Postmortem Assumptions and Unknowns Extractor Prompt

You are a senior SRE reviewing a postmortem draft for hidden assumptions and unverified claims. Your goal is to make the analysis honest about what is known versus believed, without blaming the authors or responders.

I will paste:
[POSTMORTEM_DRAFT] — the writeup, including timeline, root-cause narrative, and action items.
[KNOWN_EVIDENCE] — optional: logs, metrics, or links that back specific claims.

Do this:

1. Extract ASSERTED FACTS: pull each causal or factual claim the draft makes ("the retry storm caused the DB to fall over").
2. For each, classify as SUPPORTED (evidence cited), PLAUSIBLE-BUT-UNVERIFIED, or ASSUMPTION (stated as fact with no backing).
3. List HIDDEN ASSUMPTIONS: the unstated beliefs the narrative depends on (e.g., "the deploy and the outage are related because they were close in time").
4. List OPEN UNKNOWNS: questions the draft treats as settled but hasn't actually answered, and what evidence would resolve each.
5. Flag CONFOUNDERS: alternative explanations the draft dismisses or never considers.
6. Recommend the smallest set of VERIFICATION STEPS that would move the biggest assumptions to SUPPORTED.

Output as Markdown: a claims table (claim / classification / evidence-needed) plus the hidden-assumptions, open-unknowns, and confounders lists. Stay blameless: you are testing the analysis, not the analysts. Mark anything you infer beyond the text as [ESTIMATE]. A human review owner decides which assumptions must be verified before the postmortem is finalized; you only surface them.

Why this prompt works

Postmortems fail silently when a plausible story hardens into an accepted root cause without anyone checking whether the evidence actually supports it. The classic pattern is temporal correlation masquerading as causation: a deploy went out twelve minutes before the outage, so the deploy caused the outage, and every action item flows from that assumption. If the real cause was an unrelated dependency degrading in the background, the team ships fixes that prevent a failure that will never recur while leaving the actual fault untouched. This prompt is a dedicated adversary to that pattern, pulling apart what the draft knows from what it merely believes.

The three-way classification of asserted facts is deliberately more forgiving than a simple true/false split. Most claims in a good postmortem are PLAUSIBLE-BUT-UNVERIFIED rather than outright assumptions, and that is fine, so long as the team knows which ones they are and chooses consciously whether to verify them. Forcing every causal claim into the table makes the confidence level explicit instead of leaving it implied by fluent prose. The hidden-assumptions and confounders sections attack the harder problem: the beliefs that never appear as sentences at all, and the alternative explanations the narrative quietly excluded. These are exactly the things authors cannot see in their own draft, because they are the frame through which they wrote it.

Blameless framing is essential because this prompt is, structurally, a critique. Reviewing a colleague’s causal analysis and finding it rests on assumptions can read as an accusation of sloppiness if the language isn’t careful. By insisting the target is the analysis and not the analysts, the prompt keeps the exercise collaborative: everyone benefits from a postmortem whose conclusions survive scrutiny, and nobody should feel exposed for having written a first draft that made reasonable leaps. Good postmortems are iterative, and surfacing assumptions is part of the iteration, not a mark against the author.

The strongest guardrail is the instruction not to resolve the unknowns. An eager model will happily fill gaps with invented evidence or confidently pick a root cause, which would defeat the entire purpose by replacing the draft’s unverified certainty with the model’s unverified certainty. Requiring [ESTIMATE] markers and handing the verification decisions to a human review owner keeps the tool in its proper role: a lens that reveals where the analysis is standing on unverified ground, leaving the actual verification to people with access to the real evidence.