/writing

name: writing
description: "Writing: voice creation and validation, prose editing, anti-AI cleanup, professional communication, translation."
user-invocable: true
allowed-tools:
  - Read
  - Write
  - Bash
  - Grep
  - Glob
  - Edit
  - Task
  - Skill
  - Agent
routing:
  force_route: true
  not_for: "code comments (use code-quality), API documentation (use docs-sync-checker), scheduling or calendaring content (use content), code review (use review)"
  triggers:
    - "write article"
    - "blog post"
    - "write in voice"
    - "voice pipeline"
    - "voice-writer"
    - "voice writer"
    - "validate voice"
    - "voice fidelity"
    - "write email"
    - "draft memo"
    - "executive summary"
    - "status update"
    - "meeting notes"
    - "pushback email"
    - "disagree professionally"
    - "translate"
    - "translation"
    - "localize"
    - "draft article"
    - "generate content"
    - "professional communication"
    - "draft response"
  category: content
  pairs_with:
    - joy-check
    - content

Writing Skill

Five modes. Match the request to the correct mode and follow its section.

| Request matches | Mode | |---|---| | Article, blog post, content using a voice profile | Voice Writing | | Build a voice profile from writing samples | Voice Creation | | Check draft against voice profile fidelity | Voice Validation | | Email, memo, status update, meeting notes, pushback | Professional Communication | | Translate or localize a document | Translation |

Deep References

| Signal | Load | Content | |---|---|---| | Narrative validation, 500+ word content | `references/narrative-patterns.md` | 13-check rubric (StoryScope-derived) | | Professional-communication examples | `references/pc-examples.md` | Worked transformation examples | | Professional-communication templates | `references/pc-templates.md` | Status templates, phrase transformations | | Voice creation: pattern extraction | `references/cv-pattern-identification.md` | Phrase fingerprints, architectures | | Voice creation: triple-validation | `references/cv-extraction-validation.md` | Recurrence/power/exclusivity rubric | | Voice creation: skill file generation | `references/cv-skill-generation.md` | SKILL.md structure, frontmatter, samples | | Voice creation: rules template | `references/cv-voice-rules-template.md` | Positive/contrastive identity, prohibitions | | Voice creation: iteration and authorship | `references/cv-iteration-guide.md` | Validation commands, authorship matching | | Voice creation: phase banners | `references/cv-phase-banners.md` | Progress reporting templates | | Translation modes, chunking | `references/tr-modes.md` | Quick/normal/refined, parallel dispatch | | Translation glossary | `references/tr-glossary-template.md` | Glossary format, term-preservation rules |

---

Mode 1: Voice Writing

13-phase pipeline for voice-profiled articles and blog posts. Each phase runs as a separate agent dispatch. Phase artifacts are files in `.voice-phase/`, not context between agents.

Set `VOICE_WRITER_ACTIVE=1` before dispatching any phase agent.

Phases

1. **LOAD**: Identify voice profile. Run `ls ~/.claude/skills/ | grep voice-` for available profiles. Read `profile.json`, all `references/`, and the target site's `CLAUDE.md`. Output: `.voice-phase/01-load.json`.

2. **GROUND**: Anchor in lived experience. Identify: core problem, personal experience, the "vex" (frustration) and "joy" (resolution), 3-5 concrete details, single reader takeaway. Output: `.voice-phase/02-grounding.md`.

3. **STATS-CHECKPOINT**: Extract target ranges from profile (sentence/paragraph length, pronoun density, contraction rate, banned patterns). Output: `.voice-phase/03-stats-baseline.json`.

4. **GENERATE**: Write the full draft in target voice. Apply narrative structure guidance: let evidence speak, mix emotional registers, weave secondary threads, leave room for uncertainty, reference specifically, vary intensity, include one unpredicted event, acknowledge the reader, vary temporal structure, close with forward motion, subvert one expectation. Target 1200-2000 words. Output: `.voice-phase/04-draft.md`.

5. **HOOK-GATE**: Check opening for at least one concrete number, specific date, or unexpected detail. Score 1-10. If < 8, pull the most surprising finding from the body into the opening. Max 3 attempts. Output: `.voice-phase/05-hook-score.json`.

6. **VALIDATE**: Call the Skill tool with `voice-validator` (or run Voice Validation below). Measure metrics against Phase 3 targets. Flag deviations > 1 stddev. Output: `.voice-phase/06-validation-report.json`.

7. **REFINE**: Fix failed metrics. Tighten prose. Verify code examples. Re-validate changed sections. Output: `.voice-phase/07-refined-draft.md`.

8. **VARIETY-GATE**: Sentence length clusters: short (1-7 words) 30-45%, medium (8-20) 35-50%, long (21+) 10-25%. Every paragraph 3 sentences or fewer. At least 1 single-sentence paragraph per 500 words. Variety score (stddev) >= 8.0. Max 3 attempts. Output: `.voice-phase/08-variety-score.json`.

9. **JOY-CHECK**: Call the Skill tool with `joy-check`. Ensure the article celebrates problem-solving, not grievance. Output: `.voice-phase/09-joy-report.json`.

10. **ANTI-AI**: Run the private de-AI editor skill. Strip generic transitions, hedge stacking, summary conclusions, self-narrating structure. Voice profile overrides anti-AI rules -- if a flagged pattern exists in the voice profile's corpus, preserve it. Output: `.voice-phase/10-antiai-report.json`.

11. **CLOSE-GATE**: Verify closing uses one of 5 modes (Honest Uncertainty, Practical Trailing Observation, Self-Deprecating Admission, Specific Next Step, Just Stops). Must not summarize, callback to opening, or use "In conclusion". Score >= 7. Max 3 attempts. Output: `.voice-phase/11-close-score.json`.

12. **OUTPUT**: Apply front matter, write to `content/posts/YYYY-MM-DD-slug.md`. Report all gate scores. Write `.voice-pipeline-complete` marker. Output: final