stas2-git Agent Incident Retrospective
Use this skill to turn agent failures into improved future behavior.
Workflow
- State the incident in one sentence.
- Reconstruct expected behavior versus actual behavior.
- Identify failure type:
- intent misunderstanding,
- context drift,
- wrong/missing tool,
- weak skill trigger,
- missing verification,
- unsafe authority,
- prompt injection,
- insufficient eval,
- bad recovery.
- Find the repair surface: instruction, skill, reference, script, tool schema, guardrail, approval gate, or eval.
- Add a regression prompt/test that would have caught it.
- Propose a small fix; avoid broad "be careful" rules.
References
- Read
references/retrospective-template.mdfor root-cause and action-item templates.
Evaluation Prompts
- Positive: "The agent sent the wrong file; make a retro and prevention plan." Expected: root cause, boundary, regression eval.
- Positive edge: "Nothing bad happened, but the trace looked risky." Expected: near-miss retrospective.
- Negative: "Apologize for a typo." Expected: no incident retro unless future prevention is requested.