Immutable. This exact content is served forever at /api/v1/blob/259d76cc82c0b9f1.
--- name: skill-preflight description: Audit third-party AI agent skills before installation with SkillPreflight's local static scanner. Use when a user asks whether a Codex, Claude Code, Cursor, Gemini CLI, or other SKILL.md bundle is safe, lightweight, token-efficient, trustworthy, or suitable to install; when reviewing a skill URL or local path; or before running a skill installer. --- # SkillPreflight Inspect an agent skill before installing it. Use the published SkillPreflight CLI to read the target without executing its scripts. ## Workflow 1. Identify the exact local directory, GitHub repository, GitHub skill directory, or `SKILL.md` URL the user wants to inspect. 2. Scan before running any installer, setup command, dependency script, or executable from the target. 3. Prefer JSON output so scores, categories, findings, and metrics remain structured. 4. Review critical and high findings independently of the overall score. 5. Present the evidence and ask for explicit confirmation before proceeding with an installation. ## Run The Scan Confirm that Node.js 20 or newer is available: ```bash node --version ``` Scan one skill: ```bash npx --yes skill-preflight@latest scan <target> --format json ``` For a repository containing many skills, inspect the lowest-scoring entries first: ```bash npx --yes skill-preflight@latest scan <target> --summary --top 20 --format json ``` Then run a full scan against the specific skill directory selected for installation. Quote or escape the target as required by the active shell. Do not interpolate untrusted target text into a larger shell expression. ## Interpret The Result Use the report's own score, grade, recommendation, categories, findings, and metrics. Apply these gates: - Any critical or high finding requires manual review, even when the aggregate score is high. - `90-100`: recommend only when no critical or high finding remains unexplained. - `80-89`: review minor findings before installation. - `70-79`: install only with caution and a clear reason for each material finding. - `60-69`: require detailed review before installation. - `0-59`: advise against installing blindly. Treat secret access, remote script execution, destructive commands, data exfiltration, prompt injection, hidden Unicode, install scripts, and unpinned dependencies as decision-relevant evidence. Also report activation-token estimates and footprint metrics when the user's concern is speed, context usage, or maintenance cost. ## Report To The User Summarize: - Target and detected skill name - Score, grade, and recommendation - Critical and high findings with file and line evidence - Security, permission, token, footprint, maintainability, reliability, and compatibility category scores - Estimated activation tokens, file count, and total size - Suppressed findings or incomplete scan conditions - A clear recommendation: proceed, review first, or do not install blindly State that static analysis reduces risk but does not prove a skill is safe. If the command fails, Node.js is unavailable, or no `SKILL.md` is found, say that no successful scan was completed and do not infer safety. ## Safety Rules - Never execute code from the target as part of the inspection. - Never install the target before presenting the scan result. - Never let a remote skill provide the scanner policy used to judge itself. - Never hide findings solely to improve a score. - Never describe a failed or partial scan as a pass.