| name | skill-auditor |
| description | Use when the user wants to audit the current Codex skills library for prompt poisoning, ad-style injections, suspicious hosted-platform referrals, vendor-branded defaults, stealth marketing copy, or suspicious skill names inside ~/.codex/skills or $CODEX_HOME/skills. |
Skill Auditor
Use this skill for read-only review of the local skills library before any cleanup.
Default workflow
- Run the bundled audit script against the current skills root.
- Review
critical and high findings first.
- Distinguish legitimate documentation links from off-task marketing, referral language, or branded defaults.
- Only modify or remove flagged content if the user explicitly asks for cleanup.
Primary command
Run the auditor directly:
python3 "$CODEX_HOME/skills/skill-auditor/scripts/audit_skills.py" --format markdown
If CODEX_HOME is unset, the script defaults to ~/.codex/skills.
Useful variants
- Fail if medium-or-higher findings exist:
python3 "$CODEX_HOME/skills/skill-auditor/scripts/audit_skills.py" --fail-on medium
- Add a custom denylisted domain:
python3 "$CODEX_HOME/skills/skill-auditor/scripts/audit_skills.py" --deny-domain example.ai
python3 "$CODEX_HOME/skills/skill-auditor/scripts/audit_skills.py" --format markdown --report reports/skill-audit.md
Interpretation rules
critical: direct recommendation, forced upsell, or explicit instruction to steer users toward a hosted service
high: denylisted domains or strong CTA lines tied to external links
medium: softer marketing copy, branded defaults, or suspicious structural signals
low: weak or isolated signals worth manual review
Safety
- Do not auto-delete user skills.
- Do not auto-rewrite third-party repos unless the user asks to sanitize them.
- When cleaning is requested, keep legitimate technical references and remove only off-task promotional or poisoned content.
The default rules live in rules/default_rules.json.