Train the AI you already use on your writing style

By Jack Stovell X · Instagram · LinkedIn · YouTube · TikTok · published 2026-08-21 · updated 2026-11-17

To train ChatGPT, Claude or Gemini on your writing style, give it three pieces you wrote, describe your habits in numbers, and ask for a draft. That works today and costs nothing. It fades over a long answer. A measured voice profile, read live by the assistant, keeps it from fading.

Below is the free prompt builder: paste 120+ words you wrote and get a system prompt written in measured numbers, for any assistant. No signup, nothing stored. Use it as the quick version of everything that follows.

The three routes, compared honestly

There are three ways to get an assistant to sound like you. None of them is a scam, and each one fits a different job.

Prompting with samples is free and fast. You paste a few things you wrote and ask the model to match them. It works for a quick email or a short post. It is also shallow: the model reads your samples once and does an impression of them.

Fine-tuning is the real thing, in a technical sense. You adjust a model's weights using a large body of your text. It is heavy to set up, heavy to maintain, and it gives you no score for any single draft. You find out whether it worked by reading the output and hoping.

A measured voice profile sits between the two. We measure your writing once, as 45 attributes in 8 groups, and write the result down where you can read it. The assistant reads it live, and every draft can be checked against it. It is light to build, readable, and checkable.

You can mix them. Most people start with the first, which is fair enough, because it costs nothing and shows you quickly what the model gets wrong. The rest of this guide takes each route in turn.

Route one: prompting with samples, step by step

You need no tool for this. Here is the whole method, and it is more useful than it sounds.

  1. Pick three pieces that sound like you. Choose things you would sign without editing: a blog post, a long email, a newsletter. Avoid anything a committee touched.
  2. Paste them at the start of a fresh chat, before you ask for anything. Label them as your writing so the model knows what they are.
  3. Name your habits in numbers. Models respond better to "my sentences average about 14 words" than to "I write casually". Count your contractions, note how often you open with a question, say whether you use semicolons. Numbers leave less room for interpretation.
  4. Ask for the draft. Give the topic, the point you want to make, and the length.
  5. Correct by habit. If the draft runs long, say "shorter sentences, closer to 14 words". If it sounds stiff, say "use more contractions". Fix one habit at a time, because a vague note like "make it more me" gets you a vague change.

Our step-by-step post on training ChatGPT to write like you walks through this with more worked detail. Do it once and you will see what a model gets right quickly and what it keeps missing.

Why prompting alone doesn't hold

The first draft is often good. Then it slips. By the third paragraph of a long answer, the sentences have crept back towards the model's default rhythm: tidy, even, a bit beige. There are three reasons.

First, the samples are read once. The model takes them in at the top of the chat and treats them as background. It does not re-read them before writing each paragraph, so their influence thins out as the answer grows.

Second, nothing holds them in a long answer. A model has no ruler to check its own sentences against. It cannot tell that its average has drifted from 14 words to 22, because nobody gave it a way to count. You notice the drift by reading. The model does not notice at all.

Third, a new chat starts from zero. Close the window and everything you pasted is gone. You paste the samples again, retype the habits, and correct the same mistakes you corrected yesterday. That is the real cost of this route: the effort never accumulates.

None of this makes prompting useless. It makes it a good first step and a poor long-term system, especially if you write often or for more than one outlet.

Why fine-tuning solves the wrong problem

Fine-tuning sounds like the obvious fix. If prompting fades, change the model itself. In practice it solves a different problem from the one you have.

It needs a large body of your text, far more than most people have lying around in clean, usable form. It needs technical setup: preparing the data, running the training, and checking that the result has not picked up your worst habits along with your best. It needs a model to host and maintain, and that model ages as the underlying assistants improve. Whatever you build today has to be rebuilt later.

Then there is the part that matters most. A fine-tuned model still gives you no score for any one draft. You get output, and you judge it by eye, exactly as you would with plain prompting. If a draft drifts, nothing tells you which habit moved or by how much.

Fine-tuning changes how a model behaves in general. What you need is a way to check one draft against one voice. Those are different jobs. Our post on a fine-tuned model versus a measured voice sets the two side by side if you want the longer comparison.

What a measured voice profile holds

A profile is a written record of how you write, built from writing you already have. We measure 45 attributes across 8 groups: lexical, syntactic, tone and register, rhetorical, punctuation and format, function words, content patterns, and quirks and cadence.

Some of these are plain counts. Average sentence length and its spread. Contraction rate. Comma density. The mix of I, we and you. Others are judged: formality on a scale of 0 to 10, confidence versus hedging, your usual opening style, how you build an argument.

You can read all of it. The profile comes as a written summary, every attribute with its value, and a confidence score that tells you how far to trust the reading. If it says you write short, direct openings and you know you do not, you can see that and fix it.

Building it needs one sample of 50+ words at the minimum. We recommend three or more pieces and about 3,000 words, and you can add up to 20 samples. Building usually takes under a minute.

When the profile reads wrong, add more samples and re-analyse. A thin profile is usually a thin corpus, so the fix is more of your writing.

