# Client voice onboarding: from a client's public site to a measured profile in a day · ScriptGrain

> A one-day onboarding for a new agency client's voice: collect the pieces that sound like them, build the profile (from pasted writing or their public site with consent), read the attributes, set the threshold, brief the writers and score the first draft.

Canonical: https://scriptgrain.com/blog/client-voice-onboarding-in-a-day

# Client voice onboarding: from a client's public site to a measured profile in a day

*By Jack Stovell · 2026-09-27 · Guides*

A new client's voice can be measured and briefed to your writers by the end of the working day. Not vaguely described, actually measured: the 45 attributes, a confidence score, a narrative you can hand to a stranger and have them write like your client wrote yesterday. The trick isn't magic. It's collecting the right five or six things before lunch and letting the profile run the afternoon.

*What was measured: the 45-attribute voice profile schema (supabase/functions/_shared/voice-profile) across lexical, syntactic, tone, rhetorical and punctuation layers. Sample-count effects studied separately in SGR-003 (scriptgrain.com/research/how-many-samples-does-a-voice-profile-need, 2026-09-21).*

## Morning: collect what actually sounds like them

Here's the thing: a kickoff call gives you opinions about voice. It doesn't give you voice. What you need instead is writing the client actually produced, unfiltered by whoever wrote for them last.

So collect the founder's own emails and posts. The about page, if they wrote it themselves. The newsletter, especially the early issues before anyone polished them. Interview transcripts count too, if you've got them; people talk in their real voice even when they can't write in it. Anything they wrote without an agency involved is gold.

Exclude the opposite category entirely. Press releases, legal copy, anything the previous agency drafted: none of that is the client's voice. It's someone else's voice wearing the client's name. Feed it into a profile and you'll get a beautifully measured account of a ghostwriter you'll never meet.

You need one sample of 50+ words at the absolute minimum, but don't stop there. Three or more pieces, somewhere around 3,000 words total, gives you a profile that holds up under pressure rather than one that wobbles the moment a writer produces something slightly different in tone. If you're wondering how much sample size actually changes the result, that's exactly what [SGR-003](https://scriptgrain.com/research/how-many-samples-does-a-voice-profile-need) went and measured.

## Build the profile

Extraction takes one to three minutes. You paste in the samples, or pull them via the API (`POST /v1/profiles`, up to 20 samples at once), and out comes a narrative, a confidence score, and all 45 attributes across the lexical, syntactic, tone, rhetorical and punctuation layers.

This isn't a vibe. It's counted: vocabulary diversity, sentence length and its variance, contraction frequency, comma density, semicolon use, whether they hedge or state things plainly, whether their jokes land dry or warm or not at all. It's the difference between "the client sounds punchy" and "average sentence length 14 words, variance 35, confidence-versus-hedging 0.8". One of those you can brief a junior writer with. The other you can't.

If the client has a public site worth reading, brand mimic will read it with their consent and produce a profile from that too, useful as a second opinion or a starting point before you've got real samples in hand.

## Read the profile before anyone writes

This is the step agencies skip, and it's the one that saves the afternoon. Read the narrative. Not skim it, actually read it, because it'll tell you things the kickoff deck won't: that this client never uses exclamation marks, that they run long build-up sentences before landing a short punchy one, that their humour is dry rather than warm and a writer who doesn't know that will produce copy that reads like a different person entirely.

The confidence score matters here too. A profile built from 3,000 words across five pieces is a different animal to one built from a single 200-word about page. Both are usable. Only one of them should be treated as gospel.

Read it before the brief goes out, not after the first draft comes back wrong.

## Set the threshold with the client

Every draft gets scored against the profile: sentence length, variance, contractions, formality, commas, pronoun mix, humour, argument structure, rhythm, all weighted and calibrated into a single number between 0 and 1. 0.85 to 0.95 counts as on-voice. 0.5 to 0.8 is partial, close but noticeably off. Under 0.45 and it reads like someone else wrote it, because functionally, someone else did.

The default threshold is 0.85. That's a reasonable bar for most clients, but it's worth naming out loud in the kickoff rather than assuming. A client with a very distinctive, idiosyncratic voice (heavy fragment use, unusual rhythm, lots of rare words) might need writers working closer to 0.9 before the copy actually sounds like them rather than a competent approximation. Agree the number before drafts start landing, not after someone's unhappy with one.

