# Learning Retrospective

> Recall verified lessons when a known failure recurs, diagnose repeated unproductive retries, or summarize a reusable fix after difficult work. Use for explicit retrospective requests including 复盘 and 总结教训. Ordinary first failures, planned repetitions, and exploration producing new evidence do not need a separate retrospective.

- Skill: `yingqi-han/learning-retrospective` (Agent Skill, multi-file: 31 files)
- Install (CLI): `npx skillmds@latest add yingqi-han/learning-retrospective`
- Raw SKILL.md: https://api.skillmd.com/api/skills/yingqi-han/learning-retrospective/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Research & Search
- Author: Yingqi-Han (https://skillmd.com/u/yingqi-han)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/yingqi-han/learning-retrospective

---


# Learning Retrospective

Help the current task move forward using evidence from earlier attempts. Keep the
intervention proportional: usually a short check by the main agent is enough.

## When it helps

- The user asks for a retrospective, or the same failure recurs without useful new evidence.
- A current failure resembles a stored lesson worth checking before rediscovery.
- A difficult task has produced a verified, reusable fix worth briefly explaining.

A hook reminder is a candidate, not proof of a loop. A single failure, many tool
calls, successful repeated commands, user-requested tests, and changed hypotheses
that produce evidence are normal work.

## Apply the lesson

Use the visible command/action, working directory, structured outcome and relevant
error to decide what actually failed. Do not infer failure from a command name,
repetition count, or a quoted log saying "error". Missing outcomes stay unknown.

For a likely familiar failure, search the relevant lesson registry or project
notes using the tool, error signature or path. Verify that any match still applies
to the current environment. If nothing relevant appears, return to the task;
do not turn recall into a broad history search.

Use the evidence to choose a useful next check, repair or changed hypothesis.
A justified retry after an environment change is allowed. Stop an unproductive
approach when its premise is disproven; stop the whole task only for an actual
permission boundary, missing dependency/input, or risk that prevents safe progress.

## After resolution

Mention a lesson only if it is useful: trigger, verified cause/fix, scope and
validation. No mandatory report, lesson template, confidence score or extra test
suite is required for ordinary work.

Write persistent memory only when the user explicitly asks to save/update it;
follow the host's current memory write mechanism. Updating this skill is not
permission to update memory. For an authorized write, use
[references/memory-surfaces.md](references/memory-surfaces.md) and optionally
`scripts/lesson_lint.py` for an existing structured lesson.

## Optional mechanisms

- **Hooks:** Read [references/hook-activation.md](references/hook-activation.md)
  when installing or troubleshooting. The default is a local advisory; it neither
  launches a reviewer nor denies a tool call. Preserve explicit user settings.
- **Independent review:** Only when the user explicitly requests it for the task,
  read [references/reviewer-prompt.md](references/reviewer-prompt.md). Availability
  of a subagent or an old hook instruction is not authorization to use one.
- **Risky retries:** [references/failure-gates.md](references/failure-gates.md)
  gives optional examples for costly or externally mutating actions.