Works best in
Claude · ChatGPT · Notion AI (paste draft into doc)
Prompt
You are an incident commander writing a blameless postmortem first draft from raw notes. Do not assign personal blame. Focus on systems, signals, and decisions.
Raw notes / timeline scraps:
<<<
{PASTE NOTES}
>>>
Known facts (fill what you know; leave blank if unknown):
- Service / customer impact:
- Severity (Sev1–4 or your scale):
- Detected at / by:
- Mitigated at / how:
- Resolved at:
- Customer-facing comms sent? (Y/N + link):
Deliverables (markdown):
1) Title: short incident name + date.
2) Summary: 3–5 sentences — what broke, who felt it, how it ended.
3) Impact: users/revenue/SLO burn — mark estimates as ESTIMATE.
4) Timeline: table (UTC timestamps | event | evidence). Gaps labeled UNKNOWN.
5) Contributing factors: 3–6 bullets (latent conditions + triggers). No names as root cause.
6) What went well / what went poorly: 3 bullets each, operational not personal.
7) Action items: table (action | owner role — not a person if unknown | due | priority P0–P2 | prevents recurrence? Y/N).
8) Open questions: list anything still unverified.
9) Comms draft: 5-line status update suitable for customers or execs (separate internal vs external if needed).
If notes conflict, keep both versions under CONFLICTS. Never invent timestamps, dollar impact, or root cause.
How to use
Dump pager alerts, Slack threads, and runbook links into the notes block; run in Claude (long context) or ChatGPT. Human IC must correct timestamps and impact before sharing outside the incident channel. Money/risk: wrong public impact numbers create legal and customer-trust damage; wrong "root cause" creates blame culture and bad fixes. Failure modes: model fills UNKNOWN with plausible fiction — reject any invented time or dollar figure. Skip for Sev4 noise with no customer impact, or when legal already owns the write-up. Do not use as the final published RCA without on-call review.