界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
数据与研究 · SKILL.md

premortem

premortem

【数据与研究】Pre-mortem analysis that imagines a plan has failed, then works backward to identify causes and preventions. Use before launches, major decisions, or risky initiatives to surface hidden risks.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: premortem description: Pre-mortem analysis that imagines a plan has failed, then works backward to identify causes and preventions. Use before launches, major decisions, or risky initiatives to surface hidden risks. user-invocable: true

Pre-Mortem Analysis

Imagine the plan has completely failed, then work backward to identify what went wrong and how to prevent it.

Instructions

Set the scene: "It's [timeframe] in the future. This initiative was a complete disaster. Looking back, what happened?"

Generate failure scenarios without filtering for likelihood—get everything on the table first, then prioritize.

Output Format

The Plan Summarize what's being attempted and the success criteria.

Time Jump "It's [X months] later. This has failed completely. The outcome: [describe the disaster vividly]."

What Went Wrong

Generate 8-12 plausible failure causes across categories:

CategoryFailure ModeHow It Played Out
Execution[What failed][The story of how]
External[What failed][The story of how]
People[What failed][The story of how]
Technical[What failed][The story of how]
Assumptions[What failed][The story of how]

Risk Prioritization

Failure ModeLikelihoodImpactPriority
...High/Med/LowHigh/Med/Low1-5

Top 3 Risks & Mitigations

For each top risk:

  • Risk: [Description]
  • Early Warning Signs: What would indicate this is happening?
  • Prevention: How to reduce likelihood
  • Mitigation: How to reduce impact if it occurs
  • Owner: Who's responsible for watching this?

Pre-Mortem Insights What did this exercise reveal that wasn't obvious before?

Revised Confidence After this analysis, how confident are you in success? What would increase confidence?

Guidelines

  • Be vivid and specific—"the database corrupted" not "something went wrong"
  • Include uncomfortable possibilities (key person leaves, competitor moves, we were wrong)
  • Don't filter for "that won't happen"—the point is to surface hidden concerns
  • Assign real owners to mitigations
  • Look for single points of failure

$ARGUMENTS

我的收藏