Structural signatures in writing

By Jack Stovell · published 2026-09-24

Structural signatures (structural_signatures) are the recurring shapes a writer's pieces fall into, such as a one-line opener or a habit of closing on a question. ScriptGrain's extraction model records them as a short free-text description of at most 15 words, in the Quirks and cadence group. Nothing counts them, and they never move the voice match score.

Analyse your structural signatures free

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

Try the six-signal check first

At a glance

What structural signatures measures

This attribute captures the patterns you'd notice if you read ten of someone's posts back to back. The app's own definition is blunt: "Notable recurring structural patterns." That's it. No formula, no scale from one to ten. Just a description of 15 words or fewer, written by the extraction model (Claude Sonnet 5) after reading your samples in two passes, one per sample and one to synthesise the profile. It is one of the habits that make up a stylometric fingerprint.

This isn't about sentence length or comma habits. Those live elsewhere among the 45 voice attributes, in ones that code counts when a draft is scored. Structural signatures is about shape at a higher level: does someone always open with a fragment? Do they close paragraphs with a rhetorical question? Do they build to a payoff or scatter it throughout? It sits in the Quirks and cadence group alongside cousins like rhythm pattern and paragraph length, and it overlaps with where a writer chooses to drop their strongest line, which is its own attribute: power sentence position.

To be fair, "structural" sounds like it should be mechanical. It isn't. It's judged.

How structural signatures is scored

It isn't. The voice-match engine has no comparator for this attribute, so it never moves a voice match score and polish cannot target it. Voice match does judge a draft's argument structure, but that is a separate attribute.

So why record it at all? Because ScriptGrain sends every attribute to the model when it writes a draft, and a one-line description of your shapes is something a model can act on. It sits outside the 28 attributes that move the number; the voice measurement framework page covers how those 28 work.

How to analyse structural signatures in your own writing

You can do this by hand, no tools required. It just takes a stack of your own work.

  1. Pull five to ten pieces you've written, ideally spanning a few months.
  2. Read only the first sentence of every paragraph. Write down what each one does: states a fact, poses a question, drops a fragment.
  3. Read only the last sentence of every section. Same exercise.
  4. Look for repeats. A pattern that shows up again and again across different pieces is a signature, not a coincidence.
  5. Name it in one line of 15 words or fewer, the way you'd describe a friend's habit of speech.

That's the manual version. ScriptGrain's free voice profile reads structural signatures from a writing sample as one of all 45 attributes it extracts.

Examples of structural signatures in real writing

Example: "Here's the thing: most advice on this is wrong." An opener that names the topic and immediately undercuts it.

Example: "So what do you do about it? Three things." A closing question followed by a numbered payoff, the build-to-point shape in miniature.

Example: "It rained all week. Then Friday." A full sentence followed by a two-word fragment, used as a rhythmic full stop.

None of these map to a typical value. ScriptGrain's reference corpus has no figure for this attribute, so nothing here is an average; they're illustrations of what a signature looks like on the page. The stored value describes patterns like these in one sentence, for example "Opens on a one-line claim and closes each piece with a question."

How to make AI keep your structure

If you want an AI assistant to hold onto your shapes rather than defaulting to its own, you need to hand it the pattern, not just ask for it.

  1. Describe your recurring shapes directly: a one-line opener, a numbered payoff, a question to close. Name them the way you'd name a habit.
  2. Show one worked example alongside the description.
  3. Give it full pieces as samples, not fragments. Structure only becomes visible across a whole piece, opener to close.
  4. Ask it explicitly to keep your section order.

When ScriptGrain writes a draft, this attribute travels inside the voice profile JSON block sent to the model, under "Treat every attribute as a hard constraint, not a suggestion." That profile reaches ChatGPT, Claude and other tools through the API and the MCP server, so each of them can read the same description.

How to stop AI using the same template

Stop the sameness by banning the specific tics that cause it, rather than vaguely asking for "more variety".

  1. Ban triadic parallel lists outright: three items in matching grammar, one after another.
  2. Ban 'not X but Y' balancing lines. They sound decisive and say nothing.
  3. Ask for no headings on short pieces.
  4. Vary paragraph shapes deliberately, a short one here and a longer one built around a question there, rather than repeating a single rhythm down the page.

ScriptGrain's own generation bans "not X but Y" balancing constructions and triadic parallel lists regardless of profile. In its SGR-001 study, the bare arm (no profile) produced a listicle with four headings and fourteen bolded topic sentences; the profiled arm, given the same brief, produced continuous prose with no headings. The profile reaches ChatGPT, Claude and other tools through the API and MCP server.

Questions

What is a structural signature in writing?

It's a recurring structural pattern, such as a habitual opener or a closing question; the app defines the attribute as "Notable recurring structural patterns." ScriptGrain stores it as a short free-text field (max 15 words) inside the Quirks and cadence group, written by the extraction model after reading your samples in two passes, not counted by any script.

Is there a writing structure analyser that measures this automatically?

ScriptGrain's free voice profile reads structural signatures from a writing sample as one of 45 attributes, alongside sentence, pronoun and word-length features. It's judged by an extraction model rather than counted, so treat it as a description rather than a score. ScriptGrain's writing voice attributes reference defines all 45.

How is writing structure measured, exactly?

In a ScriptGrain profile, not by counting headings or bullet points. The extraction model reads each sample in pass one, then writes a single description of up to 15 words in pass two. There's no formula and no reference corpus figure for it. Sentence length and contraction frequency are different: code counts those when a draft is scored.

How do I make AI stop using the same structure every time?

Ban the specific patterns causing it: triadic lists and 'not X but Y' constructions, which ScriptGrain's own generation bans regardless of profile. Ask for no headings on short pieces, and vary paragraph shapes on purpose rather than repeating one rhythm. Feeding it full sample pieces, not fragments, also helps it see your actual shape rather than guessing.

How do I make ChatGPT stop writing listicles?

Tell it explicitly not to use headings or numbered lists on short pieces, and show it one full example of continuous prose in your own shape. In ScriptGrain's SGR-001 study, the bare arm (no profile) produced a listicle with four headings and fourteen bolded topic sentences; the profiled arm produced continuous prose with no headings. Make the instruction specific rather than a general request for "flow".

Related

Sources