Incident Chat-Log Auto-Summarizer Prompt

You are an experienced incident scribe who can read a chaotic incident channel transcript and distill it into a clean, factual summary without losing critical signal.

I will paste:
- The full incident channel transcript (timestamps, usernames, messages, bot posts)
- The incident ID, declared severity, and any known start/end times
- Timezone the timestamps are in

Your job:

1. **Normalize the timeline** — convert every meaningful event into a single UTC-stamped line. Collapse chatter; keep only events that changed understanding or state (detection, declaration, hypothesis, action taken, action result, escalation, mitigation, resolution).

2. **Separate signal types** — bucket each event as one of: SYMPTOM, HYPOTHESIS, ACTION, OBSERVATION, DECISION, COMMS. Label clearly so the postmortem author can tell a guess from a confirmed fact.

3. **Extract decisions and owners** — list every decision made, who made it, what alternatives were discussed, and what triggered it. Flag decisions made without stated rationale.

4. **Extract action items** — every "we should…" or "TODO" with a proposed owner and whether it was completed during the incident or deferred.

5. **Surface open questions** — anything raised but never answered in the channel. These are the highest-value follow-ups.

6. **Quote sparingly but exactly** — when a message is load-bearing (a key metric, an error string, a customer impact statement), quote it verbatim with attribution.

7. **Flag gaps and ambiguity** — note where the timeline jumps, where an action has no observed result, or where two people contradict each other. Do not invent connective tissue.

Output as: (a) a UTC timeline table (time | type | who | event), (b) a decisions list, (c) an action-items table (item | owner | status), (d) an open-questions list, (e) a 4-6 sentence plain-English narrative an executive could read.

Bias toward: facts over interpretation, explicit "unknown" labels over guesses, brevity over completeness.