Boundary
Own this recurring job: audit local skill folders, estimate usage from local evidence, recommend cleanup direction, and flag security risks before install, execution, archive, or deletion.
Do not route here for:
- generic code security review outside a skill library
- one-off file cleanup with no skill audit
- creating a brand-new skill from scratch
- real destructive cleanup unless the user explicitly authorizes it
Default Workflow
- Run
scripts/run_skill_doctor.py <root> [more-roots]. - Generate
_skill_doctor_reports/<timestamp>/withreport.html,report.json,report.md, and clickable.commandactions. - Use Usage Evidence, Cleanup Rubric, and Security Rubric to interpret the scan.
- Return the HTML report path first, then summarize inventory, usage, cleanup, and security.
Operating Rules
- Stay read-only unless the user explicitly asks for cleanup actions.
- Treat usage frequency as an estimate derived from local evidence such as modification time and top-level inventory mentions.
- Do not present inferred frequency as telemetry or exact run counts.
- Escalate to
quarantinefor secret leakage, remote shell piping, private keys, or suspicious prompt injection or exfiltration behavior. - Prefer precise path-based evidence over generic statements.
- Separate hygiene issues from security issues. A stale skill is not automatically unsafe, and a recently modified skill can still be dangerous.
- Treat generated reports as local runtime artifacts. Do not commit
_skill_doctor_reports/back into the public repo unless the user explicitly wants a sanitized sample.
Outputs
Primary artifact:
- visual HTML report with summary cards, charts, recommendation modules, and per-skill action buttons
For each skill, provide inside the report:
- absolute path
- declared skill name and one-line purpose summary
- usage estimate:
active,warm,cold, orunknown - usage confidence:
low,medium, orhigh - cleanup level:
low,medium,high, orcritical - cleanup direction:
keep,repair,backup-then-archive,backup-then-delete, orquarantine - security level:
none,low,medium,high, orcritical - top evidence and findings that justify the recommendation
Order cleanup plans by security severity, cleanup level, and confidence that the skill is stale or disposable.
Resources
- Usage Evidence
- Cleanup Rubric
- Security Rubric
scripts/run_skill_doctor.pyscripts/scan_skills.pyscripts/skill_actions.pyevals/trigger_cases.jsonevals/semantic_config.json