Write the runbook for your riskiest agent workflow

The prompt

Write a complete runbook for your single riskiest production agent workflow: prerequisites, repository instructions, identity, tools, model access, validation, security checks, approval, deployment, observability, rollback, evidence collection.

Why it's built that way

Each production workflow needs a repeatable runbook covering all of these — a workflow that only one person can operate from memory is not actually production-ready, whatever its test coverage looks like.

What to do with the answer

Have someone who didn't write the runbook try to follow it — any step they can't complete from the doc alone is the fix.