Same Brief, Four Voices: What a Voice Profile Actually Does
By Jack Stovell · 2026-07-28 · Research
Same Brief, Four Voices: What a Voice Profile Actually Does
Right, so people keep asking me this, and I've been dodging it a bit, if I'm honest, because the answer is annoying to explain in a sentence. What does a voice profile actually buy you over just... prompting a good model well? So this morning I sat down and ran an actual test. No cherry-picking, one run each, and I'm publishing exactly what came out, warts and all.
Here's what ScriptGrain actually does: it takes a handful of your writing samples and pulls out 45 measured things about how you write, sentence length, contraction rate, punctuation habits, humour, argument structure, all of it, plus a short narrative describing your voice. Then it enforces that as a hard constraint when it generates anything. That's the pitch. But pitches are cheap, so.
The method, plainly
I wrote three short content briefs from scratch this morning, all made up: a feature announcement for a fictional banking app doing round-ups into savings pots, an opinion piece on productivity advice, and a piece about killing status meetings in favour of async updates. Each brief went through two arms, byte-for-byte identical. Arm one: the same frontier model ScriptGrain runs on, called bare, no system prompt, nothing but the brief. Arm two: ScriptGrain's pipeline with a voice profile applied. One run each. No edits.
Test 1: the neobank
Profile built from a UK challenger bank's public copy.
Bare model headline:
"Say Hello to Round-Ups: The Easiest Way to Save Without Thinking About It"
Profiled:
"Small change. Add up."
Bare opening:
"We've got some good news for anyone who's ever meant to save more but never quite got round to it."
Profiled:
"You know that feeling when you pay £3.40 for a coffee and think 'well, that's basically £3.50'? We thought so too."
Bare close:
"Head to the Pots tab today and give it a try."
Profiled close:
"Open the app, switch it on, and let your spare change do something useful for once."
Same facts. Completely different person saying them. Average sentence length dropped from about 17 words to about 11, sentences under eight words doubled from 24% to 48%, contractions went up roughly 70%. And the bare version used ten em dashes. Ten. The profiled one used none. That's the tell, by the way, if you've ever wondered why AI copy all reads a bit... the same. It's the dashes.
Test 2: the blunt opinion essayist
Same productivity brief, both ways. Bare opens like it's writing for the census:
"Every January, millions of people resolve to 'manage their time better.'"
Fine. Competent. Nobody wrote that though, not really, it's just what gets produced when nobody's asked for anything specific.
The profiled version opens on a potter in Bristol. An actual named person doing an actual thing, not "millions of people". It talks like someone with a grudge, describing an app that locks your phone away
"like it's a naughty toddler".
I laughed at that one, not gonna lie.
Closings. Bare assigns homework, tidy and forgettable:
"Before optimizing your calendar, audit your attention."
Profiled reframes the whole question:
"Stop asking 'how do I find more time.' Ask instead, 'what is currently allowed to interrupt my head, and why does it have that permission.'"
Here's the bit that actually surprised me. Average sentence length was basically identical, about 15 words both. So it's not about making sentences shorter, that's not the trick. The voice lives in the tails: sub-eight-word punch sentences at 36% versus 16%. Style isn't the average. It's the distribution. I keep saying this to people and I don't think it lands until you see the numbers side by side.
Test 3: the no-nonsense B2B operator
Async-updates brief. Bare eases in gently:
"If you run operations at a 20-200 person company, you know the ritual"
Profiled opens with an order, no easing anywhere near it:
"Kill the weekly status meeting. Not soften it, not shorten it - kill it."
And later, mid-piece:
"That's it. No preamble, no 'hope everyone's week is going well.' Just the data."
Formatting turned out to be identity too, which I hadn't fully clocked until I laid them next to each other. The bare version is a scannable listicle: four headings, fourteen bolded topic sentences, the whole LinkedIn-carousel format. The profiled one is just prose. Continuous. No headings, no bold, nothing to skim. It landed at 605 words against a 600-word brief. The bare one ran to 708, padding itself out the way these things do when nobody's told it not to.
Test 4: the party trick
Right, this one I did mostly for my own amusement. I fed the exact same banking brief, the round-ups one, through a profile built from a Victorian novelist's prose. Just to see.
It opened:
"It was a Tuesday, unremarkable in every respect, when I bought a coffee for two pounds and eighty pence and thought nothing further of it."
Zero contractions in 626 words. Not one. No headings, no bullet lists, seven long paragraphs, sums of money spelled out in words rather than numerals. The actual how-to instruction, "how do I turn this feature on", got folded straight into the narrative:
"Turning it on takes rather less time than reading this sentence has done."
Same facts as the neobank brief. Same feature, same £2.80 coffee, same round-up mechanic. Entirely different century. The brief carried the information. The profile carried everything else, the rhythm, the century, the whole personality of the thing.
What it means
The bare model isn't bad, I want to be clear about that, because it's tempting to make this sound like a takedown and it isn't. It's competent every single time. Clear, organised, friendly. That's exactly the problem though: it's competently anonymous, and it's the same competently anonymous for everyone, because everyone's pulling from the same default voice. If you publish at any real volume, that default is a commons. Your competitors are drinking from the same well you are. Literally the same em dashes.
A profile doesn't make the writing better, and I want to be honest about that rather than sell it as magic. What it does is give the writing an identity: rhythm, stance, punctuation habits, formatting choices, the fingerprint a reader half-recognises without knowing why. The facts stay exactly the same. The author changes.
And to be fair about the limits, because I'd rather say this than have you find out later: a profile needs real writing samples to learn from, it can't invent a voice out of nothing. It won't rescue a bad brief; garbage in is still garbage, just garbage with better sentence rhythm. And it will not invent expertise you don't have. It's not a ghostwriter for knowledge you never had in the first place, it just makes sure the knowledge you do have sounds like you when it lands on the page.
Anyway. If you want to see what your own writing measures out to, the actual numbers, not vibes, there's a low-pressure way to try it at scriptgrain.com/voice-preview. No pitch attached to that link beyond, well, see for yourself.
---
Method note: every quote above is from an unedited, single-run output generated on 27 July 2026; briefs were byte-for-byte identical across arms.
By Jack Stovell, founder of ScriptGrain. I build the product this blog writes about; everything here comes from shipping it. See how ScriptGrain works.