Full scoring method, weights and tolerances live at [/reference/voice-measurement-framework](https://scriptgrain.com/reference/voice-measurement-framework), if you want the writers to see exactly what's being counted.

## Afternoon: brief the writers and score the first draft

So the profile's built, you've read it, the threshold's agreed. Now the writers actually write.

The brief itself should be short. Point them at the profile, not at a mood board. Let them see the preferred words, the sentence length target, the humour register, the clause ordering (build-to-point or front-loaded, it changes everything about how a paragraph should land). A one-page prompt-builder export works as a fallback if a writer's tool can't read the profile live, but reading it live, through the API or an MCP client, is the better version: the profile updates as you add samples, the fallback doesn't.

First draft goes in, gets scored. The score doesn't just spit out a number; it names which attributes moved. Sentence variance too flat. Contractions too low. Too many exclamation marks for a client whose exclamation_rate is essentially zero. That's your revision note, already written for you.

Polish then pushes the draft toward 0.9, and if the draft already passes, polish is free; you're not spending a credit to fix something that isn't broken. That's the whole afternoon: brief, draft, score, polish, done before anyone's gone home.

## What to do when the client has no writing

Some clients genuinely haven't written anything. No blog, no newsletter, a website built by the last agency, an about page that reads like every other about page. This happens more than anyone admits at the pitch stage.

First move: interview transcripts. Sit the founder down, ask them to talk about the business the way they'd talk to a friend, and transcribe it. People who can't write a clean sentence will still talk in a distinctive, recognisable rhythm, and that rhythm is exactly what the profile is measuring.

Second move: the founder's actual emails. Not the ones drafted by comms, the ones they type themselves at 11pm. These are usually the richest source going, because nobody's performing in them.

Third: brand mimic, with explicit consent that you own or may use the site. It reads what's public and gives you a working profile even when nothing else exists yet.

And here's the honesty nobody wants to say out loud in the pitch: a profile built entirely from press releases is a profile of the last agency, not the client. It'll be measured, confident, internally consistent, and completely wrong. If that's genuinely the only material available, say so, and treat the resulting profile as a placeholder to be replaced the moment real writing exists. Don't brief a whole team off it and hope nobody notices.

For agencies running this across several clients at once, drift matters as much as the initial build; the [multi-client voice drift check](https://scriptgrain.com/tools/multi-client-voice-drift-check) exists for exactly that reason, and it's worth looking at once you've got more than one profile live. A [brand voice guidelines template](https://scriptgrain.com/reference/brand-voice-guidelines-template) is also useful for turning the narrative into something a client-side reviewer can sanity-check without reading all 45 attributes cold.

## Questions

### How many writing samples does a client actually need?

One sample of 50+ words is the technical minimum and will produce a profile. Three or more pieces, roughly 3,000 words total, produces one that holds up when a writer's draft varies slightly in tone. The exact relationship between sample count and profile stability is measured in [SGR-003](https://scriptgrain.com/research/how-many-samples-does-a-voice-profile-need) rather than guessed at here.

### Can we build a profile from the client's website alone?

Yes, using brand mimic, which reads the public site with the client's explicit consent that you own or may use it. It costs one brand-mimic credit and is a reasonable starting point, but it's a proxy. Real writing (emails, transcripts, a newsletter) will always outperform a mimic built purely from marketing pages.

### What does the score actually tell a writer to fix?

The first draft's score names the specific attributes that moved off-target: sentence length, contractions, formality, humour register and so on, each weighted by how much it matters to that voice. That's the revision note. Polish then pushes the draft towards 0.9, free if the draft already passes the threshold.

### Who sets the 0.85 threshold, us or the client?

0.85 is the default and works for most voices. Agree it with the client during kickoff rather than assuming it, especially for distinctive or idiosyncratic voices where a lower score still reads noticeably off. Naming the number upfront avoids a disagreement later about whether a draft "sounds right".

### What happens if the client's previous agency wrote everything we have?

Then you don't have the client's voice, you have the previous agency's. Say so plainly rather than building a confident-looking profile from press releases and legal copy. Use interview transcripts, founder emails or brand mimic instead, and treat anything built from agency-authored material as a placeholder to replace.

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