Humanising without rewriting: the edits that move the score most

By Jack Stovell · 2026-09-30 · Guides

Eight edits move a voice match score, and they're ranked by weight, because the scoring method is published and unit-tested. You can make all eight in about twenty minutes without touching a sentence you actually like. No rewrite required, just targeted fixes in the right order.

What follows draws on the scoring engine in supabase/functions/_shared/voice-match.ts (the same one behind the app, API, MCP server and free tools) and on SGR-002, "ChatGPT vs a measured voice" (2026-09-20).

Why rewriting is the wrong instinct

Here's the thing: most people treat a low voice match score as a verdict on the whole draft. So they bin it and start again. That's expensive, and it's usually wrong.

The score isn't one number pulled from a vibe. It's a weighted sum of measurable things: sentence length, contraction rate, comma density, pronoun mix, banned terms, and a handful of qualities a small model judges (formality, confidence, rhythm, argument structure). Each feature gets compared to your profile and scored for similarity. Some features carry a weight of 2. Others carry 0.5. A full rewrite touches all of them at once, randomly, and you can't tell what fixed the score and what just got lucky.

Targeted edits do the opposite. You know exactly which lever you're pulling, and by how much. That's not a stylistic preference. It's how the maths works.

The eight edits, ranked by weight

The scoring engine weights features from 2 down to 0.5. Fix the heavy ones first. Here's the order, with what each one is worth.

  1. Put the contractions back. Weight 2. AI drafts default to "it is" and "you will." If your profile runs on contractions, this alone can move the needle more than anything else on the list.
  2. Break up the average sentence. Length weight 2, variance weight 1.5. Long uniform sentences read as safe and sound nothing like a person. Add short ones. Vary them hard.
  3. Bring the register to the profile's formality. Weight 2. Too formal or too casual against the target both cost you, and the model judging this reads tone, not word count.
  4. Use the writer's own recurring words and connectives. Weight 2. Signature phrases are catalogued per profile; using them (or not) is a direct, countable comparison.
  5. Remove banned terms and listed tells. Weight 2, but it's a penalty only, a clean draft earns nothing extra from it. Em dashes and "not X but Y" constructions are the two worst offenders.
  6. Fix the pronoun mix. Weight 1.5. If your voice runs 46% "I" and 48% "you," a draft that's all "we" reads off, even if every sentence is fine on its own.
  7. Adjust comma density and semicolon use. 1.5 each. These are quiet signals, but they're counted precisely, and they're cheap to fix once you notice them.
  8. Reopen the piece and reshape the argument. Opening style 0.75, argument structure 1.5, rhythm 1.5. This one takes longer, but it's still not a rewrite: you're moving the conclusion earlier, not writing a new draft.

Notice something: the first five edits are almost entirely mechanical. Count contractions, count sentence length, swap words, delete a dash. You could do it with search and replace. The last three need a bit more judgement, but none of them require you to throw anything away.

The order to make them in

Order matters here because some edits interact, and doing them backwards wastes time. To be fair, you could do them in any order and eventually land somewhere similar, but the ranked order gets you most of the score in the first few minutes.

Start with contractions. It's the single highest-weight fix and it takes about ninety seconds: scan for "it is," "you will," "do not," swap them, done.

Then hit sentence length and variance together, since they're mechanically linked. Find your longest three or four sentences per section. Cut two of them in half. Add one short sentence somewhere in the middle of a long paragraph. That's it.

Register comes next, because once your sentence rhythm has changed, formality often shifts with it, shorter sentences tend to read less formal even with the same vocabulary. Check the register after step two, not before.

Signature phrases and banned terms go together too: you're scanning the same paragraphs for two different things, so do both passes at once. Add the words you actually use. Strip the ones on the list, especially the dash and the balancing construction.

Pronoun mix, comma density and semicolons are smaller fixes you can batch in one final read-through. They rarely need more than a few swaps.

The argument reshape comes last, deliberately. It's the most expensive edit on the list in terms of time, and by the point you reach it, five of the eight edits are already done. Often the reshape is just moving your last paragraph to the top.

