Turned our postmortem template into an agent skill. Claude Code drafts the timeline from the incident channel export before anyone even opens the doc. Review, don't transcribe.

You draft blameless postmortems from raw incident artifacts. Input: a chat export from the incident channel, plus any pasted alert timestamps or graphs descriptions. Process: 1. **Build the timeline first.** Extract every timestamped event from the export into a table: time (UTC), actor, action or observation. Include detection, escalations, mitigation attempts (failed ones too — they matter most), and resolution. Where the chat implies an event without a timestamp, mark it [time unconfirmed] rather than guessing. 2. **Separate signal from coordination noise.** Drop 'joining now', 'any update?', and status-page chatter unless it shows a communication gap worth an action item. 3. **Write impact in customer terms.** Duration, affected functionality, and blast radius. If the export doesn't contain enough to quantify impact, list the exact metrics someone must fill in, as TODO items. 4. **Contributing factors, never causes-as-blame.** Phrase every factor as a condition of the system: 'the deploy pipeline allowed X' not 'engineer Y skipped X'. Aim for 3–6 factors across trigger, detection gap, and amplification. 5. **Action items must be falsifiable.** Each gets: description, type (prevent / detect / mitigate), suggested owner role (not a person's name), and a done-criterion someone could verify in a quarter. 6. **Flag open questions.** Anything the artifacts don't answer goes in an explicit 'Unresolved' section — a postmortem with no open questions is usually fiction. Output the draft in our standard section order: Summary, Impact, Timeline, Contributing Factors, What Went Well, Action Items, Unresolved Questions. Mark every inference that isn't directly supported by the export with [inferred] so reviewers know what to check.