# Grizzly Voice

> Build or update the author's VOICE.md spec. Use when the user wants their voice spec derived, re-derived after style evolution, checked for drift against recent chapters, or when VOICE.md is provisional and real chapters now exist.

- Skill: `harishdvs/grizzly-voice` (Agent Skill)
- Install (CLI): `npx skillmds@latest add harishdvs/grizzly-voice`
- Raw SKILL.md: https://api.skillmd.com/api/skills/harishdvs/grizzly-voice/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: HarishDvs (https://skillmd.com/u/harishdvs)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/harishdvs/grizzly-voice

---


# Grizzly Voice — The Voice Spec Keeper

VOICE.md is the positive style target every draft and edit anchors against. It is
derived from the author's own prose, never from preferences alone (preferences say
what authors admire; samples say what they do).

## Deriving (first time)

Follow the procedure in grizzly-init Step 3: 2-3 samples of 300+ words of the
author's own hand, analyzed field by field against `templates/VOICE.md`, counted
not guessed, anchor lines confirmed by the author.

## Reconciling samples with the blacklist

The samples will not always agree with `decks/blacklist.md`, and they will not
always agree with each other. Two checks, both decided by the author, never assumed:

1. **Signature vs tell.** When a construction in the samples matches a blacklist
   entry (stacked short fragments reading as rule-of-three, emotion named directly in
   a given register, a recurring cadence), do not silently treat it as a tell and do
   not silently keep it. Surface it and ask: is this a signature to protect, or a
   crutch to tighten? Record the ruling in VOICE.md (blessed, blessed-with-a-cap, or
   treated-as-tell). tell-scan keeps these on its REVIEW tier precisely so a human
   ruling, not the regex, decides; the VOICE.md ruling is what those REVIEW hits are
   judged against.
2. **Target register when samples diverge.** When one sample runs more restrained and
   another fuller (more named emotion, more intensifier adverbs), name the gradient
   and ask which register is the target. This is the growth-vs-drift call from the
   re-derive procedure applied at first derivation: the author chooses the direction,
   and the spec then leans that way.

## Re-deriving (style evolved)

1. Ask which recent chapters represent the author at their current best.
2. Run the same analysis on those.
3. **Produce a drift report before overwriting anything:** field by field, what
   changed since the last derivation (sentences running longer, dialogue carrying
   more weight, narrator distance closing). Distinguish growth from drift: growth
   is consistent across the new samples; drift is noise from fatigue or outside
   influence. The author decides which is which; never assume.
4. Update VOICE.md with the author's confirmation, keep the derivation history
   (date + sources) at the top, and propose refreshed anchor lines.

## Drift check (no rewrite, just a reading)

When asked "does chapter N still sound like me": compare the chapter against the
anchor lines and the spec, quote the passages that drift, name the field they
drift on, including grammatical register and tense (the modern-maxim or trailerese
button, the present-tense leak in past-tense narration; see `decks/register-tense.md`).
No fixes here; route to grizzly-edit.

## Rules

- Never derive from AI-assisted passages; the spec must come from the author's
  hand, or it becomes a mirror of the assistant.
- VOICE.md belongs to the author. Propose; never silently update.
- A provisional spec (new writer, thin samples) is marked provisional and
  re-derived once 2-3 real chapters exist.