Tune Voice

Tune Voice — Per-Project Voice Extraction & Enforcement

Give a project a durable, reusable writing voice. The voice is a project artifact — it lives at <project>/ref/tone-guide.md, not in global memory — so each project keeps its own voice and drafts can be checked against it.

Fork-and-adapt lineage (design borrowed, storage owned): 9-dimension profile + anti-performative guardrails from entpnomad/tone-of-voice; voice-constant / tone-flex + validate-and-explain from brand-voice-enforcement; extract-from-samples from affaan-m/brand-voice.

Model: voice is constant, tone flexes

• Voice constant — identity, beliefs, audience, traits, terminology, banned phrases. Never changes across a project's copy.
• Tone flexes on three axes — formality · energy · technical depth — per channel (slide vs. Slack recap vs. cold email vs. golden-source doc).

A profile captures both. Enforcement holds the constant and picks the tone-flex for the channel at hand.

Step 0: Resolve the profile path

1. Read $PRAXIS_DIR/config/projects.yaml. Match --project <alias> against project keys (case-insensitive); if a name is given instead, fuzzy-match name fields. If no --project, infer from the working directory (which project's praxis: path is an ancestor of cwd), else ask.
2. Profile path = <project praxis root>/ref/tone-guide.md, where the praxis root is the praxis: value for that alias (resolve /praxis/..., /repos/... aliases via .claude/paths.md).
3. Leaf projects (a praxis: path with sub-projects, e.g. Slasheo's guide lives under ai-school-tech/ref/): accept an explicit --path <dir> override, or Glob for an existing **/ref/tone-guide.md under the project root and reuse it.
4. Guards: if $PRAXIS_DIR unset → ⚠️ $PRAXIS_DIR not set. export PRAXIS_DIR="$HOME/dev/praxis". If config missing → ⚠️ No project config at $PRAXIS_DIR/config/projects.yaml.

Modes

Default when args are just file paths: extract if the profile is missing, apply if it exists.

Extract — extract <files>

Read real sample copy and reverse-engineer the voice. Prefer the project's own artifacts (slides, docs, recaps, posts, READMEs). Reject generic exemplars as input — extract from what the project actually shipped.

Analyze across these signal dimensions (report only what's actually present):

• Sentence rhythm & length; compression vs. explanation balance
• Capitalization conventions; parenthetical frequency & function
• Question usage (frequency, rhetorical purpose)
• Claim sharpness; presence of numbers / mechanisms / "receipts"
• Transition mechanics; signature rhetorical patterns (e.g. X >> Y)
• Recurring metaphor families
• Stylistic prohibitions (what never appears)

Then write ref/tone-guide.md using the profile schema below. In the response, note which sample files fed the extraction and any low-confidence dimensions (too few samples).

Interview — interview

Fallback when no representative copy exists. Ask, one at a time, and write answers into the same schema:

1. Identity — background, stance (peer? authority? builder?)
2. Core belief — the unifying, slightly provocative principle ("the billboard")
3. Audience — the real person you write for, their motivation, address form (tu/vous, first-name)
4. Voice traits — 3–5 descriptors (direct, contrarian, technical…)
5. Topics — 3–5 recurring subject areas
6. Vocabulary — preferred vs. avoided, as pairs
7. Banned phrases — cringe reflexes to eliminate
8. Channels — per-channel tone-flex (formality/energy/technical depth) + format rules
9. Examples — 3–4 reference lines that hit the voice exactly

Apply — apply <file>

1. Load the profile (Step 0). If missing → suggest extract or interview first.
2. Determine the channel (from the file, the request, or ask) → select its tone-flex.
3. Rewrite/check the draft: hold the voice constant, flex tone for the channel.
4. Honor guardrail-override sections first — a profile may carry a Garde-fous / guardrails block that overrides imported rules; those always win (see Guardrails below).
5. Validate and explain: for each material change, name the rule that drove it (one line each) — never a silent black-box rewrite. Also surface drift: places the draft violates the profile you did not auto-fix, flagged for the user.
6. Respect strictness (see Settings): strict blocks/flags every violation; advisory fixes the clear ones and lists the rest.

Show — show [section]

Print the profile, or a named section (e.g. show vocabulary). If missing → offer to extract/interview.

Profile schema (ref/tone-guide.md)

---
title: Tone Guide — <Project>
source: <files the profile was extracted from, or "interview">
method: <extract | interview>
status: ref stable
---

# Tone Guide — <Project>

> One-line purpose: what prose this governs.

## Voice constant

1. **Identité** — stance, who's speaking.
2. **Croyance-socle** — the billboard principle.
3. **Audience** — who, motivation, address form.
4. **Traits de voix** — 3–5, each with a one-line tell.
5. **Sujets récurrents** — recurring themes.
6. **Vocabulaire** — **Utilise:** … / **Évite:** … (pairs) + signature rhetorical patterns.
7. **Phrases / réflexes bannis** — hard bans.
8. **We Are / We Are Not** — contrastive identity pairs (sharper than a flat ban list).

## Tone flex (per channel)

| Canal | Formalité | Énergie | Profondeur tech | Format |
|---|---|---|---|---|
| <slides> | … | … | … | one idea/slide, visual emphasis |
| <slack recap> | … | … | … | 2–3 sentences, catalytic |
| … | … | … | … | … |

## Exemples-étalon
- "<line 1>"
- "<line 2>"

---

## Garde-fous (override imported rules — prime TOUJOURS)
<Project- or language-specific hard rules that beat any borrowed guidance.
 E.g. FR prose: emphasis via visual not markdown; verbe conjugué > nominal;
 no FR+EN doublon; métaphore filée ≤ 2 occurrences.>

⚠️ **Dérives repérées** — known drift to watch (not auto-corrected).

Anti-performative guardrails (steal verbatim into every profile)

Bake these into the Garde-fous block unless the project overrides them:

• No definition by negation — don't lead with "this isn't X, isn't Y".
• Never punch down / strawman — write peer-to-peer.
• No manifesto rhetoric or bumper-sticker closings.
• Show, don't tell — experience-first, not prescriptive.
• Don't parrot the source when responding — add, don't restate.
• Open with impact — first line is a number, a claim, or a contrarian take.

Settings (optional)

Read <project>/ref/tone-voice.local.md if present:

strictness: advisory   # advisory (fix clear, list rest) | strict (flag all)
explain_edits: true    # narrate which rule drove each change

Absent → advisory, explain_edits: true.

Key behaviors

• Per-project, not global. Never write voice profiles to Claude memory — always to <project>/ref/tone-guide.md.
• Guardrails override imports. The Garde-fous block beats any borrowed rule, always.
• Explain, don't black-box. Every material edit names its rule; unfixed violations are surfaced as drift.
• Extract from real copy. Reject generic exemplars; profile what the project shipped.
• Voice constant, tone flexes. Hold personality/terminology fixed; adapt formality/energy/depth per channel.
• Config-driven paths. All resolution goes through $PRAXIS_DIR/config/projects.yaml; unknown project → warn and suggest adding it.