| name | retro |
| description | Use this skill when the user wants to retrospect on a session, conversation, or recent stretch of work — triggers are "/retro", "let's do a retro", "review this session", "what could we have done better", "any patterns we should codify", or any moment when 3+ commits have been made or the user signals a natural pause point. Analyzes the conversation for repeated corrections, missed AskUserQuestion opportunities, recurring workflows, and effective patterns; proposes concrete updates to user preferences, skills, project memory, or scheduled tasks. Also fires reflexively when the user is "blocked on a manual action" (reading, deciding, restarting) so retrospective work happens during idle time.
|
Session Retrospective Skill
Analyze the current session and propose improvements to the Cowork configuration (preferences, skills, project memory, scheduled tasks).
Workflow
Step 0: Compile Session Context
Before analysis, gather:
- Commits in this session (if a git repo is in scope):
git log --oneline -<N> since session start
- Conversation events: corrections made, repeated instructions, workflows executed, AskUserQuestion invocations, tool permission denials, plan revisions
- Outputs produced: files in workspace folder created/modified
Format as a concise bullet list. Keep this in your context for Step 1.
Step 1: Analyze for Patterns
Look for these in the conversation:
- Repeated Corrections → preference candidate (the same mistake corrected 2+ times means it should be codified)
- Repeated Explanations → preference or project memory candidate (context Claude could have inferred)
- Repeated Workflows → skill candidate (multi-step processes executed manually 2+ times)
- Repeated Sub-agent Patterns → skill candidate
- Effective Patterns → preserve list (workflows that worked, agents that produced clean outputs)
- Existing Config Issues → modification candidate (rules ignored, skills mis-fitting)
Step 2: Deduplication Guard
If this session has already produced commits, edits to preferences, or skill creations, exclude proposals that map to those changes. Mark them "Already codified in [hash / edit]".
Step 3: Present Findings
Two sections:
Section A: Patterns to Reinforce — what worked. Preserve these in future config edits.
Section B: Improvement Proposals — numbered list. Each:
## Proposal N: [Short title]
**Type**: preference / skill / project-memory / scheduled-task
**Target**: [where it should go]
**Pattern observed**: [what happened in this session — concrete example]
**Proposed change**: [exact text to add or behavior to codify]
**Priority**: high / medium / low
Step 4: User Selection
Use AskUserQuestion (multiSelect) to let the user pick which proposals to implement.
Step 5: Implement
For each approved proposal:
- Preference: surface the exact text the user should paste into Cowork User preferences (Cowork preferences are user-managed; Claude cannot edit them programmatically)
- Skill: draft a
SKILL.md and supporting files in the workspace folder
- Project memory: write to the relevant project memory file (Cognee
cognify if connected, or a CLAUDE.md/AGENTS.md in the project repo)
- Scheduled task: use the
schedule skill
After implementation, present computer:// links for any new files.
When NOT to use
- Sessions with < 3 meaningful interactions (nothing to retrospect)
- The user has explicitly declined retro work in this session
- The session is mid-task and not at a natural pause