Preferred words analyser

By Jack Stovell · published 2026-09-24

Preferred words are the recurring, distinctive words a writer keeps reaching for, the ones that make a piece sound like them. In a ScriptGrain voice profile the extraction model stores them as a list of up to 5 words or phrases per analysis, quoted verbatim from your samples, and you can hand-edit the list on the Profile page.

Analyse your preferred words free

Your free voice profile measures preferred words among all 45 attributes, from a sample of your own writing. No card needed.

Try the six-signal check first

At a glance

What preferred words measures

Preferred words measures which words you use often enough, and distinctively enough, that they've become part of your signature. Not "the" or "and", obviously. Words like "genuinely" or "distinctive" if that's your habit, or a two-word phrase you drop into every third piece without noticing. The app's own definition is blunt: "the most-recurring distinctive words across your samples." Nothing more mystical than that.

Here's the mechanism. The extraction model (Claude Sonnet 5, running two passes: one per sample, one to synthesise across all of them) reads your writing and pulls out the words verbatim, spelling and all. It doesn't correct you: if you spell a word your own way, that spelling is what lands in the profile. This attribute is one of the 45 voice attributes and sits in the Lexical group alongside five others, including vocabulary diversity, word length distribution and rare word rate; the lexical layer reference covers the group as a whole.

The free writing style analyser does something more mechanical: it lists the 8 most frequent words over 3 letters that aren't on its function-word list, with counts. That's frequency, a proxy for distinctiveness: a word can also be frequent because of the topic.

How preferred words is scored

Preferred words does move your voice match score, with weight 2. It doesn't sit alone, though. It's scored through a shared "signature phrases" feature: the pool draws from preferred_words, discourse markers, filler phrases and preferred phrasings, keeps anything multi-word or any single word of 6+ letters, caps the pool at 15 items, and needs at least 3 to work at all.

When a draft is scored, code counts how many pool entries appear in it. Similarity is hits divided by the smaller of 4 and the pool size, capped at 1. So if your pool has 4 or more items, a draft that uses any 4 of them maxes out that component. Miss most of them and the score drags. This is one of 28 attributes (out of 45) that can shift the number; the other 17 never move it. Full mechanics live on the voice measurement framework page, and the voice match glossary entry defines the wider 0 to 1 score.

How to analyse preferred words in your own writing

You can find your preferred words by hand with a few pieces of your writing and a highlighter, metaphorically speaking.

  1. Pull together three or more pieces you've written, at least a few hundred words each.
  2. Strip out function words: the, and, of, to, a, in, that, it, for, with, and so on.
  3. Count what's left. Anything appearing more than a handful of times across separate pieces is a candidate.
  4. Check it's actually yours, not just topic-driven. "Mortgage" showing up a lot in a mortgage article doesn't count. "Genuinely" showing up everywhere you write, does.
  5. Shortlist five at most, the same cap a ScriptGrain analysis uses.

That's the manual version. ScriptGrain's free voice profile reads this straight from a writing sample, as one of all 45 attributes.

Examples of preferred words in real writing

Example: a writer who always says "here's the thing" before a pivot, in almost every longer piece, has a preferred phrase, not a coincidence.

Example: someone who reaches for "distinctive" and "genuinely" across unrelated pieces (blog posts, emails, LinkedIn updates) is showing a lexical fingerprint rather than a topic echo.

Example: a marketer who consistently writes "the short version" before a summary has a preferred phrase doing double duty as a structural signal too.

How to make AI use your words

Getting a model to sound like you starts with telling it, explicitly, which words are yours.

  1. List your recurring words and phrases somewhere the model can see them.
  2. Tell it to use them where they genuinely fit, never forced into a sentence that doesn't want them.
  3. In a ScriptGrain profile, add the exact phrasings you want favoured under Preferred phrasings on the Profile page.
  4. Feed the model two or three real pieces of yours so it sees the words sitting in actual context, not floating in a list.

When a ScriptGrain profile generates text, preferred_words gets sent inside the voice profile JSON block of the system prompt, under an instruction to treat every attribute as a hard constraint, not a suggestion, and the generator's bans on stock AI phrases are lifted for anything on the list. Preferred phrasings render as hard constraints above that block. The profile reaches ChatGPT, Claude and other tools through the API and MCP server, so the list travels with you to wherever you draft.

How to stop AI overusing the same words

Stopping a model's favourite crutch words means banning them outright, not asking nicely.

  1. Name the specific words the model keeps reaching for. Be exact, not vague.
  2. In a ScriptGrain profile, put them under Banned terms on the Profile page; a banned term overrides the voice profile, even when the same word sits in preferred_words.
  3. Ask for one editing pass that swaps each repeated pet word for a plainer alternative.

To be fair, this is the same lever in reverse. Preferred phrasings and Banned terms both render as hard constraints above the voice profile JSON: one says what to favour, the other what to drop.

Questions

How do I find my most used words in writing?

Gather several pieces, strip out function words (the, and, of, a, to), then count what's left. Anything recurring across separate, unrelated pieces, not just one topic, is a candidate. ScriptGrain's free analyser does this automatically: it lists the 8 most frequent words over 3 letters, excluding function words, with counts, from one sample.

What words do I overuse in my writing?

Overused words are usually the ones you don't notice: fillers, hedges, or a favourite adjective that turns up regardless of topic. The fix is the same as finding preferred words, just read differently. If a word shows up often and isn't doing real work in the sentence, that's your overused word, not your distinctive one.

How is preferred word use scored?

It's scored with weight 2 through the shared signature phrases feature, pooling preferred_words with discourse markers, filler phrases and preferred phrasings. The pool keeps multi-word phrases and single words of 6+ letters, caps at 15, and needs 3 or more to function. Similarity is hits divided by the smaller of 4 or pool size, capped at 1.

How do I make ChatGPT use my vocabulary?

List your recurring words, tell the model to use them where they fit naturally, and feed it real samples so it sees them in context. In ScriptGrain, preferred_words sits in the generator's system prompt as a hard constraint, and ChatGPT can read the same list through the API or MCP server instead of a manual prompt each time.

How do I stop ChatGPT using the same words repeatedly?

Name the specific repeat offenders and ban them explicitly rather than asking the model to "vary its language". In a ScriptGrain profile, Banned terms override the voice profile, including preferred_words. Follow with one edit pass asking for plainer swaps, and check the result by counting your words again.

Related

Sources