Scribe Plugin - Contributor Guide

What This Is

Scribe is a Claude Code plugin that orchestrates specialist agents to produce high-quality longform writing. It catches AI-default writing patterns before they ship.

Structure

scribe/
├── .claude-plugin/     # Plugin manifest (plugin.json, marketplace.json)
├── agents/             # Flat directory, auto-discovered by Claude Code
├── commands/           # /scribe:* slash commands
├── skills/             # Scribe pipeline and focused humanize skill
├── anti-patterns/      # Base + research-sourced AI-ism library
├── state/              # Append-only audit log (state/draws.jsonl)
└── tools/              # Draw simulator and validators

Conventions

• All internal paths use ${CLAUDE_PLUGIN_ROOT}/ prefix
• Agents are flat in agents/ (no subdirectories - breaks auto-discovery)
• Anti-patterns at repo root (not nested in skills/) because many agents read them
• Specialists use Haiku (fast, cheap, constrained output)
• Implementation agents use Sonnet (code generation)
• Conductor, voice-critic, and critique-synthesizer use Opus (complex judgment, synthesis)
• Conductor never writes or critiques - only coordinates

Focused Humanize Skill

${CLAUDE_PLUGIN_ROOT}/skills/humanize/SKILL.md provides /scribe:humanize for direct edits of existing prose. It runs in the current conversation without the conductor, pattern draw, audit log, or Sameness Score. Keep this separate from the existing pipeline commands. Use the writer's voice and preserve claims instead of enforcing blanket pattern substitutions.

Adding a New Agent

1. Create agents/your-agent.md with YAML frontmatter (name, model, description). Set model to a tier alias (haiku, sonnet, opus), never a pinned model ID
2. Reference files using ${CLAUDE_PLUGIN_ROOT}/path/to/file
3. Define clear inputs, outputs, and constraints
4. If the agent participates in the pipeline, update skills/scribe/references/orchestration/pipeline.md

Adding a New Anti-Pattern

Community-sourced patterns go in anti-patterns/sourced-from-research/ via the research-scout agent. Base patterns in anti-patterns/language.md, structure.md, or rhetoric.md are manually curated.

Each new pattern needs a manifest.json entry with id, categories, severity_default, source_file, anchor, personas, site_signals.

Adding a New Persona

Create a file in skills/scribe/references/personas/ following the existing format. Update skills/scribe/references/personas/README.md and skills/scribe/SKILL.md.

Code Style

• No em-dashes in any content (this rule applies recursively - scribe must not exemplify what it flags)
• Writer-to-writer voice: specific, direct, no marketing register
• Conventional commits (feat:, fix:, docs:)
• Keep agent prompts under 100 lines unless the role genuinely requires more