How to create a personal AI writing style
By Jack Stovell X · Instagram · LinkedIn · published 2026-09-20 · updated 2026-09-30
Most guides tell you to write a persona: "You are a witty, plain-spoken product marketer who values clarity." The assistant reads it and produces the same witty, plain-spoken draft it produces for everyone who wrote that sentence. A personal style is not a persona. It is a set of measurements only your writing produces, stated so precisely that a model cannot drift from them without noticing.
Here is how to build one, from the one-page version to the full profile.
Build the one-page version now. Paste 120 words or more of your own writing below and the prompt builder measures it and writes the system prompt in numbers: sentence length and variation, contraction rate, punctuation habits, formality, humour, how you open and how you argue.
Step 1. Choose the writing that is you
Three or more pieces, ideally 3,000 words between them, written without help: a newsletter, a long email, a post you were proud of. Not the pieces an editor smoothed. The measurement is only as personal as the sample.
Once you've got that sample together, the measuring is the easy part. What you do with the profile is where it pays off. For a one-person business, a measured voice is the one asset competitors can't copy, and we cover the full loop on ScriptGrain for solopreneurs: profile once, generate posts, newsletters and emails, check the score, publish.
Step 2. Measure, do not describe
Sentence length and its variation. Contractions per 1,000 words. Commas per sentence. Asides in brackets. Semicolons, exclamation marks, ellipses. The pronoun mix. Then the judged attributes: formality on a 0 to 10 scale, humour register, confidence against hedging, opening style, argument structure, rhythm. A full profile measures 45 of these; the prompt builder does the ones that matter most for a prompt.
Step 3. State the rules the model breaks by default
No em dashes. No generic transitions. No "not X but Y". A short sentence in every paragraph. The measured numbers are the ceiling for elevation, not the floor. These go into every prompt because every model reaches for the same tells when nothing forbids them.
Step 4. Put it where the assistant reads it
ChatGPT: Settings, then Custom instructions. Claude: a Project's instructions or a Style. Gemini: Saved info. Any assistant: the system prompt. The connector version skips the pasting: ScriptGrain plugs into ChatGPT and Claude directly and hands them the full profile, then scores what comes back.
Step 5. Score, edit, repeat
A style you never check is a style you slowly lose. Score each draft against the measurements, fix the two features that moved, and keep your edits: a profile that learns from them gets closer every month.
Mistakes people make building a writing style
Most failed profiles fail before the measuring starts. They fail at the sample.
The mistakes are boring, and they're all avoidable. Four come up again and again.
- Edited or ghostwritten samples. If an editor touched it, the numbers describe the editor. Use pieces you wrote without help, even if they're rougher than you'd like. Rough is the point. Rough is you.
- Too few words. One short email tells a profile almost nothing. A single sample of 50+ words will technically run, but three or more pieces and about 3,000 words is the recommendation for a reason. Sentence length variation and comma habits only settle once there's enough text to average over.
- Mixing registers. A breezy newsletter plus a formal client proposal gives you a profile of a person who doesn't exist. Formality lands somewhere in the middle, and neither piece sounds like it. Build one profile per register, then keep them apart.
- Describing instead of measuring. "Punchy but warm" is a mood. "Average sentence length of 12 words, wide variance, high contraction rate" is something a draft can be checked against. A model can shrug off the first and can't hide from the second.
Fair enough, you can get by with a decent description for a quick job. But if you want to know why a draft feels wrong, you need numbers to point at.
Keeping the style current
Your writing moves. You change jobs, take on a new audience, get bored of your own semicolons. A profile built two years ago describes the person you were two years ago.
So refresh it. The trigger is simple: when scores on drafts you'd happily publish start slipping, the profile has aged. On-voice drafts typically land 0.85 to 0.95. If your own good work keeps landing at 0.7, the profile is stale, or it was never quite right.
The per-feature deltas tell you which. Each score comes back with the draft value set against the profile value, so you can see whether one feature drifted (say, contractions) or the whole thing did. One feature off? Edit the draft and carry on. Everything off, across several pieces? Add fresh samples.
A refresh needs no ceremony. Pick two or three recent pieces you wrote unaided, add them to the sample pool, and re-run the extraction. It usually takes under a minute (median 39 seconds, measured 28 September 2026). Then keep saving your edits as you go, because the app learns from them. That's the slow, quiet version of a refresh, and it does most of the work.
At the end of the day, a profile is a habit, not a file.
One style across several assistants
Say you draft in ChatGPT, tidy in Claude, and now and then poke at Gemini or Cursor. Pasting a style block into each one is tedious. Worse, the copies drift apart the moment you tweak one.
The alternative is one profile, read live. The MCP server and the REST API hand the same profile to ChatGPT, Claude, Gemini, Cursor, Meta Muse and any other MCP client each time they need it. Change the profile once and every assistant sees the change. Both routes are on every plan, including Free. It's never a pasted prompt, and that matters, because a pasted prompt is a snapshot and a live read isn't.
To be fair to the assistants, their built-in tools are good at what they do. ChatGPT's custom instructions and Claude's Styles are quick to set up and fine for a light steer. The catch is that each one lives inside a single product, in your words rather than measurements. Use them for a nudge. Use the profile when you want the same voice, checked and scored, wherever you happen to be working.
Same numbers everywhere. That's the whole trick.
Questions
How much of my writing do I need?
120 words is enough for the one-page prompt; the numbers steady from about 300. A full profile works best on three or more pieces totalling 3,000 words, because it also keeps the samples for the model to anchor on.
Is a persona paragraph useless?
Not useless: it sets the intent. It just cannot carry the habits that make a voice yours, because you cannot see most of them. Put the persona in one line and the measurements underneath it.
Does the prompt builder store my writing?
No. The text is measured once (one call for the judged attributes), the numbers come back, and the text is discarded. The fingerprint that comes with it keeps scores only.