How to preserve your voice when using AI

By Jack Stovell X · Instagram · LinkedIn · published 2026-09-20 · updated 2026-09-30

You wrote the brief. The assistant wrote the draft. It is fine, and it is not you. The sentences are a little longer, the hedges are back, the contractions are gone, and a "furthermore" has appeared from nowhere. Nothing is wrong enough to point at, so it ships, and the next one drifts a little further.

Preserving a voice is not a matter of asking nicely. Models regress to the middle of their training data unless something holds them to a target. Here is what holds.

Check the drift first. Paste something you wrote yourself and the AI draft below. The draft is measured feature by feature against your piece, so you get the number and the list of what moved rather than a feeling.

1. Know which habits are yours

A voice is mostly habits you cannot see: how often you contract, how many asides you put in brackets, whether you open with a question, how short your shortest sentences get. Ask an assistant to "keep my voice" and it keeps the adjectives you might use to describe yourself, which is not the same thing. Measure the piece you are proudest of and you have a target: average sentence 11 words, 21 contractions per 1,000, no exclamation marks, dry humour. That is a brief a model can follow.

2. Give the assistant numbers, not adjectives

"Conversational but authoritative" produces the same draft for everyone. "Average sentence 12 words, one sentence under eight words in every paragraph, contractions everywhere, no em dashes, conclusion first" produces yours. The prompt builder on the ChatGPT page writes that prompt from a sample of your writing. A full profile does it with 45 attributes and keeps your samples so the model has real sentences to anchor on.

3. Score before you ship

The step most people skip. A voice-match score compares the draft to the measured profile and names the features that moved: "formality 7.5 against your 4.8", "contraction rate 0 against your 21". Fix those two and the draft is back. Without the score you edit by instinct, which is slower and less reliable than a number.

4. Keep the edits

Every edit you make to a draft is information: this word, never; this rhythm, always. ScriptGrain reads the difference between the draft and what you finally published and proposes changes to the profile, which you accept or reject. Six months in, the profile knows things about your voice that you would never have thought to write down.

5. Watch the published work, not only the drafts

Voice drifts in public too: the pricing page a contractor rewrote, the help centre a new hire filled in. The monthly voice monitor reads the pages you choose and scores them against the profile, so drift on your own site becomes a number before a reader notices.

Start with the check below. If the draft reads under 0.55, the AI did not keep your voice; the list underneath says exactly where it let go.

Signs your voice is slipping

Drift rarely announces itself. Nobody reads a draft and thinks, "that isn't me." They think it reads fine, and publish it. So you need tells, small and checkable.

Here's the short list:

  1. Your contractions thin out. "It is" creeps in where you'd always have written "it's".
  2. Your shortest sentences get longer. The five-word punch turns into a twelve-word explanation.
  3. Hedges return: "can often help", "may suggest", "in many cases".
  4. Stock transitions appear that you'd never say aloud.

None of these is fatal alone. Together they're a pattern. Read the last three things you published, then the last three drafts you only lightly edited. If the second pile sounds calmer, politer and a bit more like everyone else, that's drift. Another sign: a colleague says a piece "sounds a bit off" and can't say why. That's your readers noticing the middle of the training data.

A monthly voice check

Once a month is enough. Put it in the diary and keep it boring.

Pick a few recent pieces, ideally a mix of drafted-with-AI and written-by-hand. Score each against your profile. Every draft comes back with a score from 0 to 1, and the deltas tell you which features moved: draft value against profile value. On-voice drafts typically land between 0.85 and 0.95. Partial matches sit from 0.5 to 0.8. Under 0.45 is a clearly different voice.

Then read the notes, not just the number. A score of 0.78 with one big delta on contraction rate is a five-minute fix. A score of 0.78 with six small deltas is a different problem, and you want to know which one you've got.

Now the real question. Did your style move on purpose? Maybe you've started opening with questions, or your sentences have tightened since you took on a new editor. Fair enough. If the change is deliberate, refresh the profile with your newer samples so the target follows you. If it isn't, leave the profile alone and fix the drafts. Mixing those two up is the classic mistake: you either freeze your voice in amber or bless every accident.

Preserving one voice across a team

Teams drift faster than individuals. Five people, five prompts, five slightly different versions of "our tone". Nobody's wrong. The page just sounds like a committee.

The fix is one shared profile, read live by every tool. It reaches ChatGPT, Claude, Gemini, Cursor, Meta Muse and any MCP client through the MCP server and the REST API, and it's never a pasted prompt that someone can quietly reword. Everyone's assistant starts from the same measured target, so the brief no longer depends on who wrote it.

If you're weighing this against a lighter tool, we've set the two side by side. StyleMimic counts a handful of style features in the browser and scores drafts against your samples, which suits quick jobs. Our StyleMimic comparison shows what changes when 45 attributes sit behind the score, how we publish the method with weights and tolerances, and why the same voice then travels into Claude, ChatGPT and other tools.

Then use the score as a shared bar. Agree a number, say 0.85, and treat anything below it as a draft that needs polish, not a draft that needs a debate. Polish revises toward a target score (0.9 by default), which takes the argument out of "does this sound like us?" and puts a measurement in its place.

Two practical points. Build the profile from published work the team is proud of, three or more pieces and around 3,000 words, rather than from a style guide nobody reads. And run the monthly check on the team's output as well as your own. A sample from the newest hire tells you more about drift than another from the founder.

At the end of the day, a shared voice is a habit with a number attached. The number just makes the habit hard to ignore.

Questions

Why does AI writing drift away from my voice even when I give it samples?

Samples in a prompt shape the first paragraph and then fade: the model returns to its statistical middle as the draft goes on. A measured target (numbers for rhythm, punctuation, register) holds for the whole piece, and a score afterwards catches whatever slipped.

What score counts as 'still my voice'?

On the free check, 0.8 and above reads as the same writer, 0.55 to 0.8 is drifting, under 0.55 is a different voice. On a full profile the scale is the same and the comparison is against the measured profile, where 27 of the 45 attributes carry a weight, rather than one piece.

Can I preserve my voice inside ChatGPT or Claude rather than switching tools?

Yes. That is the point of the connector: the assistant you already use writes from the measured profile and scores its own drafts. The prompt builder is the no-account version for custom instructions.

Related