Style guide generator

By Jack Stovell X · Instagram · LinkedIn · published 2026-08-05 · updated 2026-09-30

Most style guides don't change how anyone writes. This one does, because it's built from measurement, not a workshop.

Build a one-page style guide from your own writing, free. Paste 120 words or more that you wrote yourself. The generator below counts your sentences, paragraphs, punctuation, contractions, paragraph openers and the words you lean on, sets each number against 299 recent public posts, and writes a rule from it. It adds the AI tells it found and a short block of rules for an AI writer, and you can copy the whole guide as Markdown. It runs in your browser, so nothing is uploaded.

Templates describe a personality. Nobody actually has one.

"Friendly but authoritative." "Bold, human, clear." Fine words. Utterly useless to a writer at 4pm trying to decide whether that sentence needs a comma. Template style guides give you adjectives. Adjectives don't enforce anything. That's why the guide gets written once, praised in a meeting, and ignored by everyone within a fortnight.

Most teams are already writing one of these, and almost none of them call it a style guide. It's the pile of adjectives you hand a model, hoping it writes like you instead of like everyone else. We define an AI style guide plainly, along with why that pile usually fails and what the model needs from you instead.

We measure instead of describing

Upload real writing, yours or your client's. ScriptGrain runs a two-pass stylometric analysis across 45 attributes: sentence rhythm, contraction rate, comma density, hedging versus confidence, the lot. Then it writes a narrative explaining what those numbers mean in practice. Not vibes. Numbers, with a narrative attached.

What you actually get

A working voice profile, not a PDF nobody opens. It states exact contraction frequency, sentence length variance, punctuation habits, the works. Every attribute is a constraint your writers (human or AI) can actually follow, because it tells them precisely what to do, not what to feel.

Every draft gets scored against the profile

Generate something, and you get a voice-match score straight away. No guessing whether a freelancer nailed the brief or missed it by a mile. Either the numbers line up or they don't.

Built for teams that hand voice to other people

Onboarding a new writer? Give them the profile, not a 20-page deck they'll skim once. Running an agency? Hand a client's voice to freelancers without three rounds of "make it sound more like us" emails. That's the whole point: the standard is shared, and it still sounds like you when someone else is writing.

Pricing, plainly

Free: one voice profile, one full analysis, no card needed.

1. Writer, $17/month: 1 profile
2. Operator, $40/month: 3 profiles
3. Studio, $134/month: 10 profiles
4. Agency, $405/month: unlimited profiles, 10 team seats, per-member credit caps

Measuring is free on every tier, and so is your first draft. Regular generation is what you pay for.

Start with the free analysis

You don't need a card, and you don't need to trust our copy about what this does. Upload a sample, get your 45 attributes and your narrative, see if it actually sounds like you. It either will or it won't. That's the test.

Fair enough if you're skeptical. Every template style guide promised the same thing once. This one just shows its working.

What goes in a measured style guide

Start with the numbers. A measured guide should open with the ones a writer can check against a sentence: average sentence length and how far it swings, contraction rate, comma density, how often you reach for a semicolon or a dash. The generator on this page covers those, plus paragraph openers and the words you lean on. Keep all of it.

Then the rules. Every number needs a rule beside it, written as an instruction. "Average sentence 17 words, wide variance" becomes "Follow a long sentence with a fragment." Short. Checkable. Nobody has to interpret it.

Add the things to avoid. A good guide lists the AI tells it found in your writing, because those are the habits that creep in when a draft gets polished too smoothly. And if an AI writer will touch your copy, it needs its own short block of rules, separate from the human ones.

What stays out? Adjectives. Mission statements. Anything a writer can't test.

Finally, give it a sample. One real paragraph, in your actual voice, sitting at the bottom. So a new writer sees the numbers turned into prose. Fair enough, it won't cover every case. It doesn't need to. The full 45-attribute profile exists for that.

Rolling the guide out to a team

A guide lands better when people can see where it came from. So don't send a PDF with a cheery note. Send the guide alongside the writing it was measured from, and let people check the claim themselves.

Start small. Pick one writer, one channel, one month. Ask them to draft against the guide and mark every rule that felt wrong or unclear. Rules that survive that test go to everyone. Rules that don't get rewritten.

Freelancers need a different handoff. They don't want your history; they want a standard they can hit first time. Give them the one-page guide, the sample paragraph and the AI-tell list, then tell them plainly that drafts get scored against the numbers. That removes the vague "make it sound more like us" round. They know what the target is before they type a word.

If someone on the team uses an AI writer, the profile shouldn't travel as a pasted prompt that goes stale. ScriptGrain's MCP server and REST API pass it to ChatGPT, Claude, Gemini, Cursor, Meta Muse and any MCP client, read live each time. Both are on every plan, Free included.

Keeping the guide honest over time

Guides drift. Not because anyone's lazy; because voices move and nobody re-measures. So build a check into the routine.

Every draft gets a voice-match score from 0 to 1. On-voice work typically lands between 0.85 and 0.95. Partial matches sit around 0.5 to 0.8. Anything under 0.45 is clearly a different voice. Those bands give you a simple rule: a draft under the bar goes back, a draft over it ships.

The score arrives with per-feature deltas, draft value against profile value. That's the useful part. A single number tells you something's off; the deltas tell you what.

1. Sentences averaging far longer than the profile says.
2. Contraction rate sitting well below the target.
3. Hedging creeping in where the profile is direct.

Polish can revise a draft toward a target score (the default is 0.9), and the app learns from the edits a user saves.

And re-measure now and then. Take your latest three published pieces, roughly 3,000 words in all, and run them again. If the numbers have moved and the work still feels right, update the guide. The method behind the scoring is published on the voice measurement framework page, so nobody has to take the number on trust.