AgentSkillsCN

hindsight

在每次 Claude Code 会话结束时，于用户层级（~/.claude/LEARNINGS.md）记录会话收获，包括哪些方法奏效、哪些尚待改进，以及自我评估的会话得分。当用户说“end session”、“wrap up”、“session review”、“record learnings”、“what did we learn”、“hindsight”、“/hindsight”，或当会话明显即将结束时，即可触发此技能。此外，当用户希望查看过往的学习成果或趋势时，也可使用此技能。

SKILL.md
--- frontmatter
name: hindsight
description: Track session learnings and maintain a persistent LEARNINGS.md at the user level (~/.claude/LEARNINGS.md). Use this skill at the END of every Claude Code session to record what worked, what didn't, and a self-assessed session score. Trigger when the user says "end session", "wrap up", "session review", "record learnings", "what did we learn", "hindsight", "/hindsight", or when a session is clearly ending. Also use when the user asks to view past learnings or trends.

Hindsight

Record session outcomes, patterns, and scores to build a persistent knowledge base across all Claude Code sessions.

File Location

All learnings are stored in ~/.claude/LEARNINGS.md (user-level, shared across all projects).

When to Record

Record a session entry when:

  • •The user explicitly asks to end/wrap up a session
  • •The user asks for a session review or to record learnings
  • •The user invokes /session-learnings

Session Entry Format

Append each entry to ~/.claude/LEARNINGS.md, creating the file if it doesn't exist.

markdown
---

## Session: YYYY-MM-DD HH:MM — [Brief Title]

**Project:** [repo/project name or "General"]
**Branch:** [git branch if applicable]
**Duration estimate:** [short/medium/long based on conversation length]

### What Worked
- [Specific thing that went well — be concrete]
- [Tool/approach/pattern that was effective]

### What Didn't Work
- [What failed, was slow, or caused friction]
- [Approaches that were abandoned and why]

### Key Learnings
- [Reusable insight for future sessions]
- [Pattern discovered, gotcha avoided, or technique learned]

### Session Score: X/10

| Dimension         | Score | Notes                        |
|-------------------|-------|------------------------------|
| Task completion   | X/10  | [Did we finish what was asked?] |
| Code quality      | X/10  | [Clean, tested, follows patterns?] |
| Efficiency        | X/10  | [Minimal tool calls, no wasted effort?] |
| Communication     | X/10  | [Clear, concise, helpful?] |
| **Overall**       | **X/10** | |

### Follow-ups
- [ ] [Any unfinished work or next steps]

Score Dimensions

  • •Task completion (40%): Did the session accomplish what was asked? 10 = fully done, 1 = nothing completed.
  • •Code quality (25%): Is the code clean, tested, following repo conventions? 10 = production-ready, 1 = broken.
  • •Efficiency (20%): Minimal unnecessary steps? 10 = direct path, 1 = excessive searching/retries.
  • •Communication (15%): Clear, concise, helpful? 10 = perfect clarity, 1 = confusing/verbose.
  • •Overall: Weighted average, rounded to nearest 0.5.

Process

  1. •Read ~/.claude/LEARNINGS.md if it exists (check format and see past entries)
  2. •Reflect on the current session: what was accomplished, what tools/approaches worked, what caused friction
  3. •Generate the entry using the template
  4. •If the file doesn't exist, create it with this header before the first entry:
markdown
# Session Learnings

Persistent record of Claude Code session outcomes, patterns, and scores.

| Date | Project | Score | Title |
|------|---------|-------|-------|
  1. •Add a row to the summary table at the top for the new session
  2. •Append the full entry after the table (newest entries first, right after the table)
  3. •Show the user a summary of the entry

Querying Past Learnings

When the user asks "show my learnings", "what patterns have I seen", or "learnings summary":

  1. •Read ~/.claude/LEARNINGS.md
  2. •Summarize: average scores, recurring themes in "What Worked" / "What Didn't", top learnings
  3. •Highlight score trends (improving/declining over time)