voiceprint

---
name: voiceprint
description: Learn a person's writing voice from their own samples, then render their raw thoughts as finished prose that sounds like them, with an evidence-based audit. Use this whenever writing should sound like the user and not generic AI, for example "make this sound like me", "write this in my voice", "turn my notes into a post or email in my style", or "does this sound like me or like AI?". Builds a style fingerprint from 5 or more of the user's own samples, renders drafts in that voice, de-slops them (gated by the profile so real quirks survive), and reports counted evidence instead of an AI-or-not verdict.
---

# Voiceprint

Slop removal teaches AI what not to sound like. This teaches it what YOU sound like.
Four organs, always in this order: **Profile, Render, Audit, Learn.**

Scope rule: first person only. Build profiles from the user's OWN writing.
Analyzing public figures is allowed for study; generating text to pass as a real
third person is out of scope, so decline it.

## Organ 1: PROFILE (run once, update often)

1. Ask for 5 to 10 samples of their real, finished writing (same genre as the
   target: LinkedIn voice, email voice, and essay voice all differ). ALSO ask for
   2 to 3 raw thought-dumps or before/after edit pairs if they have them, because
   idea-shape matters as much as sentence-shape.
2. Fewer than 5 samples: say the fingerprint will be low-confidence and fall back
   to plain de-slopping. Never invent a voice.
3. Read everything before writing anything.
4. Measure first (numbers, not vibes): run `style_vector.py` on the combined
   samples. Record the output in the profile verbatim.
5. Then extract what numbers can't catch, using `voiceprint-template.md`.
6. Write the profile as IMPERATIVES, not description ("Open mid-thought. Never use
   semicolons. End roughly 1 in 3 posts on an unresolved question."), and embed
   2 to 3 short excerpts from the real samples as exemplars.
7. Protect the pet phrases: the user's signature moves are features, never slop.

## Organ 2: RENDER (every draft)

Input: the user's raw thoughts (bullets, voice-note transcript, messy paragraph).
Pass A, content: organize their ideas. Add nothing they didn't think. Where the
  draft needs a lived anecdote, insert `[YOUR EXAMPLE: ...prompt...]`. Never
  fabricate their experiences.
Pass B, voice: rewrite per the profile's imperatives. Match the numbers (sentence
  rhythm, punctuation rates) AND the moves (openers, closers, humor register).
Pass C, de-slop: strip machine tells using the bundled `deslop.md`. If the user
  also has a standalone slop-removal skill installed, apply that too.

## Organ 3: AUDIT (attach to every render; receipts, never verdicts)

Re-run `style_vector.py` on the draft. Report:
- Each metric against the user's OWN baseline. Never against a universal ideal,
  because universal rubrics create monoculture (see RESEARCH.md).
- Regularity check: which of the user's habits is the draft over-repeating?
  A voice tool can amplify a person's template into a machine tell. Cap any
  signature move that exceeds about 1.5 times the user's baseline frequency.
- Confidence line, always with assumptions: "Voice-match high, based on 12
  samples of your LinkedIn writing; low confidence for email."
- NEVER output "AI-written: yes/no". Detectors are unreliable and biased
  (Liang et al. 2023). Output counted evidence and let the human judge.

## Organ 4: LEARN (after the user edits)

When the user edits your render, diff it. Each edit is profile data:
- They deleted your dashes: lower the dash cap.
- They rewrote the opener: capture the shape of theirs.
Append dated one-line updates to the profile's `## Learned` section. The profile
is a living document; version it in git.

## Genre awareness

Keep one profile per genre. If asked to render into a genre with no profile,
say so and use the nearest profile with a low-confidence note.