# What Is an AI Style Guide? · ScriptGrain

> The instructions that make a model write in a specific voice, why adjective versions drift, and what a measured alternative looks like.

Canonical: https://scriptgrain.com/blog/what-is-an-ai-style-guide

# What Is an AI Style Guide?

*By Jack Stovell · 2026-09-16 · Guides*

Here's the thing: everyone's writing one now, and almost nobody's calling it that. You know the document. "Write like us, but friendlier." "Confident, not arrogant." "Sound human." That's an AI style guide, whether you've named it or not, and it's usually a mess.

An AI style guide, defined plainly: it's the set of instructions you give a model so it writes in a specific voice, rather than the flat default it reaches for otherwise. Today that mostly means a hand-written prompt full of adjectives and a few pasted examples. It works, sort of, for about a paragraph. Then it drifts.

Drift is the failure mode worth naming early, because it's the whole problem. You write "warm but professional" into a prompt. The model reads that and produces something. So does a different model. So does the same model tomorrow, with slightly different phrasing in your request. Three outputs, three interpretations, none of them wrong exactly, none of them you either.

What people currently put in one

Open most AI style guides floating around a company's shared drive and you'll find the same ingredients. A list of adjectives: clear, bold, approachable, occasionally "punchy" if someone in marketing had a good week. A handful of do's and don'ts. Maybe three paragraphs of "good" copy pasted in as reference, with a note saying "match this tone." Sometimes a banned-words list, because someone got burned by "delve" once and never recovered.

None of this is stupid. It's just underspecified. And underspecification is the actual disease here, not laziness, not bad intentions.

Why adjectives don't hold up

Take "confident." Ask two writers what confident means and you'll get two different pieces of writing, both defensible, neither identical. One writer hears confident and starts cutting hedges, short sentences, full stops doing the heavy lifting. The other hears confident and reaches for longer, assertive sentences stacked with subordinate clauses, because to them that's what authority sounds like. Same word. Two readings. Neither writer is misreading the brief; the brief just never had enough information in it to be read only one way.

Now hand that same adjective to a model. It doesn't have your shared history, your inside jokes about the last three drafts, the unspoken sense of "we tried that once and it read try-hard." It has the word "confident" and whatever training data taught it that word means. So it picks a reading. Could be either one. Could be neither. That's interpretation drift, and it's not a bug you can prompt your way out of, because the input was never precise to begin with.

Fair enough, you might say, so tighten the adjectives. Add more examples. That helps, a bit. But you're still describing style with words that don't have fixed values. "Punchy" isn't a number. "Warm" doesn't have units. At some point you need something that does.

The measured alternative

This is where a style guide stops being a mood board and starts being data. Instead of "confident," you measure confidence-versus-hedging as an actual ratio across your existing writing. Instead of "punchy," you measure average sentence length and how much that length varies, sentence to sentence, because punchy writing tends to swing wildly between a five-word line and a twenty-five-word one. Contraction rate. Comma density. How often you open a paragraph with "so." Whether you reach for semicolons or avoid them like a personal insult. Forty-five separate attributes, each with a value, each pulled from writing you actually produced rather than adjectives you guessed at afterwards.

That measured profile is exportable. It becomes a tone-of-voice prompt you can drop straight into Claude, ChatGPT or Notion AI, and because it's specified as attributes with values rather than adjectives with vibes, two different models reading it land in roughly the same place. That's the whole point. Not a nicer-sounding brief. A brief that can't be misread three different ways by three different sittings.

This is what scriptgrain.com does with its free profile tool: one analysis, no card needed, and a browser-based style check at /tools/writing-style-analysis if you just want to see your own numbers before committing to anything. Paid plans start at £12 a month if you want the full export and ongoing re-measurement.

Keeping it current

Here's the bit people forget. Your writing changes. You get better, you get lazier on deadline days, you pick up habits from whatever you've been reading lately. A style guide written in January and never touched again is measuring who you were, not who you are.

The fix is boring and correct: re-measure periodically. Feed in recent samples, check whether the attributes have shifted, update the exported prompt if they have. It's maintenance, not ceremony.

So, what is an AI style guide? At the end of the day, it's not a vibe document and it was never supposed to be one. It's a specification, ideally a measured one, that travels between tools without losing your voice in translation. Adjectives get you a guess. Numbers get you you.

[More from the ScriptGrain Journal](https://scriptgrain.com/blog)