smoother and stiffer than you wrote it. So a paraphraser fits one job: rewording a passage when you only care that the meaning survives.

A humanizer works on a different problem. It rewrites a draft to strip out the measurable tells of machine writing: evened-out sentence lengths and stock connective phrases. That's the right tool for a draft that reads machine-made. There's a longer explanation on the humanize page. To be fair, it does what it says. But it aims at "human" in general, never at a particular writer.

A voice rewriter is what this page does. It rewrites toward a measured style: your own, or one you may use. The result keeps your sentence rhythm, your contractions, your punctuation habits and your register. So it fits anything that will go out under your name. Emails, posts, a client report. Anywhere a reader who knows you would notice the join.

Here's the honest route on this page, in three steps.

1. The free builder reads your own writing and writes the rewrite instructions, in numbers, for whichever assistant you use.
2. Any assistant does the rewrite. ChatGPT, Claude, Gemini, whatever you already have open.
3. The free comparison scores the result against your own writing. At 0.80 or above it reads like the same writer; 0.55 to 0.79 is drifting; under 0.55 is a different voice.

The result names the top differences and the top matches, feature by feature, with the two values side by side. So you don't just learn that a draft missed. You learn that the sentences ran long, or that the contractions vanished. Fair enough, that's a lot more useful than a vague "doesn't sound like you".

The builder stays free. And you don't need the app to use it.

If you want the rewrite done for you, the paid route is Transfer, in the app (Writer plan and above, 1 credit). Paste a draft of 20 words or more, pick a voice, and it's rewritten in that voice, keeping the ideas, facts and structure. Polish goes a step further: it revises toward a full profile, measuring as it goes.

Pick by the job. Rewording a passage: paraphraser. A draft that sounds machine-made: humanizer. Something with your name on it: voice rewriter. Only one of the three has ever heard of you.

Why 'rewrite this in my style' fails

"My style" isn't an instruction. It's a vibe. And large language models are bad at vibes when there's no measurement underneath them.

Ask an assistant to sound more "confident" or "punchy" and it'll guess. It might add exclamation marks. It might shorten a few sentences and call it a day. What it won't do, because it can't, is know that your writing runs 0.5 contractions per 1,000 words less than it should, or that your average sentence length needs to drop from 17 to 11.

SGR-002 tested this directly. A one-page measured prompt (asking for roughly 14 contractions per 1,000 words, formality at 3.7) got GPT-5 to obey the countable rules: no em dashes, more short sentences. But on the things that weren't spelled out as numbers, it drifted hard. It produced 0.5 contractions per 1,000 words instead of 14. Judged formality landed at 7.2, nearly double the target. The result scored 0.62 against the study profile. ScriptGrain's generation, working against the full profile rather than a static prompt, scored 0.83.

So the prompt worked, up to a point. It just wasn't the same thing as measurement continuing through the whole rewrite.

The free workflow: measure, rewrite, score

Three steps, and none of them cost anything.

1. Measure. Paste a sample of 120 words or more into the builder below. It runs in your browser, counts sentence length and variation, contractions per 1,000, punctuation rates, pronoun mix, and the judged features (formality, humor, structure) too.
2. Rewrite. The builder turns those numbers into a system prompt: not "sound casual" but "formality 3.7, contractions around 14 per 1,000, sentences under eight words for roughly a third of the total." Paste that prompt alongside your draft into any assistant.
3. Score. Run the result back through the free off-voice check. It compares two pieces with no fixed profile: 0.80 or above reads as on-voice, 0.55 to 0.79 is drifting, under 0.55 is off. It also needs 120+ words per piece, and it stores nothing.

That third step matters more than people expect. A rewrite that feels right can still measure wrong, and you won't know which until you check.

For the mechanics behind all of this, the full method sits at /reference/voice-measurement-framework.

Rewrite ChatGPT output in your own style

You rewrite ChatGPT output in your own style by measuring your own writing first, then setting the builder's target to ChatGPT, then scoring what comes back against the thing you actually wrote. Skip step one and you're just guessing at what "your style" even means. So don't.

Here's the sequence. Paste 120+ words of your own writing into the free prompt builder (up to about 3,500 words if you've got it). That gets measured the same way ScriptGrain measures any voice profile: one small-model judge call, nothing stored. Set the target switch to ChatGPT. That matters, because the prompt gets written differently depending on which assistant is going to read it. Then take the ChatGPT draft you already have, the one that sounds like it was written by a competent stranger, and paste it alongside the generated prompt back into ChatGPT, asking for the rewrite.

Now the part people skip: score it. Run the free comparison tool, put your own original piece in as the main sample, drop the rewrite in as the one being checked. You'll get a band back. At 0.80 or above, it reads like the same writer. Between 0.55 and 0.79, it's drifting, recognisable but the habits are slipping. Under 0.55, it's a different voice wearing your name. Whatever comes back, it names the actual differences, not just a number. That's what you fix.

To be fair, this loop only gets you so far. The free comparison is pairwise: your piece against the rewrite, nothing more. If you want the rewrite polished against a full profile, one that's measured 45 attributes across everything you've written and keeps your samples to anchor on, that's the paid step. Or you use Transfer in the app directly. Either way, the free version tells you where you stand before you pay for anything.

