AgentSkillsCN

Premortem

事前回顾式审查——假设交付物在生产环境中遭遇失败，深入探寻失败原因并制定应对措施。

SKILL.md
--- frontmatter
description: Premortem review - assume the deliverable failed in production and find causes + mitigations
argument-hint: "What to review: PR link, file path(s), or short summary. You can paste the ADR/spec too."
disable-model-invocation: true
allowed-tools: Bash(git status:*), Bash(git diff:*), Bash(git show:*), Bash(rg:*), Bash(ls:*), Bash(cat:*)

Context

Deliverable to review: $ARGUMENTS

Repository quick context:

  • •git status: !git status --porcelain
  • •recent diff (if any): !git diff --stat

Your task (Premortem)

You are a reliability-focused senior reviewer. Run a premortem: Assume this deliverable shipped and failed badly in production.

Important constraints

  • •Do NOT rewrite the design.
  • •Do NOT list generic best practices.
  • •Focus on realistic failure modes and actionable mitigations.

Steps

  1. •Summarize what the deliverable changes (2-4 bullets). If unclear, state assumptions explicitly.
  2. •List Top 5 failure scenarios. For each scenario, include:
    • •Failure mode (what breaks)
    • •Trigger (what causes it)
    • •Blast radius (who/what is impacted)
    • •Detection (how we notice; metrics/logs/alerts)
    • •Mitigation (prevention + response)
  3. •Add a short section: "Hidden Couplings / Edge Cases"
    • •concurrency/race
    • •retries & idempotency
    • •partial failure
    • •ordering
    • •backward compatibility
    • •data migration / rollback
  4. •Add a short section: "Kill-switch & Rollback Plan"
    • •feature flag? safe fallback?
    • •rollback steps (including data)
    • •safe-to-disable behavior
  5. •Finish with "Minimal Fix List (P0/P1)"
    • •P0 = must fix before shipping
    • •P1 = should fix soon after

Output style

  • •Use headings and bullets.
  • •Keep it crisp. Prefer concrete checks/tests over opinions.