Check your AI skills before your agents run them.
Skills are the instructions your agents follow. The free CLI grades them, scans them for security problems, tests them and compiles them for five agents. Watch, on pre-order and not built yet, is planned to re-run those checks when a model, an agent or a tool you use changes.
npx skillhawk review Pre-order Watch
Watch: Planned: re-run these checks when things change.
Models, agents and the tools you install change. Today you can re-run these checks yourself. Watch is our plan to run them for you when something changes and send any fix as a pull request.
- New-model check: Today: re-run your skills' evals on a new model yourself, with your own API key. Planned for Watch: run them when a model ships and open a pull request with the drafted fixes.
- Supply-chain check: Today: scan the skills, MCP servers, hooks and scripts you use. Planned for Watch: re-scan them when they change upstream.
- Agent format check: Today: compile to five agents and check each file against its limits. Planned for Watch: re-check when an agent changes its format or limits, and open a pull request.
Build: Write a skill once. Run it in five agents.
Turn an interview, a runbook or a document into a skill with trigger tests, then compile it to each agent's own format.
- Create: Interview, document or session log in; a reviewed, secured and tested skill out.
skillhawk create --answers answers.json - Generate: Six generators (procedure, checklist, voice, playbook, decision, metric) filled from your own document.
skillhawk generate checklist --from runbook.md - Compile: Native files for Claude Code, Codex, Cursor, Copilot and Gemini, each checked for size, frontmatter and lost content.
skillhawk compile release-notes --to all
Check: Catch the errors before agents do.
Grade every skill A to F with line-level findings, and audit the skills, MCP configs, hooks and scripts around them.
- Review: Contradictions, vague wording, broken links, skills that crowd each other out. --fix applies the safe fixes.
skillhawk review skills/ - Secure: Permissions, MCP servers, hooks, dependencies and remote scripts. A non-zero exit makes it a CI gate.
skillhawk secure . --gate high - Clean: Scripts shipped inside skills: error handling, hard-coded paths, exec bits.
skillhawk clean scripts skills/
Improve: Learn from how your team really works.
Find the corrections people keep repeating, turn them into skill changes, and keep a change only when it measurably helps.
- Chat review: Reads your own agent sessions locally and proposes edits. Nothing changes without your approval.
skillhawk review --chats - Optimize: Cuts tokens and keeps the rewrite only if the grade and trigger cases hold.
skillhawk optimize notes-v2 - Radar: Corrections repeated across people become a drafted, tested skill, ready for a pull request.
skillhawk radar logs/ --team - A/B: Split sessions between two versions; promote B only at p < 0.05 with 30 sessions per arm.
skillhawk ab report deploy - Model watch: Re-run every skill's evals on a model you name and an older one, report the drops and draft fixes.
skillhawk watch run skills/ --model <new> --allow-model
Save: See where your tokens go.
Give agents a map to query instead of opening every file, see where the tokens go, and estimate what each skill saves.
- Map: Writes an index of files, symbols, imports and skills (map.json, MAP.md) that agents can query.
skillhawk map - Tokens: Tokens and estimated cost from Claude Code logs, per day, repo, skill and session.
skillhawk tokens --days 7 - Savings ledger: Estimated dollars saved per skill from corrections, redo turns and tokens, with its confidence stated.
skillhawk savings --statement out.md
Protect: Your skills are an asset. Keep them.
A private vault with roles and an audit trail, copies that name the person they were issued to, and access you can take back.
- Vault and roles: Versions, diffs and rollback; publishing must pass review and security (an org admin can override; it is logged).
skillhawk token create --org acme --role editor - Audit log: Pulls, publishes, rollbacks and access changes, hash-chained.
skillhawk audit verify - Signed skills: `skillhawk verify` fails if the files differ from what the org signed.
skillhawk verify deploy-safely --pub org.pem - Marked copies: Every SKILL.md copy is marked for the token holder it is issued to, so a leaked fragment can often be traced to who it was issued to.
skillhawk trace leaked.md --marks marks.json --secret $SKILLYARD_SECRET - Evidence pack: Signed report with the trace, pull history, audit chain and a drafted takedown notice.
skillhawk evidence leaked.md --org acme - Call-home link: A per-copy link that records when a leaked copy is opened: time, IP address and browser.
skillhawk org settings acme --beacon on --i-have-legal-review - Revocable streaming: Agents load skills from the vault over MCP; Skillhawk installs no skill files (the agent's own session logs may still hold text it loaded).
skillhawk connect claude --stream - Offboarding: Revoke a leaver's tokens and list every skill and version they pulled.
skillhawk offboard alice --org acme --out off.md - Exit interview: Capture what only they know as private draft skills, checked like any other.
skillhawk exit-interview pat --answers a.json - Governance export: Inventory, approval trail, access and integrity, labelled with the SOC 2, ISO 42001 and EU AI Act controls they support. Evidence for your audit; it does not make you compliant, and Skillhawk itself is not certified.
skillhawk governance export --org acme --framework all
Measured, not claimed
The scan, lift, reviewer and trace numbers come from runs you can repeat from the repository; the token week is one developer's own logs.
Skill library
277 public skills graded A to F and security-checked: 120 grade A, 36 grade B, 83 grade C, 33 grade D, 5 grade F. Browse the library.
Answers
Planned, not built
Leak-watch crawler, Stronger collusion resistance (merged copies already rank both colluders), Single sign-on, Slack alerts for vault events (Studio already posts review proposals to a Slack webhook), VS Code extension.