ops-runbook

Write an operational runbook any qualified operator can execute under pressure, with verification, failure paths, and escalation.

SKILL.md

GitHub

Runbook

When to use

Use this to write a step-by-step operational runbook for a recurring task or incident response, so any qualified operator — including on-call under pressure — can execute it correctly without tribal knowledge.

Not for: documenting a multi-role business process with a RACI (use ops-process-doc), or high-level policy. A runbook is one operator, one task, executable top-to-bottom.

Method

  1. Define task and trigger. State what the runbook covers, when it runs (schedule, alert, or condition), and the expected outcome.
  2. List prerequisites. Access, credentials, tools, and preconditions that must be true before starting.
  3. State safety and impact. Note what this affects, who to notify, and any destructive steps. Decision: flag every irreversible step with a confirm-before-proceeding gate.
  4. Write the procedure. Numbered, copy-runnable steps. Per step: the action, the expected result, and how to tell it worked.
  5. Add verification. Specify the checks that confirm end-to-end success.
  6. Cover failure paths. For likely failure points, give symptom -> recovery/rollback action; include an escalation contact. Decision: if a step can fail silently, add an explicit check rather than assuming success.
  7. Record metadata. Owner, last-tested date, review cadence — a runbook untested in 90 days is suspect.

Example

Title: Clear stuck payment queue. Trigger: alert queue_depth > 500 for 10m. Prereqs: prod DB read/write, PagerDuty access.

  1. Check queue depth (expected: >500). 2. Identify the blocking message ID (expected: one poison message). 3. Move it to the dead-letter queue (destructive — confirm ID first); expected: depth drops. Verify: depth < 50 within 5 min, no new errors. **Failure:** depth stays high -> restart the consumer; still high -> escalate to payments on-call (@payments-oncall).

Pitfalls

  • Assuming context the reader lacks. On-call at 3am has none of your mental model; spell out prereqs and expected results.
  • No expected result per step. Without "you should see X," an operator can't tell success from silent failure.
  • Untested runbook. Steps that were never rehearsed drift from reality; stamp and re-test on cadence.
  • Missing failure paths. The happy path is the easy half; the value is knowing what to do when step 3 errors.

Output format

Header:         title | purpose | trigger | severity
Prerequisites:  access | tools | preconditions
Impact:         what it affects | who to notify
Procedure:      numbered steps, each with expected result
Verification:   end-to-end success checks
Troubleshooting: symptom -> action table
Escalation:     who | when | how
Metadata:       owner | last tested | review cadence