AgentSkillsCN

postmortem-and-learning-extractor

在重大事件、故障或项目里程碑之后，提炼可复用的经验教训，用于更新技能体系，避免未来重蹈覆辙。

SKILL.md
--- frontmatter
name: postmortem-and-learning-extractor
description: Extracts reusable lessons after incidents, failures, or project milestones. Use to update skills and prevent future mistakes.
triggers: [incident-resolution, mission-failure, project-completion]
outputs: [postmortem-report, skill-refinement-proposal]

Postmortem and Learning Extractor

Purpose

Converts failures into fuel for the system's evolution. It ensures that the same mistake is never made twice by codifying lessons into skills, specs, or orchestration rules.

When to use this skill

  • •After an incident or production bug
  • •After a failed migration attempt or rejected PR
  • •After the completion of a major delivery to document "what went right"

Extraction Steps

  1. •Identify Root Causes: Use the "5 Whys" method to get past surface-level symptoms.
  2. •Separate Failure Types:
    • •System Failure: A skill or orchestration rule was missing or flawed.
    • •Human Failure: Approval was given to a flawed plan or spec.
  3. •Extract Reusable Rules: What instruction could be added to a SKILL.md to prevent this?
  4. •Update Specs or Skills: Trigger the appropriate update skill (e.g., skill-evolution-engine).

Decision Tree

mermaid
flowchart TD
    A[Analyze Incident] --> B{Root Cause Found?}
    B -->|No| C[Deep Dive: 5 Whys]
    B -->|Yes| D{Skill-Based?}
    D -->|Yes| E[Propose Skill Update]
    D -->|No| F{Process-Based?}
    F -->|Yes| G[Update Orchestration Rules]
    F -->|No| H[Flag for Human Training]
    C --> B

Review Checklist

  1. •Blamelessness: Does the report focus on "What" and "How" rather than "Who"?
  2. •Actionability: Are the recommendations concrete enough to implement?
  3. •Breadth: Does the lesson apply to other modules or projects?
  4. •Verification: How will we know if this lesson prevents a future failure?

How to provide feedback

  • •Be specific: "The postmortem for incident #104 identifies 'bad code' but not the missing boundary guard check."
  • •Explain why: "Attributing it to 'bad code' doesn't help the AI prevent it in the next task."
  • •Suggest alternatives: "Add a rule to implementation-boundary-guard to check for recursive depth."

Blameless, actionable, reusable.