MCP Playbooks
Curated skills and workflows agents can run with run_playbook or solve_task.
Playbooks are markdown skills plus a mapped workflow (or local content path). Prefer run_playbook(skillId) over loading the markdown and inventing steps.
Agent contract
plan_task → solve_task | run_playbook → verify_taskHigh-value playbooks
| Skill id | Goal phrases | Workflow / mode |
|---|---|---|
ship-gate | ship gate, go live check | ship-gate-job |
fix-verify-ship-gate | verify ship gate after fix | ship-gate-job + verify_task |
fix-verify-security-headers | verify headers after fix | security-headers-job + verify_task |
fix-verify-social-preview | verify OG after fix | social-preview-audit-job + verify_task |
fix-verify-core-web-vitals | verify LCP/CLS after fix | core-web-vitals-job + verify_task |
fix-verify-seo-audit | verify seo after fix | full-seo-audit + verify_task |
fix-verify-email-auth | verify SPF/DKIM after fix | email-auth-security-job + verify_task |
frontend-supply-chain | SRI, mixed content, CORS | frontend-supply-chain-job |
fix-verify-frontend-supply-chain | verify supply chain after fix | frontend-supply-chain-job + verify_task |
auth-token-hygiene | decode jwt, token hygiene | auth-token-hygiene-job |
developer-ship-checklist | before deploy, prod readiness | developer-ship-checklist-job |
seo-site-audit | full seo audit | full-seo-audit |
page-performance | core web vitals, LCP/CLS | core-web-vitals-job |
social-preview | open graph, twitter card | social-preview-audit-job |
web-security-audit | security audit URL | full-security-audit |
content-ship | ship this html, local content | local HTML (free unless enhance:true) |
dns-email-security | SPF DKIM DMARC | email-auth-security-job |
secrets-and-auth-hygiene | secret leak scan | secrets-hygiene-job |
Fix → verify loop
After applying fixes in the harness:
verify_task(goal, { url }, baselinePreviousReport)
# or async
verify_task(goal, input, baseline, async: true) → get_run(runId) → read resultStatus
# or
run_playbook("fix-verify-ship-gate", { url })Baseline can be a prior solve_task result, verify_task.after, a get_run payload, or a raw jobReport. Prefer verify_task when you still have that baseline — responses are deltas only. Score direction uses metric status (good / needs_improvement / poor), not raw “higher number wins.”
Credits
plan_task estimates credits (heuristic: ~2–3 per tool step). Exact billing matches REST tool costs (1–10 credits).