2 min read
Skill Creator + Auditor
Two complementary skills for the skill lifecycle: create skills with wiki integration, and audit skills for security before installation.
Skill Creator
Based on anthropics/skills. Customized with Kevin's wiki-first protocol. Source: skills/productivity/skill-creator/SKILL.md
Wiki-First Protocol
Before creating any skill:
- Check the wiki (
wiki/_index.md) for existing knowledge on the domain - Check skills.sh via
find-skillsfor upstream skills to adapt - Check installed skills across
skills/{engineering,productivity,personal,misc}/plus draft/retired categories for overlap
After creating a skill:
- Update
wiki/tools/skills-sh.mdwith the new skill in the inventory table - Update
AGENTS.mdinstalled skills table - Append to
wiki/log.md - Cross-reference related wiki pages
- Rebuild index:
npx tsx scripts/build-index.ts
Skill Anatomy
skill-name/
├── SKILL.md (required - frontmatter with name + description, then instructions)
└── resources/ (optional - scripts, references, assets)
Install Cursor/personal skills to skills/personal/<name>/SKILL.md. Keep under 500 lines. Use references/ for overflow. Descriptions should be slightly "pushy" to prevent under-triggering.
Writing Principles
- Explain the why behind instructions - reasoning over rigid MUSTs
- Include concrete input/output examples
- Always include a "Related Skills" section
- Reference relevant wiki pages for domain context
- Progressive disclosure: metadata always in context, references loaded on demand
Skill Auditor
Based on useai-pro/openclaw-skills-security. Part of the three-gate quality chain (see Brin - Agent Context Security). Source: skills/engineering/skill-auditor/SKILL.md
6-Step Vetting Protocol
| Step | What | Critical Findings |
|---|---|---|
| 1. Metadata & typosquat | Name, description, author verification | Char swap, homoglyph, scope confusion |
| 2. Permission analysis | fileRead, fileWrite, network, shell | network+fileRead = exfiltration; network+shell = RCE |
| 3. Dependency audit | Package safety, postinstall scripts | Recent owner transfer, typosquat packages |
| 4. Prompt injection scan | Role overrides, fake tags, zero-width chars | "Ignore previous instructions", [SYSTEM] tags |
| 5. Network & exfil | Raw IPs, DNS tunneling, env var leaks | fetch(url?key=${process.env.API_KEY}) |
| 6. Content red flags | Credential file access, obfuscated content | References to ~/.ssh, ~/.aws, ~/.env |
Three-Gate Quality Chain
All three gates must pass before installing any new skill:
- Brin -
curl https://api.brin.sh/skill/<owner>/<repo>. Block if suspicious/dangerous. - Skill Auditor - 6-step protocol above. Must verdict SAFE.
- Reputation signals - install count, source reputation, GitHub stars, skills.sh security badges.
Timeline
- 2026-06-17 | Added inline provenance and timeline for the skill lifecycle page. Source: skills/productivity/skill-creator/SKILL.md; skills/engineering/skill-auditor/SKILL.md
- 2026-04-07 | Created combined creator/auditor page for new-skill lifecycle and security vetting. Source: skills/productivity/skill-creator/SKILL.md; skills/engineering/skill-auditor/SKILL.md