# How to develop a writing voice you can actually measure · ScriptGrain

> A method for developing a writing voice with a measurement at every step: baseline your habits, pick the three to change, write, measure, keep what moved. With the numbers 299 human pieces sit at, so you know where you stand.

Canonical: https://scriptgrain.com/blog/develop-a-writing-voice-you-can-measure

# How to develop a writing voice you can actually measure

*By Jack Stovell · 2026-09-24 · Guides*

A writing voice develops fastest when you can measure it. "Find your voice" gives you nothing to change on Tuesday. "Your sentences average 24 words with 6% under eight" does. One is a mood. The other is a target you can hit or miss, and know which.

*What was measured: 45 attributes across lexical, syntactic, tone, rhetorical, punctuation, function word, content and cadence layers, extracted in two passes from a corpus of 299 public pieces (2026-09-20). The schema lives in scriptgrain.com/writing-voice and the underlying voice-profile extraction. Anchors below come from that same corpus.*

## Start by measuring what you already do

Here's the thing: most writers guess at their own style. They think they're "concise" because they'd like to be. So the first move isn't picking a new voice. It's finding out what your old one already looks like on paper.

Take five pieces you've published and run them through something that counts. Average sentence length, for instance: the corpus median sits at 16.8 words, with the 10th percentile at 11.6 and the 90th at 23.4. If you're at 24, you're writing longer than nine in ten people in that sample. That's not good or bad. It's just a fact you didn't have yesterday.

Do the same for short sentences. The corpus median for sentences under eight words is 20%, ranging from 8% to 40% across pieces. If your number is 5%, your writing probably reads breathless, one long clause after another with no landing spot. If it's 45%, you might be choppier than you intend.

Contractions matter too. Median across the corpus is 19.5 per 1,000 words, with a spread from 8.1 to 32.2. That range alone tells you formality is a dial, not a switch. You can run this whole check for free at [writing style analysis](https://scriptgrain.com/tools/writing-style-analysis), which reports your numbers against those percentiles rather than against a vibe.

None of this requires new writing. You're measuring the voice you already have. That's the baseline everything else gets compared to.

## Pick three habits, not a personality

You can't fix forty-five things at once, and trying to will make your prose worse, not better. So pick three. Concrete ones.

Good candidates: sentence length variance, comma density, and contraction rate. All three are measurable, all three change fast, and none of them require you to invent a new personality on the page. "Be more authentic" isn't a habit. "Get 25% of sentences under eight words" is.

Here's a short example of the difference in practice:

Before: "The report indicated that engagement had declined across most channels during the period under review, which the team attributed to a combination of seasonal factors and reduced spend."

After: "Engagement dropped. Most channels, most of the quarter. The team blames seasonality and a smaller budget."

Same information. Different rhythm. The second version has two sentences under eight words in a stretch of three, which is exactly the kind of burst the corpus anchors reward, sitting near that 20% median rather than nowhere close to it.

To be fair, three habits sounds arbitrary. It isn't. Attention is finite, and a longer list just means you measure everything and change nothing. Three gives you enough room to notice progress without drowning in metrics you'll ignore by the second paragraph.

## Write, then measure the draft against the baseline

Write the thing. Don't think about the numbers while you're doing it, that's what editing is for. Once the draft exists, measure it the same way you measured the baseline, and put the two side by side.

This is where most advice stops, which is a shame, because the comparison is the useful part. If your baseline average sentence length was 24 words and the new draft comes in at 19, you've moved two thirds of the way to the corpus median of 16.8. If your under-eight-word share went from 5% to 15%, you're closing on that 20% median but not there yet. Progress you can see is progress you can trust.

There's a published study on exactly this kind of shift. SGR-001, "Same Brief, Four Voices" (28 July 2026), ran the same briefs bare and through a voice profile on the same model. On the banking brief, average sentence length fell from about 17 words to about 11 once the profile was applied. Sentences under eight words doubled, from 24% to 48%. Contractions rose by roughly 70%. Ten em dashes became none. One run per arm, unedited, but the direction is the point: measurable inputs produce measurable, repeatable shifts in output. That's not "the AI understood my brand." That's arithmetic.

