skill-review-audit（技能审核审核）

Skill Review & Audit

Produce a systematic, multi-dimensional review of any skill directory (a SKILL.md plus optional scripts/, references/, assets/, and install metadata).

Non-Mutating Constraint (Default)

• Default is read-only: do not create, delete, or modify any files (including apply_patch, sed -i, overwriting configs, auto-fixes, git commit, etc.).
• If the user wants changes applied, ask for explicit human authorization first and wait for a clear “yes, apply these edits” (or equivalent) before touching the filesystem.
• You may still propose fixes as text (recommendations or patch snippets), but do not apply them without authorization.

Outcomes

• A clear description of what the skill teaches and *what it does not*.
• A map of tooling + side effects the skill may cause when followed (commands, network, file writes, permissions).
• A risk assessment (security, privacy, safety, supply chain) with mitigations.
• A quality assessment (correctness, completeness, maintainability, UX) with prioritized improvements.
• Optional scoring using references/scoring-rubric.md.
• A report formatted using references/report-template.md.

Inputs To Request (If Missing)

• Skill identifier: name and/or filesystem path to the skill directory.
• Target agent environment (e.g. Codex CLI / Claude Code / other) and any constraints (offline, no web, sandboxed, etc.).
• Intended usage context (what kinds of user prompts should trigger it; what “done” looks like).

Workflow (Do In Order)

0) Scope The Review

• Confirm whether the review is (a) informational only (read-only) or (b) includes proposing patches (still read-only unless explicitly authorized to apply).
• Define what “safe enough” means for the target environment (network allowed? can write files? secrets present?).
• Restate the Non-Mutating Constraint and ask for authorization if the user requests edits to be applied.

1) Inventory & Provenance

1. List the full directory tree and file sizes.
2. Identify install/provenance files (common: .openskills.json, package.json, pyproject.toml, git submodule markers).
3. Record:

- Skill root path - Total file count - Presence of scripts/, references/, assets/ - Any external source URL + install timestamp (if present)

1. Flag anything unexpected (executables, binaries, obfuscated blobs, huge files, symlinks pointing elsewhere).

Optional helper: run scripts/scan_skill.sh (read it first; it is intended to be read-only). Note: scan_skill.sh may surface sensitive strings (e.g., tokens, private keys) depending on the target directory. Treat its output as sensitive; redact before sharing.

2) Trigger Contract (Frontmatter Audit)

Read SKILL.md YAML frontmatter and assess:

• Name: unique, stable, correctly scoped (not overly broad).
• Description (primary trigger): includes concrete triggers/symptoms; avoids vague “does everything”.
• False positives/negatives: prompts it might match incorrectly vs fail to match.
• Overlap risk: collisions with other skills (same domain, similar trigger phrases).

Output: “Trigger Strength” rating + rewrite suggestions.

3) Capability Model (What It Teaches)

Extract and summarize:

• Core tasks it claims to support.
• Preconditions and assumptions (tech stack, tools installed, access levels).
• Deliverables (expected outputs, formats, artifacts).
• Anti-scope (“When NOT to use”) and limitations (explicit or missing).
• Degree-of-freedom: where it’s prescriptive vs heuristic.

If the skill includes references, don’t assume the main SKILL.md is complete—sample or selectively read reference files to confirm scope.

4) Tooling & Side-Effects Map

Build a table of *everything the skill instructs the agent to do*:

• Shell commands (including examples).
• Network access (curl/wget, HTTP clients, package installs, API calls).
• File system writes (what paths, destructive operations, deletes).
• Privilege/permissions (sudo, elevated access, credential usage).
• External dependencies (libraries, CLIs, SaaS).

For each, record: intent, required permissions, risk, and safe alternatives (sandbox, dry-run, allowlists).

5) Security / Privacy / Safety Risk Assessment

Use references/risk-taxonomy.md to assess:

• Prompt injection exposure (especially if the skill fetches external content).
• Command injection risks (string interpolation into shell; unsafe copy/paste patterns).
• Destructive operations (rm -rf, overwriting, migrations, irreversible actions).
• Secrets handling (API keys, env vars, logs, redaction).
• Supply-chain risks (install scripts, unpinned deps, untrusted sources).
• Data exfiltration (uploading files, telemetry, “paste logs here” patterns).

Output: severity × likelihood per risk + mitigations + “safe-by-default” recommendations.

6) Quality & Correctness Review

• Verify examples for internal consistency (missing imports, wrong prop precedence, mismatched ARIA ids, etc.).
• Check for missing edge cases (cancellation, cleanup, concurrency, accessibility, i18n).
• Check “progressive disclosure” quality: is SKILL.md lean and navigational, with details in references/?
• Check for outdated or unstable advice (versions, APIs likely to change); suggest pinning and dates.

7) Maintainability & Operational Fit

• Structure: clear headings, searchable keywords, minimal duplication.
• Update strategy: versioning, ownership, changelog expectations (even if no file).
• Testability: are scripts tested? is there a validation workflow?
• Portability: OS assumptions, shell assumptions, tool availability.

8) Improvement Plan (Prioritized)

Provide:

• Quick wins (low effort / high impact).
• Structural changes (refactor into references, add scripts, add checklists).
• Safety hardening (guardrails, confirmations, allowlists).
• “Definition of Done” for the next iteration.

9) Produce The Report

Use references/report-template.md and keep:

• Facts separated from recommendations
• Explicit uncertainty markers when you did not verify something
• Concrete examples (commands, paths, prompts) where useful

Red Flags — Stop And Re-check

• Only read SKILL.md and ignored scripts/ / references/.
• Listed risks without mapping concrete commands/side effects.
• No provenance/supply-chain notes.
• No severity/likelihood distinction (everything “risky”).
• Suggested running scripts you did not read.
• Gave recommendations without tying them to a specific observed gap.

Deep Checklist (Optional)

If you need a more exhaustive pass, use references/review-checklist.md and score with references/scoring-rubric.md.