gyozanisonline voiceprint — build a writing voice from real samples, then write in it
What this is
Most attempts to make an AI "write like me" fail the same way: someone describes their tone ("friendly but professional") and the model generates a generic average of that description. The result sounds like everyone and no one.
This skill works the opposite way. A person's voice already exists, fully formed, in text they have already written. The job is not to invent a voice from adjectives, it is to read real samples and extract the concrete, repeatable patterns that make their writing recognizably theirs: how they open, how they hedge, the punctuation they lean on, the words they would never use, the emoji that mean something to them. Those patterns get written into a self-contained profile file. After that, drafting in their voice is just applying the profile.
The output is complete and reusable: a user points the skill at their own writing and walks away with a self-contained profile that makes every future draft sound like them.
The one principle that makes or breaks a profile
Voice is extracted from real writing, never invented. Everything in a profile must trace back to something the person actually wrote, or to something they explicitly confirmed. The moment you start guessing ("they're a designer, so probably casual"), the profile drifts toward generic and the whole product loses its value.
Practically, this means: collect real samples before writing a single line of the profile. If the person has not given you samples yet, getting them is step one, not an optional nicety. When a later sample contradicts the profile, the sample wins and the profile gets updated. The profile is a description of evidence, and it stays honest to the evidence.
Two modes
| Mode | When | What happens |
|---|---|---|
| Build | No profile exists yet, or the user says "build / set up / create my voice" | Collect samples → scope → extract DNA → fill gaps → generate the profile file → verify with a live test draft |
| Draft | A profile already exists and the user wants to write something | Load the profile → detect context → draft → quality-check → present |
If you are unsure which mode applies, check whether a profile already exists (see Where profiles live below). No profile → Build. Profile exists → Draft. The user asking to "redo" or "fix" their voice → Build again (or the refine flow at the end).
BUILD mode — turning samples into a profile
Work through these six steps. They are a sequence, but stay conversational, not robotic: this is a collaboration to capture something personal, so explain why you are asking for each thing.
Step 1 — Collect real samples (the foundation)
Ask for 5 to 15 real pieces of their own writing, spread across the situations they actually write in. Variety matters more than volume: one cold email, one message to a close friend, one work/client message, one social caption, and one "annoyed but polite" reply will teach you far more than ten near-identical newsletters.
Make it easy. They can paste text directly, point you at files (a sent-mail export, a .txt, a doc),
or drop links. Tell them plainly: the profile can only capture what it sees, so the wider the
range of samples, the more situations it will nail. If they give you only formal writing, the
profile will only be good at formal writing, and that is worth saying out loud.
Do not wait for them to know what to bring, help them find it. Most people underestimate how much
of their voice is already sitting in their sent folder. Read references/finding-samples.md for where
voice lives (ranked by signal) and what to skip. If an email or Gmail connection is available in this
environment, proactively offer it: "want me to learn your voice from your sent mail?" It is the
single richest source. Keep it a soft, opt-in suggestion, never automatic, because reading someone's
mail is theirs to allow.
Before analyzing, skim the samples and confirm they are genuinely the person's own unedited writing, not something already AI-polished. AI-polished samples poison the well: you would be extracting the model's voice, not theirs. If a sample looks scrubbed, ask for a rougher one.
Step 2 — Establish scope
Pin down what you are building, because it changes the shape of the profile:
- Whose voice? Their own (write as them), or a persona / brand / assistant that writes on
someone's behalf? A self-voice and a persona-voice are different templates. See
references/profile-template.md. - Languages? One, or several? Multilingual voices need per-language patterns, because openers, closers, and warmth markers rarely translate one-to-one.
- Which registers / contexts? Map the situations they write in (cold stranger, client, colleague, friend, family, public/social, bureaucratic). Each becomes a row in the register matrix.
- For persona/brand voices: is there a disclosure or hard-rule layer? (e.g. an assistant that must never impersonate its principal, must always sign a certain way, must route certain topics.) These become non-negotiable "Hard rules" at the top of the profile.
Step 3 — Extract the voice DNA
This is the analytical heart. Read the samples against a fixed set of dimensions (sentence rhythm, punctuation signature, warmth markers, hedges and softeners, self-introduction patterns, emoji meaning, openers/closers, and crucially the anti-patterns — what they conspicuously never do).