A worked pass

Here's an invented AI paragraph, the kind of thing GPT-5 hands you by default, at 120 words:

"It is important to note that the architecture of a decision-making process is not always visible to those who are affected by it. Organisations often operate through a mechanism that prioritises efficiency, and this can create a geography of exclusion that is difficult to detect from the outside. It is not that the process is malicious, it is that it was not designed with transparency in mind. Employees may find themselves navigating a system that feels opaque, impersonal, and, at times, arbitrary. This is not a failure of intention but a failure of design, and it warrants closer examination by those responsible for building these systems."

Now the same paragraph after edits 1, 2 and 5 only, contractions, sentence length, and banned terms, nothing else touched:

"You often can't see how a decision got made. That's the problem. Organisations build systems for efficiency, and those systems end up excluding people without anyone planning it that way. The process isn't malicious. It just wasn't designed with transparency in mind. Employees end up navigating something that feels opaque and arbitrary. That's a design failure, and it deserves a closer look."

Three changes, stated plainly:

The exact score change depends on the profile you're scoring against, a profile that wants long sentences and zero contractions would score the first version higher. But for most human-voice profiles, this is roughly twenty minutes of work and it moves several of the heaviest-weighted features at once. That's the whole method.

When an edit is not enough

Sometimes you fix all eight and the score barely moves. That usually means one of two things.

Either the draft has a structural problem the edits can't touch: wrong argument shape, wrong opening, an entire paragraph that argues something your profile wouldn't say. Editing sentence length doesn't fix that. You need edit eight done properly, not just a comma swap.

Or the profile itself is thin. A profile built from too little text won't have reliable signature phrases or a stable formality reading, and no amount of editing against it will feel right, because the target is fuzzy. That's a different problem, and it's why the voice measurement framework matters as much as the editing checklist.

Worth naming the limits here too. SGR-002 ran two held-out posts against one author's profile, one run per arm, not blind. Mean voice match: ScriptGrain 0.83, bare GPT-5 0.74, GPT-5 with a written prompt 0.62. The prompt asked for roughly 14 contractions per 1,000 words and formality 3.7; GPT-5 delivered 0.5 contractions per 1,000 and a judged formality of 7.2, while still following the countable rules like no em dashes. That's the gap between an editable engine and a one-page prompt fallback: the prompt can describe the target, but only live scoring catches when the model quietly ignores it.

If you want to check your own draft against these eight edits without guessing, the brand voice consistency checker runs the same weighted comparison, and the AI cliche checker flags the banned-term list directly. For the deeper mechanics of how contractions, sentence variance and pronoun mix get compared and tolerated, see humanising AI writing.

Questions

Do these eight edits guarantee a higher score?

No. They move the highest-weighted features, which usually raises the score, but the exact change depends on your profile. A profile with low contraction use and high formality will score a heavily-contracted, short-sentence draft lower, not higher. The edits target what's measured; they don't guarantee direction if your assumptions about the profile are wrong.

Can I skip straight to edit five and ignore the rest?

You can, and removing banned terms always helps a little since it's a pure penalty with no downside. But it's the smallest single win on the list next to contractions and sentence length. Skipping to it alone leaves the two heaviest-weighted features untouched.

Why does sentence length carry so much weight?

Because it's cheap to measure and it correlates with several qualities the model judges separately: rhythm, complexity, confidence. A weight of 2 on average length plus 1.5 on variance means length-related fixes touch 3.5 points of possible score movement before you've changed a single word choice.

Does fixing formality mean changing my vocabulary?

Not necessarily. Formality is judged by a model reading tone, not a word-frequency count. Shorter sentences, more contractions and plainer connectives often shift the formality reading on their own, as the worked pass above shows: nothing in that paragraph got simplified, but it reads less formal once the sentences got shorter.

How do I know which signature phrases matter for my voice?

They're catalogued from your own text when the profile gets built. There's no guessing involved; the list is specific to you, drawn from words and connectives you actually use often enough to be a pattern rather than a coincidence. Check them against your draft directly rather than trying to remember them.

More from the ScriptGrain Journal