Consistent Writing Voice With AI: A Style Brief Method

You keep a consistent writing voice with AI by turning your own published posts into a short style brief, then reusing that brief as constraints on every draft instead of retyping a fresh prompt each time. The brief is the system. The prompt is just the delivery.

Why do AI drafts stop sounding like you?

AI drafts drift generic because the model has no memory of your published work. It has your instructions for this session and nothing else. Ask for a post about pricing and you get a competent post about pricing, written in the average of every pricing post ever published.

The flattening is predictable. It shows up in four places.

Rhythm. Model output tends toward even, mid-length sentences. Your writing probably isn’t even. You may open with a three-word fragment and follow it with a long, comma-heavy sentence that carries the argument. That contrast is part of how readers recognise you.

Hedges. Real writers hedge specifically: “I’ve run this twice and got the same result,” or “this might not hold for enterprise buyers.” Models hedge generically: “results may vary depending on context.” The second version is true and useless.

Specific references. You mention the client who asked a strange question, the tool you abandoned, the number from your own dashboard. Models substitute abstractions unless you force the specifics in.

Opinion. You take positions. Models summarise positions. A post that surveys five views without picking one reads like a briefing, not a blog.

Our own measurement on 2026-09-13, scoring 23 hand-written DeAIze guides and 12 unedited model passages with the DeAIze detector (local ensemble, lite tier, 30% flag threshold), put the human sample at a mean AI score of 21% and the unedited model output at 33%. The gap is real but modest — and the more useful finding was the tell frequency: 4 per 100 words in the human sample against 1 per 100 words in the model output. Human writing carries more of the small irregularities that make a voice identifiable. That is what a style brief is trying to reproduce.

How do you build a style brief from your own best posts?

Pick five to eight posts that did well and that you would still defend. Not your most recent — your most you. Then read them with a pen and extract four things.

1. Sentence length pattern. Count words in the first ten sentences of each post. Write down the shortest and the longest. If your range is 4 to 41, that range is a constraint, not a coincidence.
2. Vocabulary. List the words and phrases you actually repeat. Not your topic vocabulary — your habit vocabulary. The transitions you favour, the way you open a paragraph, the way you close one.
3. Recurring moves. Do you always start with a concrete scene? Do you end sections with a one-line verdict? Do you interrupt yourself with a parenthetical? These moves are your signature more than any word list.
4. Positions and hedges. Note where you commit and where you deliberately leave room. “I think X is overrated” is a different move from “X has trade-offs.”

Write it up in under 400 words. A brief that runs long will not survive contact with your own workflow. Here is the shape:

VOICE BRIEF — [your name]
Rhythm: mix of very short (3-6 word) and long (30+ word) sentences. Never three mid-length sentences in a row.
Openings: concrete scene or a blunt claim. Never a definition.
Vocabulary I use: "honestly", "the catch is", "here's what I'd actually do"
Vocabulary I avoid: "leverage", "seamless", "robust", "in today's landscape"
Moves: end each section with a one-line verdict. One parenthetical per 300 words max.
Opinion: take a position in every section. Hedge only when I'd hedge out loud.
References: one specific example per section — a client, a tool, a number from my own work.

That document is the asset. Everything below is about using it.

What should the brief actually constrain?

A brief that only lists vocabulary does very little. The constraints that change output are structural.

Constraint type	Weak version	Strong version	What it fixes
Rhythm	“Write naturally”	“Mix 4-word and 35-word sentences; no three even sentences in a row”	Flat, uniform pacing
Openings	“Be engaging”	“Open with a scene or a blunt claim, never a definition”	Generic intros
Opinion	“Be helpful”	“Take one clear position per section”	Survey-style posts
Specifics	“Add examples”	“One named example per section from my own work”	Abstract filler
Hedging	“Be honest”	“Hedge only where I’d hedge out loud”	Vague qualifiers

Notice that every strong version is checkable. You can read a draft and ask: did this open with a definition? Are there three even sentences in a row? If the constraint can’t be checked, it won’t be followed.

