Ghostwrite in five voices. Keep them all separate.

By Jack Stovell X · Instagram · LinkedIn · published 2026-08-03 · updated 2026-09-30

One profile per client, measured from their real writing. No bleed, no guesswork, no "does this sound like them?" panic at 11pm.

If you're weighing up other tools that measure style, My Writing Twin is the fairest comparison we have. Both of us measure rather than describe. The difference is what comes out the other end: it hands you a profile, a prompt and a connector for the models you already use, while ScriptGrain also drafts in the profile and scores each draft against your voice.

The problem you already know about

You're not one writer. You're five, ten, maybe more, depending how many clients you've got on retainer this month. Client A writes short and blunt. Client B loves a rhetorical question. Client C hasn't used a semicolon since 2019 and never will.

Keep switching between them by memory and instinct, and eventually one voice leaks into another. Client B starts sounding a bit like Client A. You notice. Client B notices. That's the whole business, right there, quietly eroding.

One measured profile per client

ScriptGrain runs a two-pass stylometric analysis on each client's real writing: lexical choices, syntax, tone, rhetorical habits, punctuation, function words, content patterns, quirks. Forty-five attributes, measured, not guessed.

Each profile sits in its own lane. Nothing crosses over. Generate inside Client A's constraints and you get Client A, full stop, every time.

Evidence, not vibes

Tell a client "this sounds like you" and they want proof, not reassurance.

Every draft ships with a voice-match score against that client's profile. It's not a testimonial. It's a number you can put in front of them before they even open the doc. That's the pitch that lands in the second meeting, not the first.

One profile, every format

LinkedIn post on Monday. Newsletter on Wednesday. A client email that needs sending in the next ten minutes. Same client, same voice, three completely different formats.

The profile doesn't care what you're writing. It holds the voice constant and lets the format change. That's the whole point of measuring the person and not the platform.

Plans built for how ghostwriters actually work

  1. Free: one profile, one full analysis. No card required. Try it before you commit to anything.
  2. Writer, £12/month: one profile. Fine if you've genuinely only got one client right now.
  3. Operator, £29/month: three profiles. Most solo ghostwriters land here.
  4. Studio, £99/month: ten client profiles. Built for the ghostwriter juggling a full roster.
  5. Agency, £299/month: unlimited profiles, ten team seats included. For agencies where more than one person is writing in the same client's voice, and it needs to hold.

Try it before you trust it

Build one profile free. See the 45 attributes it pulls from a real sample. Then decide if it's worth paying to generate in that voice.

Ten clients, ten voices, zero bleed. That's the job. This is the tool that keeps it that way.

Onboarding a new client in an afternoon

Start with their own words. Not your notes from the intro call, not the brand deck. Published writing: articles, newsletters, LinkedIn posts, anything they wrote themselves rather than something a previous ghostwriter tidied up.

A profile needs one sample of 50 words or more. That's the floor. The recommendation is three or more pieces and about 3,000 words, and you should treat that as the real target. Why? One short post tells you how someone sounded on one day. Three pieces tell you how they sound.

So ask for a spread. A punchy post, a longer essay, maybe an email they sent to their team. Drop them in, run the analysis, and wait. Extraction usually takes under a minute (the median we measured was 39 seconds). Go and make tea; it's done before the kettle is.

What comes back is a profile of 45 attributes across eight layers: sentence length and its variance, contraction rate, comma density, semicolon and dash habits, formality, opening style, pronoun mix. Read it with the client beside you. Fair enough if half of it surprises them. Most people have never seen their own habits measured, and "you never use a semicolon" lands harder than any style guide ever will.

What the score gives you in a client conversation

Every draft is scored from 0 to 1 against the client's profile. That's the voice match. On-voice drafts typically land between 0.85 and 0.95. Partial matches sit from 0.5 to 0.8. A clearly different voice drops under 0.45.

"This doesn't sound like me" is a feeling, and you can't argue with a feeling. You can, however, look at the deltas.

The score comes back with per-feature notes: the draft's value set against the profile's value. Say a client flinches at a paragraph. You check, and the draft's average sentence length runs well above theirs, while hedging is higher than they'd ever allow. Now the conversation changes. You're no longer defending taste. You're pointing at two numbers and saying, "Here's the gap. Let's close it."

And to be fair to the client, sometimes the profile is right and the draft is wrong. Sometimes it's the reverse, and their recent writing has drifted from the samples. Either way you've got something concrete to work from.

Polish revises a draft toward a target score (the default is 0.9), and the app learns from the edits you save. So the fix is rarely a rewrite. It's a nudge, then a second look at the number.

Keeping several voices from bleeding into each other

Bleed happens in your head first. You write Client A's blunt fragments in the morning, then carry that rhythm into Client B's afternoon piece without noticing.

The fix is structural: one profile per client, each held apart. Operator holds 3 profiles, Studio holds 10, and Agency is unlimited. Pick the tier that matches your actual roster, then give every client their own lane and leave it there.

The profile also travels with you. It reaches ChatGPT, Claude, Gemini, Cursor, Meta Muse and any MCP client through the MCP server and the REST API, read live each time. It is never a pasted prompt, which matters more than it sounds. A pasted prompt goes stale, gets copied into the wrong chat, and quietly ends up describing the wrong person. A live read can't.

A few habits keep the lanes clean:

  1. Name each profile after the client, nothing cleverer.
  2. Score every draft against the profile it was written for.
  3. Rebuild a profile when the client's own writing has clearly moved on.

That last one matters. People change how they write. A profile built from old samples will eventually describe someone who no longer exists.

So keep the lanes separate, and let the numbers do the arguing for you. That's the job.