Micro-habit detection: your voice is not your vocabulary

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

Ask people what makes their writing theirs and they name words. The ones they like, the ones they avoid. Vocabulary is real, and it is the least of it. Stylometry, the discipline that attributes anonymous texts to authors, barely looks at content words. It looks at function words, punctuation habits, sentence-length distribution, the shape of clauses: things a writer produces without deciding to, and could not fake if asked.

Those are micro-habits. They are why a friend can tell your email from a colleague's before reading the signature, and why an AI draft "in your voice" still is not: the model matched your words and none of your habits.

Measure twelve of yours. Paste something you wrote below. Each habit is counted in your browser and placed against a reference corpus of recent public writing, so you see which ones are typical, which are distinct, and which are yours alone.

What a micro-habit looks like in numbers

Contractions, per 1,000 words. The reference corpus runs from 8 (formal) to 32 (spoken) between the tenth and ninetieth percentile. A writer at 120 is unmistakable in one paragraph.
Sentence variation, standard deviation over mean. Human prose usually sits above 0.5; a machine's average sits lower, which is why detectors score it first.
Short sentences, the share under eight words. Some writers never write one. Some write one in three.
Asides in brackets, per 300 words. A habit of thinking that survives every editor.
Semicolons, exclamation marks, ellipses. Each one is a yes or a no for most writers, which makes each one a fingerprint.
Who the writing addresses. The share of "you" against "I" and "we". Teachers and marketers write to you; diarists and analysts do not.
How you open, how you argue, whether you hedge. These need a judge rather than a count; the fingerprint adds them.

Why this matters for AI

A prompt can carry adjectives. It cannot carry habits you do not know you have. That is the whole reason "write like me" produces something generic: the model got the vocabulary and guessed the rest, and its guess is the average. Measure the habits and you can hand them over as numbers, then score every draft against them. ScriptGrain's profile does that with 45 attributes; the report below does the countable dozen with no account.

Where word choice does come in

Preferred words, signature phrases and the words a writer never uses are measured too, and they matter in a draft. They are just not what identifies you. A draft can use all your favourite words and still read as somebody else if the rhythm is wrong. Start with the habits.

Five micro-habits worth checking in your own writing

Start with the ones you can count in a minute. Here's the short list.

Contraction rate. Read two lines aloud. "It is not going to work" and "It's not going to work" say the same thing, yet a writer who always picks the first has a different voice from one who always picks the second. The habit holds steady across topics, which is why it's so useful.
Comma density. Some writers comma every clause. Others let a long sentence run clean to the end. Neither is wrong; each is consistent.
Paragraph openers. Look at the first word of each paragraph. Do you start with "So"? With "You"? With a noun and a fact? The pattern repeats more than you'd guess.
Brackets. An aside in parentheses is a little thought that escaped. Writers who use them use them everywhere; writers who don't, never do.
Pronoun mix. Count "you" against "I" and "we". It tells you who the writing is talking to.

Semicolons and dashes belong on the list too. Most people either reach for them or don't. Few sit in the middle.

None of this is a choice. You made these calls before you made a single word choice.

How micro-habits get lost in a rewrite

Ask a model to tidy your draft and it tidies. That's the problem. Tidying means moving toward the average, and the average has no habits.

Take a short example. You write: "We shipped it. Late, sure, but it works." A rewrite comes back: "We delivered the project, which, although delayed, functions as intended." Same content. Different writer. The contractions are gone and the fragment has gone with them. The sentence has also grown a comma pair you never asked for.

That's the pattern. Rewrites lengthen short sentences and smooth out the rhythm, which is why sentence variation drops first. They expand contractions because full forms read as "correct". They swap your opener for a gentler one. They drop the bracket, because a bracket looks like an edit note, and they quietly trade "you" for something more neutral.

To be fair, the model isn't being careless. It's doing what it was trained to do: produce plausible, polished prose. Polish is the enemy of a fingerprint.

And you won't notice in the moment. Read the rewrite once and it seems fine, even better. The loss only shows when you put the numbers side by side. So measure before and after, not just by eye.

Reading the per-feature notes

Every draft scores between 0 and 1 against your profile. On-voice drafts typically land between 0.85 and 0.95. Partial matches sit from 0.5 to 0.8, and a clearly different voice falls under 0.45. That's the headline. The useful part is underneath.

Each feature comes back with a delta: the draft's value against yours. Say the profile shows a high contraction rate and the draft shows half of it. That's one line to fix, and you know which way to push. Short sentences too rare? Break up two of the long ones. Pronoun mix tilted toward "we"? Check whether you meant to address the reader at all.

Read the deltas in this order:

Start with the biggest gap, not the first one listed.
Fix habits that cluster together, since contractions and sentence length often move as a pair.
Re-score after each change. One edit can fix two features or break a third.

One caution. A small delta isn't a failure. No writer hits their own numbers on every draft, and a score of 0.9 with two soft misses is a good result. Chasing 1.0 is how you end up writing like a spreadsheet.

So treat the notes as a map, not a verdict. They tell you where the draft drifted and by how much. What you do about it is still your call.

Questions

How accurate is a report from one piece of writing?

Indicative from 120 words, steady from about 300, and better across several pieces. A single piece written for an unusual purpose (a legal notice, a eulogy) measures that purpose, not you.

What is the reference corpus?

Two hundred and ninety-nine recent posts from forty-seven public blogs, newsletters and essay sites, measured with the same code and discarded. Percentiles say where you sit against that set; they are not a judgement.

Can an AI reproduce these habits if I give it the numbers?

Largely, yes, and that is the point: a number is a target a model can hit, an adjective is not. The remaining gap is what the voice-match score is for.