Guidance for AI agents working in this repository. (Human contributors: see CONTRIBUTING.md and SKILL-AUTHORING-STANDARD.md.)
A library of professional Agent Skills — each a plain SKILL.md file that teaches an assistant to do one professional task well. Skills live in skills/<name>/SKILL.md; bundles in plugins/<bundle>/; the marketplace in .claude-plugin/marketplace.json.
- New skill / bundle: scaffold with
node scripts/new-skill.mjsornode scripts/new-bundle.mjs— don't hand-wireplugins/andmarketplace.json. - Every skill must pass the L3 / SkillSpec gate. Validate with
node scripts/skillcheck.mjs. - Keep counts honest. After adding skills, run
node scripts/check-drift.mjs(skill/bundle counts across docs must match) and regenerate derived files withnpm run check. - Don't hand-edit generated files (
web/skills.json,exports/**,web/catalog.html,web/coverage.html) — they're rebuilt. - Ship via a branch + PR, not direct pushes to
main.
Everything you write — skill bodies, docs, and the prose a skill produces — follows this. Full version: docs/prose-style.md; as a persona: output-styles/plain-honest-prose.md.
- Write for a specific reader and the action the text is for.
- Match voice to genre: docs and summaries stay neutral; only personal/persuasive writing shows a person. When unsure, neutral.
- Give every paragraph a checkable detail — but never invent a number, quote, date, or fact to sound specific. Hedge honestly or ask.
- Plain words over inflated ones; drop filler (leverage, robust, seamless, delve, game-changer).
- Cut the staged parts: no ceremonial openers ("Let's dive in", "It's worth noting"), no keynote rule-of-three, no "Great question!", no "In conclusion" restatement.
- Don't fake humanity: no manufactured typos, forced slang, random "But" openers, or sentences chopped just to vary length. Earned, or absent.
- Keep useful structure (headings, steps, tables); remove decoration, not architecture.
- Revise by reading and cutting — it should end shorter than it started.