runbook-creator

Say "save what I just ran" and the Skill picks up the auto-persisted script from the session, lints it, checks name collisions, writes the source to the team-shared runbooks directory, and deploys a copy to .claude/workflows for by-name invocation.

Use a runbook only for genuine fan-out, multi-stage orchestration, or resumable iteration. One-off tasks should just run directly, event-triggered behavior belongs in hooks, and sequential procedures without fan-out belong in a plain skill.

No. Workflow script bodies are sandboxed to standard JS built-ins only; Date.now(), Math.random(), and fs or shell calls throw at runtime. All file and shell work must happen inside spawned agent() calls that return structured results.

The Workflow tool can deliver args as a JSON string or a parsed object depending on invocation path. Always start with the normalization line const parsedArgs = typeof args === 'string' ? JSON.parse(args) : args, then validate parsedArgs before any agent call.

Runbooks are user automation with no proof-of-behavior guarantee and must never invoke the locked harness pipeline or write harness machinery paths. Code changes needing verified behavior should go through the locked pipeline via start-task instead.