plugin-builder

Plugin Builder

Create new Cowork plugins from scratch. Workflow: describe capability -> scaffold structure -> author with hard rules applied -> audit -> package -> install.

The skill enforces the production-tested conventions documented in this plugin's lib/ files. Skills built here pass audit out of the box and dispatch reliably under Cowork.

References

• lib/plugin-builder-workflow.md - detailed steps 3-9, idempotent re-run, guard rails, error cases, release maintenance
• lib/plugin-json-schema.md - minimum manifest schema, forbidden fields, naming
• lib/skill-md-frontmatter.md - required frontmatter, description format, trigger surface
• lib/cowork-hard-rules.md - no em-dash, curly placeholders, structured prompts, marker sections, idempotence, sandbox shadow, two-tier context model, condensed pitfalls table
• lib/build-and-audit.md - audit.py rules, build script, dist convention, CHANGELOG, versioning
• lib/release-maintenance.md - HARD RULES for post-publish maintenance: per-plugin versioning, description prefix, CHANGELOG-only-what-moved, per-bump checklist
• lib/user-prompts.md - structured prompt patterns
• lib/consent.md - consent-gate trigger phrases

Read all five hard-rule lib files (plugin-json-schema, skill-md-frontmatter, cowork-hard-rules, build-and-audit, release-maintenance) before generating any plugin content. They are load-bearing.

User input rules

Every user prompt in this skill MUST use the host client's structured input UI (single-pick, multi-pick, or text-input form). Never freeform plain-text yes/no questions. Per lib/user-prompts.md.

When to use

• Creating brand-new plugins with custom skills
• Packaging custom workflows
• Building org-specific capabilities
• Scaffolding skills beyond the bootstrap plugins in this marketplace

Do NOT use for:

• Modifying existing plugins (edit directly)
• Installing plugins from catalog (use /runesmith-core:install)
• Creating simple one-off scripts (use chat or a temp file)

Workflow overview

See lib/plugin-builder-workflow.md for detail on every step below.

1. Pre-flight read - load the four hard-rule lib files into context. If any are missing, the skill is broken; surface and stop.
2. Describe the plugin (structured prompts) - plugin name, one-line description, author, number of skills.
3. For each skill (structured prompts, per skill) - name, description seed, trigger phrases, workflow summary, mutations flag. See lib for full prompt list.
4. Generate plugin scaffold - create plugins/{plugin-name}/ with .claude-plugin/plugin.json, per-skill SKILL.md, empty lib/, and README.md. Plugin.json uses minimum schema only (name, version, description, author).
5. Generate each SKILL.md - YAML frontmatter (single-line quoted description, compatibility field, name matches folder), body with References / User input rules / Pre-flight / When-to-use / Workflow / Idempotent re-run / Guard Rails / Error cases. Mutation skills include consent gate; file-writing skills use marker-bounded sections.
6. Review and iterate - surface generated files, prompt looks-good / edit / cancel.
7. Run audit - scripts/audit.py if present. Do not proceed on