The ChatGPT writing style prompt the builder writes

The ChatGPT writing style prompt is a one-page system prompt built from three blocks: numbers, judgments, and rules, in that order, ending with an instruction to check the draft against every one of them.

Block one is VOICE MEASUREMENTS: counted things. Average sentence length, the share of sentences under 8 words, contractions per 1,000 words, commas per sentence, brackets per 300 words, semicolons and exclamation marks per 1,000, whether questions turn up, ellipsis habits, your I/we/you mix, your short/medium/long word split.

Block two is REGISTER: judged things. Formality on a 0 to 10 scale. Humor. Confidence versus hedging. Opening style, argument structure, rhythm, clause ordering, specificity, metaphors per 1,000.

Block three is HARD RULES, identical for every writer. Quoted verbatim: "Never use em dashes or en dashes. Use commas, colons, semicolons, full stops or brackets." And: "Burstiness: in every paragraph of three or more sentences, at least one sentence under eight words, and adjacent sentence lengths that differ by three times at least once." One more: "The measurements are the ceiling, not the floor."

The ChatGPT version opens with: "You are writing as me. Treat every line below as a hard constraint on how the text is built, not as a mood." It closes by telling the model to check the draft against every number and fix what's off.

Where does it go? Settings, then Custom instructions, pasted straight into "How would you like ChatGPT to respond?" That's the entire setup, covered properly on make ChatGPT sound like you. This is the one-page version of a profile; the full profile measures 45 attributes, keeps your samples to anchor on, and scores every draft against the profile.

What the paid rewrite adds

The free route gives you a prompt. That's honest, and it's useful, but a prompt is static: it doesn't watch what the model actually does with it. It fires once and hopes.

The paid tools measure the output too, then adjust.

Polish (1 credit up to 5,000 words) revises a draft toward a profile's target of 0.9, running up to three measured passes and returning the best one. Instead of one shot at matching a voice, it checks itself against the profile after each pass and course-corrects. Rewrite-page (1 credit per 1,500 words) does the same thing across a whole web page, images considered, not just the copy block. Both return before-and-after scores, so you're not guessing whether it worked.

This is the gap SGR-002 measured directly. GPT-5 with the free prompt hit 0.62 against the author's profile. ScriptGrain, scoring each pass against the full profile rather than firing a prompt and hoping, hit 0.83. The free route is a start. It isn't the ceiling.

If you want to see how a mimicked voice compares to a genuinely measured one, /tools/ai-style-mimic and /research/chatgpt-vs-a-measured-voice both go into more of the same study.

Where the line is

Rewrite into your own voice. Rewrite into a house voice you write for. Rewrite into a voice you have permission to use. Fair enough, that covers most legitimate reasons anyone wants this tool.

What it doesn't cover: a named living person's voice, without their say-so. That's not a legal disclaimer bolted on for the look of it, it's the actual boundary the tools are built around.

To be fair, most competing tools don't even raise the question. Hyperwrite's own ai-style-rewriter page, as it stood on 2026-09-20, ran 652 words with no date, no author, no example, and no evidence the rewrite matched anything. No measurement, so no way to know if it worked, and no attempt to ask whether it should.

HyperWrite is a different kind of tool, and we say so. It matches the style of the examples you paste. We measure 45 named attributes across everything you've published, then keep that profile as something you can inspect, refine with more samples and re-check. The HyperWrite comparison shows what each approach gives you, and why the measurement is the product here.

If you're rewriting brand or organizational copy rather than personal work, /tools/brand-voice-consistency-checker is the more direct fit.

Questions

Does the free rewriter actually change anything, or just suggest tweaks?

It produces a full system prompt built from your sample's measured numbers, sentence length, contractions per 1,000, punctuation rates, judged formality and structure. You paste that prompt with your draft into any assistant, which does the actual rewriting. The builder measures; your assistant executes.

How much text do I need to paste in?

120 words minimum, for both the sample you're measuring and the off-voice check afterward. Shorter than that and the counted features (especially sentence length variance) get unreliable. There's no upper limit, though very long samples take a little longer to process in the browser.

What's the difference between the free check and the paid score?

The free check compares two pieces against each other with no fixed profile: 0.80 and above reads on-voice, under 0.55 is off. The paid route scores against a stored profile, aiming for 0.85 to 0.95, and does it across multiple passes automatically rather than once.

Can I use this to write as someone else?

Only with permission, or if it's your own voice or a house voice you write for. Not a named living person's voice without their consent. The tool doesn't police intent beyond that, but the design assumes legitimate use: your own drafts, your team's brand voice, or a profile you've been given access to.

Why does the paid version run multiple passes instead of one?

Because a single generation, even from a well-specified prompt, drifts on the features that weren't explicitly numbered. SGR-002 found GPT-5 obeyed countable rules from a prompt but still missed contractions and formality by a wide margin. Multiple measured passes let the paid tool check itself and pick the best result rather than the first one.