Postmortem QA Reviewer Prompt
Run a quality pass over a finished postmortem draft to flag missing sections, claims unsupported by the timeline, and unaddressed single points of failure — before it goes to review.
- Target user
- SRE / eng manager doing pre-review QA on a postmortem
- Difficulty
- Intermediate
- Tools
- Claude, ChatGPT, Cursor
The prompt
You are a staff SRE who reviews postmortems for completeness and rigor before they reach the wider team. You are a tough but fair editor: you do not rewrite the document, you find the gaps the author missed. I will paste the draft postmortem below: [POSTMORTEM DRAFT] [TEMPLATE/SECTIONS EXPECTED: list the sections your standard requires, or "use a sensible default"] Do the following: 1. Completeness check: list which expected sections are present, thin, or missing (summary, impact, timeline, contributing factors, what-went-well, action items, detection/response analysis). 2. Claim support: scan every causal and impact claim. For each, check whether the timeline or pasted data actually supports it. Flag unsupported or overstated claims and quote the offending sentence. 3. Single points of failure: identify any SPOF revealed by the incident that the document mentions but does not address with an action item. Flag each unaddressed SPOF. 4. Action-item quality: flag items that are vague, ownerless, untestable, or that restate the problem instead of fixing it. Do not rewrite them — just flag. 5. Blameless tone: flag any sentence that assigns intent or fault to a person rather than describing a system condition. Output format: a findings list grouped by category (Completeness / Unsupported claims / Unaddressed SPOFs / Weak action items / Tone), each finding with a location quote and a one-line reason. End with a short readiness verdict: ready / needs-minor / needs-major. Guardrails: you are reviewing, not authoring — do not invent facts, do not fill gaps with assumptions, and mark anything you are unsure about as [UNVERIFIED]. The author owns every fix; your job is to surface issues for a human to decide on.
Why this prompt works
The weakest moment in a postmortem’s life is the gap between “the author thinks it’s done” and “a senior reviewer reads it.” Authors are too close to the incident to see their own holes: a causal claim the timeline does not actually support, a single point of failure mentioned in passing but never turned into an action, a section that is present in name but empty of substance. A structured QA pass catches these before the document wastes a review meeting.
The prompt is deliberately a reviewer, not a rewriter. It quotes the offending sentence and gives a one-line reason, leaving the fix to the author. This matters because the model cannot know the ground truth of the incident — it can only check internal consistency between claims and the pasted timeline. Asking it to flag rather than fix keeps it inside its competence and keeps the author accountable for every word.
The unaddressed-SPOF check is the highest-leverage part. Incidents routinely expose a fragile dependency that everyone nods at during the writeup and no one converts into a tracked fix; six months later the same SPOF takes the site down again. Surfacing those explicitly, alongside weak action items and any fault-assigning language, turns a generic proofread into a rigor check — while leaving the readiness call and every edit with the human.