SKILL.md
/retro — CP-7 close the V
Every non-tiny shipped or escalated run gets a retro. Retro is advisory but strongly expected: it is how the harness improves.
When to run
- After CP-6 ship (approved publish, merged PR, vault deliverable done)
- After escalate (capture why verifier failed)
- End of an
/ultragoalphase (one retro per phase) - Manual:
/retro <run-dir or spec path>
Phase 1 — Gather evidence bundle
Read:
- Spec or card description + acceptance criteria
04-projects/harness/runs/<id>/evidence/(ledger, checkpoints.tsv, CP-* files).claude/logs/loop-ledger.tsv(last rows for this run)04-projects/<goal>/STATUS.mdif an ultragoal phase
Phase 2 — Checkpoint audit
For each CP the lane required, mark Expected vs Actual vs Gap:
| CP | tiny |
normal |
full |
|---|---|---|---|
| CP-1 spec | skip | required | required |
| CP-3v component | skip | required | required |
| CP-4 integration | skip | skip | required if multi-task |
| CP-5 acceptance | if mutate | required | required |
| CP-6 ship | if external | if external | required |
Phase 3 — Evidence quality pass
Ask:
- Any AC-n shipped without a PASS evidence row? (traceability violation)
- Any PASS row where observation is tool-return not artifact? (confident-but-unchecked)
- Any criterion that wasn't falsifiable? (fix in next spec)
Phase 4 — Write retro
Copy references/retro-template.md to:
04-projects/harness/retro/YYYY-MM-DD-<slug>.md
Fill all sections. Action items get IDs (AI-01…).
Phase 5 — Feed forward
- Harvest: append action items + lessons to today's harvest staging
- Backlog: add
AI-nrows to04-projects/harness/BACKLOG.mdif harness work - STATUS.md: if an ultragoal phase, advance the phase state + log open
AC-n - Spec checkpoint log: update spec
## Checkpoint logCP-7 row - Record:
bash .claude/lib/checkpoint.sh record <run-dir> CP-7 PASS "retro: <path>"
Output to the user
TL;DR: outcome, top lesson, top action item, retro path.