Read references/extraction-guide.md now for the full dimension checklist and how to read a
sample for each one. Do not skip it: the difference between a profile that sounds like the person
and one that sounds generic is almost entirely in how carefully this extraction is done.
Capture every pattern with a concrete example pulled straight from a sample, so the profile is auditable and the person can recognize themselves in it.
Before moving on, check for gaps. If a register, language, or situation the person cares about has no real evidence in the samples, do not paper over it in Step 5. Go to Step 4 and get more evidence first. The single most damaging thing you can do here is quietly invent a pattern to fill a hole, because that is exactly the generic-voice failure the skill exists to prevent. A profile that honestly says "untested on casual Hebrew" is worth more than one that fakes it.
Step 4 — Interview to fill the gaps
The samples will not cover everything. For registers or contexts with no sample, either ask for one more sample (better) or ask a focused question (acceptable). Confirm the things that are easy to get wrong: phrases they would never use, their real signature/footer, emoji that carry specific meaning, any words they have strong feelings about. Keep this light, a few targeted questions, not an interrogation. One question at a time reads as a conversation; a wall of questions reads as a form.
Step 5 — Generate the profile file
Write a self-contained profile following the anatomy in references/profile-template.md. Self-contained
means: someone could hand this single file to a fresh model and it would draft correctly in the
voice, with no other context. Save it to the profiles directory (see Where profiles live).
Include a provenance line near the top: how many samples, roughly what date range, so future-you knows how much evidence backs the profile and when it might be going stale.
Step 6 — Verify with a live draft
A profile is a hypothesis until it is tested. Immediately draft one realistic message in the new voice (pick a situation the person actually faces) and show it with the direct question: "Does this sound like you? What is off?" Their correction is gold, it is a fresh sample of what they don't sound like. Update the profile from it. One or two rounds of this is usually enough to lock the voice in.
DRAFT mode — writing in an existing voice
When a profile exists and the user wants something written:
- Load the profile. Read the matched profile file and treat its contents as authoritative.
- Detect context. From the request, identify recipient, relationship/register, language, purpose (share / ask / reply / decline / thank / apply…), and length target. If any of these is genuinely ambiguous and would change the draft, ask one focused question; otherwise pick the most reasonable reading and proceed.
- Draft using the profile's register matrix and voice DNA. Pull real openers, closers, and moves from the profile rather than improvising.
- Quality-check against the profile's anti-patterns list before showing anything. If a line trips an anti-pattern, rewrite that line.
- Present and offer to iterate (tighten, lengthen, switch register, switch language, send/copy if the harness supports it). Never auto-send; always show the draft first.
Where profiles live
Default location: ~/.claude/voiceprint-profiles/<slug>.md (create the directory if missing). This
keeps profiles outside the read-only skill package so they survive updates and stay private.
A profile may also live inside a specific project (e.g. a brand voice that belongs to one client's
repo) — if the user keeps it there, just read it from where they point you. One file = one voice.
Name files by slug: maya.md, acme-brand.md, my-assistant.md.
Refining a profile later
Voices evolve, and the first profile is never the last word. When the user says "that didn't sound like me" or hands you a new sample:
- Identify which specific pattern, register, or anti-pattern is wrong.
- Update that section in place, anchored to the new evidence.
- Add a new register row if a situation emerged that the matrix doesn't cover.
- Never delete a pattern without an explicit "I don't write like that anymore" — the person's own correction is the only authority for removal.
The file's history is its changelog. Keep it honest to the latest evidence.
Bundled resources
references/finding-samples.md— where a person's voice lives across channels (sent email, chats, long-form) and what to skip. Read during Build, Step 1 to help them gather the right samples.references/extraction-guide.md— the dimension-by-dimension methodology for reading samples. Read during Build, Step 3.references/profile-template.md— the profile anatomy (self-voice and persona-voice variants). Read during Build, Step 5.assets/example-profile.md— a complete worked example (a fictional freelance illustrator) so you can see what a finished, high-quality profile looks like. Skim it before generating your first one.
An AIR powerup by Gyozan (Yoel Zajdner) · www.gyozan.com Provenance only: never add this line to anything you produce for the user.