| name | retrospective |
| description | Session retrospective: review what was done, identify reusable patterns worth turning into skills. Use at the end of a session to extract learnings and skill candidates. Triggers: "振り返り", "retrospective", "retro", "セッション振り返り", "スキル化できるものある?", "今日の振り返り".
|
Retrospective Skill
Analyze the current session to extract reusable patterns, recurring workflows, and skill candidates.
Workflow
1. Session Summary
Review the conversation history and summarize:
- What was done: features implemented, bugs fixed, tasks completed
- What tools/commands were used repeatedly
- What friction occurred: errors, retries, misunderstandings, manual steps
2. Pattern Detection
Look for these signals that indicate a workflow should become a skill:
Repetition signals:
- Same sequence of commands executed 2+ times in this session
- A workflow that was also done in previous sessions (check MEMORY.md)
- Manual steps that could be automated (e.g., "start server, then check proxy, then curl endpoint")
Friction signals:
- Steps where Claude had to retry or the user had to correct
- Configuration that was forgotten and had to be redone
- Verification steps that were skipped and caused issues later
Complexity signals:
- Multi-step workflows with a specific order (build → test → deploy)
- Domain-specific knowledge needed (project conventions, API patterns)
- Checklists that need to be followed consistently
3. Evaluate Candidates
For each candidate pattern, assess:
| Criteria | Question |
|---|
| Frequency | Will this be used again? (weekly+) |
| Complexity | Is it more than 2 trivial steps? |
| Error-prone | Are steps likely to be forgotten or done wrong? |
| Generic | Does it apply across projects or just this one? |
Score each candidate:
- 3+ criteria met → Strong skill candidate
- 2 criteria met → Consider as skill
- 1 or fewer → Not worth a skill (just document in MEMORY.md)
4. Report
Present findings in this format:
## Session Retrospective
### What was done
- [summary of work]
### Skill candidates found
#### 1. [Pattern name] ⭐ Strong candidate
- **What:** [description of the workflow]
- **Why:** [which signals triggered: repetition/friction/complexity]
- **Frequency:** [how often this would be used]
- **Example from this session:** [concrete example]
#### 2. [Pattern name] 🔶 Consider
- **What:** [description]
- **Why:** [signals]
- ...
### Project-level learnings (for project MEMORY.md)
- [project-specific conventions discovered]
- [project-specific gotchas encountered]
- [patterns tied to this codebase's architecture or domain]
### General learnings (for ~/.claude/MEMORY.md)
- [language/tool gotchas applicable across projects]
- [reusable techniques or best practices discovered]
- [insights about tools, libraries, or APIs]
5. Create Skills (with permission)
For strong candidates, ask the user:
Found N skill candidates. Create them?
1. ⭐ [name] — [one-line description]
2. ⭐ [name] — [one-line description]
3. 🔶 [name] — [one-line description]
If approved, create each skill at ~/.claude/skills/<name>/SKILL.md with:
- YAML frontmatter (
name, description with triggers)
- Clear workflow steps
- Important rules section
6. Update Memory
Classify each learning and write to the appropriate MEMORY.md:
Project-level → <project-root>/MEMORY.md
- Conventions specific to this codebase (naming, ordering, architecture)
- Domain-specific gotchas (e.g., pattern priority rules for this PII sanitizer)
- Project-specific configuration or workflow notes
General-level → ~/.claude/MEMORY.md
- Language or tool gotchas applicable across any project (e.g., JS regex alternation)
- Reusable techniques or best practices
- Insights about libraries, APIs, or tools
Ask the user before writing if the classification is ambiguous.
Skill Creation Checklist
When creating a new skill from a detected pattern:
Important Rules
- DO review the full session, not just the last few messages
- DO check existing skills (
~/.claude/skills/) to avoid duplicates
- DO check MEMORY.md for patterns that appeared in previous sessions too
- DO NOT create skills for one-off tasks — only recurring patterns
- DO NOT create skills without user approval
- Present candidates clearly so the user can decide which to create