If you're working with AI drafts specifically, the same logic applies whether the model is generating from scratch or revising your work. A [personal AI writing style](https://scriptgrain.com/personal-ai-writing-style) profile does this by feeding your measured attributes back in as constraints, rather than asking the model to guess at your voice from a paragraph of examples.

## Keep what moved and let the rest be

At the end of the day, not every number needs to change, and chasing all forty-five is how you end up sounding like nobody at all.

Some attributes are load-bearing. Sentence length, contraction rate, comma density: these shift the felt experience of reading you, and they're worth defending draft after draft. Others barely register. Your exact ratio of "the" to "a" to "an" isn't something a reader consciously notices, even though it's part of the full profile. Most writers could be off by a few points either way and nobody would flag it in a comment.

So keep the three you picked. Check them each time you publish. Let the fine-grained stuff (rare word rate, exact parenthetical density) sit in the background as description rather than target. You're not trying to hit forty-five numbers exactly. You're trying to make the two or three that carry your rhythm stay put across fifty pieces instead of drifting.

This is also where consistency checking earns its keep, especially if more than one person writes under the same byline or brand. A [brand voice consistency checker](https://scriptgrain.com/tools/brand-voice-consistency-checker) compares new pieces against the established profile and flags drift before it compounds. Catching it at 0.6 is easier than catching it at 0.3, once six months of pieces have quietly wandered off in the same direction.

## What the numbers cannot tell you

And here's the honest caveat: the numbers describe. They don't decide.

A profile can tell you that your average sentence length is 14 words and your contraction rate is 22 per 1,000. It won't tell you whether the joke in paragraph three lands, or whether the analogy about tax returns actually clarifies anything for a reader who's never filed one. That judgment is still yours. It always was.

I keep my own numbers in a spreadsheet I check maybe once a month, not every article. That's a personal habit, not a rule, but it's saved me from over-correcting after one bad draft. A single piece with too many long sentences isn't drift. Six in a row is.

The full attribute list, if you want to see what's actually being tracked underneath all this, sits at [writing voice attributes](https://scriptgrain.com/reference/writing-voice-attributes). It's long. Most of it you'll never consciously adjust. But knowing it exists is different from guessing what "voice" even means when someone says find yours. There's also a plainer definition of the underlying idea at the [stylometric fingerprint](https://scriptgrain.com/glossary/stylometric-fingerprint) glossary entry, if the jargon's getting in the way of the point.

## Questions

### How long does it take to develop a measurable voice?

Not long to get a baseline, that's one pass through five pieces. Change is slower: expect a few drafts before your three chosen habits move noticeably and hold. SGR-001 showed shifts within a single generation pass on a briefed model. For a human writer editing by hand, three to five pieces is a realistic window to see stable movement rather than noise.

### Do I need to write more to develop a voice?

No. Volume helps you gather more samples to measure, but it doesn't fix an unmeasured habit by itself. Five well-chosen pieces analysed properly beat fifty written blind. What matters is comparing draft against baseline each time, not simply producing more text and hoping a pattern emerges on its own.

### Can AI help me write in my own voice?

Yes, if it's given your actual numbers rather than a vague description. SGR-002 compared a measured voice profile against a plain ChatGPT prompt using one author's held-out posts, and the profiled version tracked the author's real attributes far more closely. Feeding a model your sentence length, contraction rate and structure beats asking it to "sound like me."

### Will measuring my voice make me sound like everyone else?

No, and that's the opposite of what it does. Corpus anchors show where a typical piece sits, they're not a target to copy. Your three chosen habits are what make you distinct from that median, not what pull you toward it. A shared measurement system doesn't flatten difference, it just makes the difference visible and checkable.

### How many samples do I need to get a reliable baseline?

Five is a reasonable working minimum, enough to average out one unusually short or unusually long piece. Fewer than three and a single off day skews everything. The corpus behind these anchors runs to 299 pieces, but you don't need anywhere near that for your own baseline, just enough that one weird Tuesday doesn't define your whole profile.

[More from the ScriptGrain Journal](https://scriptgrain.com/blog)