AgentSkillsCN

debrief

会话结束后的自我审视。扫描潜在的修正点、失败之处以及缺失的上下文信息,随后提出切实可行的规则更新建议。

SKILL.md
--- frontmatter
name: debrief
description: "Session-end self-review. Scans for corrections, failures, and missing context, then proposes concrete rule updates."
last_updated: 2026-02-07
tools_required: [Read, Edit, Glob, Grep]
agent_type: main_agent

Session Debrief

Review this session for mistakes, corrections, and learnings. Generate concrete rule updates.

[GOAL]

Identify what went wrong, what was missing, and what went well in this session. Convert findings into specific, actionable rule updates across CLAUDE.md, Chief-of-Staff.md, memory files, and skills.

[PROCESS]

Step 0: Self-Review (scan the conversation)

Scan the full conversation for each category below. List every instance found, not just the first.

0a. Corrections - Where the user said "no", "that's wrong", "I meant...", "not that", or redirected you.

  • What you did vs what they wanted
  • Rule: when X, do Y instead of Z

0b. Failed approaches - Things tried 2+ times before working, tool calls that errored, dead ends.

  • What failed and why
  • Rule: avoid X because Y; instead do Z

0c. Repeated instructions - Things the user had to say more than once, or context they had to re-provide.

  • What should have been known from existing docs
  • Rule: always check X before doing Y

0d. Frustration signals - Pushback, exasperation, "I already said...", course corrections, user taking over.

  • What triggered it
  • Rule: don't X without asking first

0e. What worked well - Approaches the user praised, things that went smoothly, good patterns to reinforce.

  • What made it work
  • Rule: continue doing X when Y

If no instances found for a category, say "None detected" and move on.

Step 1: Reflect on missing context

What context was missing that would have helped from the start?

  • Commands or patterns discovered mid-session
  • Environment/configuration quirks encountered
  • Code style patterns that weren't documented
  • Testing approaches that worked
  • Gotchas or warnings encountered

Step 2: Find target files

Identify where each finding belongs:

TargetWhen to use
CLAUDE.md (project)Team-shared patterns, project-specific rules
~/.claude/CLAUDE.md (global)Cross-project behaviors, persona rules
Cockpit/Chief-of-Staff.mdFacts, preferences, project context used daily
memory/topics/*.mdDetailed context for specific domains
skills/*.mdSkill-specific improvements
rules/*.mdCoding guideline updates

Read each target file before proposing edits.

Step 3: Draft updates

For each finding, draft a concrete update:

code
### Update: <filepath>

**Finding:** <what happened>
**Rule:** <when X, do Y instead of Z>

\`\`\`diff
+ <the addition - one line per concept, keep it brief>
\`\`\`

Keep it concise - one line per concept. These files are part of the prompt, so brevity matters.

Avoid:

  • Vague self-improvement ("be more careful")
  • Obvious information
  • One-off fixes unlikely to recur
  • Duplicating rules already present

Step 4: Present summary

Show the user:

code
## Session Debrief

### Corrections (X found)
- [list each with one-line rule]

### Failed Approaches (X found)
- [list each]

### Missing Context (X found)
- [list each]

### What Worked Well (X found)
- [list each]

### Proposed Updates (X total)
[Step 3 output]

Step 5: Apply with approval

Ask {USER_NAME} which changes to apply. Only edit files they approve.

After applying, confirm what was changed and where.

[IMPORTANT]

  • Be specific - "when working with Supabase MCP, always start tunnels first" not "remember to prepare"
  • Be honest - If the session went perfectly, say so. Don't manufacture findings.
  • No self-flagellation - State facts, propose fixes. Skip the apologies.
  • Cite the moment - Reference what the user said or what failed, so they can verify
  • Deduplicate - Check existing rules before adding new ones
  • One line per rule - These go into prompt context. Every word costs tokens.