If you’re going to run this loop every week, you need a way to see which sentences drifted. DeAIze scores detection sentence by sentence, so you can spot the lines that read generic and rewrite only those — the free tier gives you 200 words on signup, no card. See pricing if you’re publishing at volume.

What does the same post look like with and without the brief?

Here is an illustration, not a real post. Same topic, same argument, two drafts.

Without the brief. “Content consistency is important for audience retention. When readers know what to expect, they are more likely to return. Establishing a clear tone across your posts helps build trust over time.”

Three sentences, all mid-length, no position, no specific, one hedge. It is not wrong. It is also not anyone’s writing.

With the brief. “Readers don’t come back for your topics. They come back for you. I learned this the slow way — I published eleven posts in a row that could have been written by anyone with a keyword list, and my reply rate dropped to almost nothing.”

Short sentence, long sentence, a specific count from the writer’s own experience, and a position. Same argument, different author.

The second draft will still trip a detector sometimes. That is the honest part. Detector scores are probability estimates, not proof of authorship, and no brief guarantees a particular reading on GPTZero or anywhere else. What the brief does is make the draft recognisably yours to a human reader, which is the thing that actually affects whether people keep subscribing.

How do you feed the brief back as constraints, not a one-off prompt?

The mistake is pasting the brief into a chat once and expecting the model to remember it next week. It won’t. Constraints have to travel with every request.

Three ways to do that, roughly in order of effort:

1. Paste the brief above the task every time. Crude, works, costs you thirty seconds.
2. Save it as a project instruction or custom instruction. Better, because the model applies it before you ask.
3. Keep the brief as a file and reference it in your template. Best if you publish often, because the file becomes the single source of truth and you stop maintaining two versions.

Whichever you pick, add one instruction that matters more than the rest: flag anywhere you had to guess at my opinion, rather than inventing one. You want the model to leave gaps for you, not fill them with plausible-sounding positions you don’t hold.

Then run your own pass. Read the draft aloud. Where you stumble, the rhythm is wrong. Where you’d never say the sentence out loud, cut it. Where there’s no position, add one. This is the same work described in our guide to editing AI drafts, applied to voice rather than to flags.

When should you update the brief?

Voices drift, and the brief should drift with them — slowly, on purpose, not every week.

Update it when one of these happens:

• You’ve published ten posts since the last revision. Read the most recent three against the brief. If they’d fail it, either the brief is stale or you’ve drifted. Decide which.
• A reader comments on your tone. Reader perception is the signal that matters. If someone says you sound different, you probably do.
• You change format. Moving from long essays to a weekly newsletter changes rhythm. The brief has to reflect the format you’re actually publishing in.
• A phrase in the brief starts feeling like a tic. Habits become crutches. Retire the phrase, keep the move.

Do not update the brief to chase a detector score. The brief is about voice, not about passing anything. If you’re getting flagged on writing you did yourself, the problem is elsewhere — false positives on your own work is a different problem with a different fix.

One more thing worth saying plainly. The ideas, the reporting, the data and the conclusions in your posts have to be yours. A brief keeps your voice on AI-assisted drafts of your own thinking. It does not make someone else’s work yours, and it shouldn’t be used that way.

The short version

Build the brief from five to eight posts you’d still defend. Extract rhythm, vocabulary, moves and positions. Write it in under 400 words. Constrain structure, not just word choice, because structure is what readers actually notice. Paste it into every request. Read the draft aloud and fix what you stumble on. Revisit the brief every ten posts or when a reader tells you that you sound different.

That routine takes an afternoon to set up and about ten minutes per post afterward. It is the difference between using AI to write faster and using AI to write as someone else.

---

Measurement note: figures in this article come from our own detector run on 2026-09-13 — n = 23 hand-written guides versus n = 12 unedited model outputs, median AI score 22% and 34% respectively. Method: /methodology/.