Set it up in ChatGPT, Claude or Gemini

We work with the assistants you already use. The profile is read live through an MCP server (41 tools) and a REST API, both on every plan including Free. You can connect it in Claude, in ChatGPT (where custom connectors are on your ChatGPT plan), Cursor, Gemini CLI or any MCP client.

  1. Build the profile. Paste your samples and let the analysis run.
  2. Connect it. In your assistant, add the connector. Where a connector is not available, paste the prompt builder's output at the start of the chat instead. The Claude setup guide and the ChatGPT setup guide cover the exact steps for each.
  3. Ask for a draft in the profile's name. Say something like "draft this in my ScriptGrain voice". The assistant calls `get_profile` to read your measurements and `generate_content` to write with them (1 credit).
  4. Ask for the score. The assistant only scores a draft when it calls `check_voice_match`, so ask it to.

The profile is read live each time, which is why a new chat no longer starts from zero. You are not pasting anything. The assistant reads the same measurements on Monday as it did on Friday.

Score the draft, then fix what moved

Every draft is scored from 0 to 1 against your profile: the voice match score. On-voice drafts typically land between 0.85 and 0.95. Partial matches sit between 0.5 and 0.8. A clearly different voice comes in under 0.45.

The number alone is only half of it. Each score returns deltas: for every habit, the draft's value beside the profile's. As an example, a delta might read: average sentence length, draft 21 words, profile 14. That tells you exactly what to ask for. "Cut the sentences down towards 14 words" is a precise correction, and it is the same correction you would have made by hand, only found in seconds.

When you want the fix done for you, `polish` revises the draft toward a target score (default 0.9) over up to three measured passes, and returns the best version with before and after scores. It costs 1 credit up to 5,000 words, and it is free when the draft already passes. It changes how a thing is said, never what is said: it adds no facts and no first-person stories.

Two limits are worth stating plainly. The score is a style check only. It never checks facts, so a person still reads every draft. And when you make a final edit, `save_edit` keeps it, which feeds the profile's learning. Scoring any text of 20+ words against a profile is free, up to 200 scores an hour.

A memory of what you have already said

A voice is more than rhythm. It is also the set of things you have already argued, and writing the same argument twice in different words is a quiet way to bore your readers.

That is what the Voice Brain is for. It is the memory behind a voice, and it keeps what your writing has already said: ideas, claims, stories and frameworks, each with its exact quote. Before you write, a novelty check (`check_novelty`) scores a brief against it and tells you whether it repeats, builds on, or is new, then proposes angles. "What next" (`what_next`) proposes five angles when you are short of one.

Once a voice has 10 indexed pieces, new drafts build on the memory and do not re-argue points already made. You can edit or delete any memory. It is on every plan, free. The Voice Brain explainer covers how it works in more detail.

What it costs, and what happens to your writing

Free gets you one voice profile and one full analysis, plus the API and MCP server, the Voice Brain and one free draft for the life of the account. No card is needed. Paid plans hold monthly credits:

A draft up to 5,000 words is 1 credit, taken only after delivery. Scoring is free.

On your writing: we never train AI models on it, and our terms with the AI providers forbid them from training on it. Your writing is measured, never trained on. It is stored in the EU (Ireland), and you can download or delete everything from Settings. We are a UK company. ScriptGrain does not publish or schedule anything, so nothing goes out under your name unless you send it.

Questions

Do you fine-tune a model on my writing?

No. We measure your writing and never train on it. The measurements become a profile of 45 attributes that your assistant reads live, and every draft is scored against it. Your samples are not used to train any model, ours or anyone else's, and our terms with the AI providers forbid them from doing so either.

How many samples do I need?

One sample of 50+ words is the minimum, but it gives a thin reading. We recommend three or more pieces and about 3,000 words, and you can add up to 20 samples. If the profile reads wrong, add more of your writing and re-analyse. The confidence score tells you how far to trust the current reading.

Can it learn a brand's voice?

Yes, if you have enough writing in that voice. Build a profile from the brand's best pieces and it measures them the same way it measures a person. Paid plans hold more than one profile: Operator has 3, Studio 10, and Agency has unlimited, so you can keep each brand or client separate.

What happens when my style changes?

Add newer samples and re-analyse, and the profile moves with you. The edits you save with save_edit also feed the profile's learning, so it picks up your corrections over time. If you want to compare old and new, keep the earlier writing as its own profile on a plan that allows more than one.

Does it work inside ChatGPT and Claude?

Yes. The profile is read live through an MCP server and a REST API, so you can connect it in Claude, in ChatGPT (where custom connectors are on your ChatGPT plan), Cursor, Gemini CLI or any MCP client. Where a connector is not available, you can paste the prompt builder's output at the start of the chat.

Is my writing used to train a model?

No. Your writing is measured, never trained on. We never train AI models on it, and our terms with the AI providers forbid them from training on it. It is stored in the EU (Ireland), and you can download or delete everything from Settings. We are a UK company.

What is free?

The prompt builder is free with no signup and nothing stored. The Free plan gives you one voice profile and one full analysis, plus the API and MCP server, the Voice Brain and one free draft for the life of the account, with no card. Scoring any text of 20+ words against a profile is free too.

Related