Distill Voice

---
name: distill-voice
description: >
  Distill HOW a newsletter is written into an operational, checkable Voice Profile (rules +
  verbatim examples + anti-AI-texture checklist), then calibrate it. Steps 7–8 of the Context Kit
  (Milestone 2). The profile is a GUARDRAIL that protects the writer's voice and flags AI texture —
  NOT a prose generator (D21). Run after the Brief. Use for "distill my voice", "build the voice
  profile", "voice guardrail", "how do I write".
---

# Distill Voice + Calibrate (steps 7–8)

An **actionable** Voice Profile a writing skill can apply and check — grounded in real prose,
shaped by the writer's answers.

## Precondition
Raw **full-prose** corpus (D12 — never summaries), scoped to the current voice era. The Brief
(keeps voice in service of identity). The writer's "never" list + disowned eras.

## Procedure
1. **Scope** to the current voice era (exclude disowned/other-language; recency-weight, Q4).
2. **Close-read the raw BODY** (subagent for large corpora) for operational patterns at every level: structure (open/pivot/close) · paragraph & rhythm · sentence patterns · diction/lexicon · rhetorical devices · titles. Each rule = **directive + verbatim quote**.
3. **Anti-slop / "never" list:** generic AI tells to forbid + the writer's own disavowals.
4. **Naturalness layer (D21 — overrides the rest):** device-density limits, no manufactured similes, ground-in-specifics, ban the escalating-triplet crescendo, allow unevenness, restraint over performance. Add the read-aloud test.
5. **Assemble** skill-consumable: rules + examples + a runnable **VOICE CHECKLIST** (binary items).
6. **Calibrate (step 8):** generate a *restrained* sample → writer reacts (dead-on / generic / "I'd never") → revise. Optional **blind test** (the Q1 proof). Gate before any skill builds on it.

## Use-mode (D21) — defensive
The profile's primary job is a **guardrail**: polish, flag AI texture, **preserve idiosyncrasy,
never sanitize**. The writing stays the writer's. Any drafting is rough scaffolding the writer
rewrites — never reader-facing prose.

## Output
`newsletter-context/voice-profile.md` — rules + examples + checklist + naturalness layer. The substrate for the
guardrail/assist skills (step 9).

## ▶ Milestone 2 complete — hand off (the "you're set up" moment)
With the Brief + Voice Profile both done, the setup bulk is finished. Tell the writer clearly that their context is **ready to use**, and how to use it from now on:
- **Where it lives:** this folder's `newsletter-context/` — `newsletter-brief.md` (source of truth), `voice-profile.md` (how you write), plus `seed-context.md` / `identity-stated.md` / `corpus/` (raw inputs).
- **Every time you write:** open Claude Code **in this folder** (that's what points the skills at *your* context), then use **`brainstorm`** (angle) → *you draft* → **`editorial-review`** (substance) → **`polish`** (voice). 
- **Keep it sharp:** **`compound`** at session end to fold in what you learned; **`prune`** occasionally to keep it lean; if the newsletter pivots, say so and we update the Brief first.

---
_**Language:** if the newsletter isn't in English, converse with the writer — and write outputs — in the newsletter's language._