| name | retrospective |
| description | After completing a significant task or experiment, extract lessons learned and update the project knowledge base. Captures what worked, what failed, and what to remember for next time. |
| argument-hint | ["task-description"] |
| disable-model-invocation | true |
| allowed-tools | Read, Edit, Write, Bash, Grep, Glob |
Retrospective: Capture Lessons Learned
After completing "$ARGUMENTS", extract insights and update the project knowledge base.
Steps
-
Gather context: Read recent changes, results, and any errors encountered during the task. Look at:
- Recent git commits:
git log --oneline -10
- Modified files:
git diff --name-only HEAD~5
- Any error logs in
logs/
- Result files if experiment-related
-
Extract lessons in these categories:
What worked — approaches, tools, or patterns that were effective
What failed — approaches that were tried and didn't work, with WHY
Surprises — unexpected findings or behaviors
For next time — specific recommendations for similar future tasks
-
Update knowledge base:
If there are failed approaches: Append to docs/failures.md following the existing format:
### Title
**When**: Date
**What happened**: Description
**Root cause**: Why it failed
**Fix/workaround**: What to do instead
If there's a decision: Append to docs/decisions.jsonl using the established schema.
If there are experiment findings: Consider whether docs/findings.md needs updating.
-
Summarize the retrospective in 3-5 bullet points for the user.
Rules
- Be specific. "It was hard" is not useful. "XGBoost training fails with >100k samples on MPS due to memory limit" is useful.
- Include exact error messages or config values when relevant.
- Failed attempts with details are MORE valuable than successes — they prevent repeated mistakes.
- Don't update files unless there's genuinely new information to capture.
