kai-retro

Run a learning retrospective on the Kai harness. Mines gate-failure logs and 30-day performance results into lessons, triages candidate lessons (promote/keep/retire), and graduates repeated lessons in

SKILL.md

GitHub

Kai root note: knowledge/, harness/, and scripts/ paths in this skill live in the Kai install, not the user's project. Resolve them against the first ancestor directory of this SKILL.md that contains a knowledge/ folder (the Kai plugin root, ~/.claude/kai, or the kai-cmo-harness repo). MARKETING.md, memory/, and any output files live in the current project. If a referenced scripts/ command is not available in this install, say so, skip it, and continue with the file-based guidance — never fabricate its output.

Run the Kai learning retrospective. This is how the harness gets smarter: raw failure logs become lessons, repeated lessons become enforced checks. Read memory/MEMORY.md first for the graduation ladder.

When to run

  • Monthly, or after any sprint that produced 5+ gated pieces
  • Whenever the same gate failure shows up twice in one session
  • When a 30-day performance check grades new underperformers

Step 1 — Mine the gate logs

python scripts/self_improvement/lesson_capture.py mine

This groups recurring failure signatures from data/learning/gate_runs.jsonl. Append candidates with --write. If the log is empty, note it and move on — the gates only log when they run.

Step 2 — Diagnose underperformers

python scripts/self_improvement/lesson_capture.py losers

This lists pieces whose 30-day grade is underperformer (the grade vocabulary is winner/average/underperformer; older docs said "loser") with no diagnosis yet. For each one, read the piece and its content_log.json entry, write a one-line diagnosis (hook type, persona mismatch, seasonality, thin proof — name the cause, not the symptom), and add it to memory/what-doesnt-work.md under "Measured losers" with the piece id. Check seasonality and competitor moves before blaming the content (see memory/edge-cases.md EC-15).

Step 3 — Triage every lesson

Go through memory/lessons.md:

Verdict Criteria Action
Promote Fired 3+ times, or checkable by a regex/threshold Graduate it (Step 4), mark (promoted)
Keep True, useful, not yet recurring Upgrade candidate → active if verified
Merge Near-duplicate of another lesson Combine into the more general one
Retire No longer true (platform changed, gate fixed) Mark (retired) with the reason — never delete

Step 4 — Graduate promoted lessons

Pick the strongest enforcement target, in this order:

  1. Lint rule / contract check — new entry in a banned-word tier, a new overclaim regex in scripts/quality_gates/seo_lint.py, or a deterministic_checks line in the format's skill contract.
  2. Checklist line — the relevant knowledge/checklists/*.md.
  3. CLAUDE.md / framework rule — only for judgment calls code can't check.

Non-negotiable: any change to a gate script requires a matching case in evals/golden/manifest.json (one sample proving the new check fires, and confirm the existing pass samples still pass), then:

python scripts/quality_gates/golden_check.py

A gate change without a golden case is not a promotion — it's a regression waiting to happen.

Step 5 — Refresh the memory index

  • Update the "Current standing lessons" section of memory/MEMORY.md (keep the file under 200 lines).
  • Cross-check memory/edge-cases.md: mark any entry whose Enforcement: none you just fixed; add new edge cases discovered this cycle.

Step 6 — Report

Output a retro summary:

## Kai Retro — [date]

**Mined:** [N] recurring failure signatures ([gate]: [signature] ×[count], ...)
**Underperformers diagnosed:** [N] ([id]: [one-line diagnosis], ...)
**Promoted:** [lesson] → [enforcement target] (+ golden case [id])
**Retired:** [lesson] — [reason]
**Edge cases:** [new/closed entries]
**Open risks:** [lessons at 2 occurrences — one more and they must promote]

Commit the memory and gate changes together so the diff shows the lesson and its enforcement side by side. If a promotion changes publishing behavior (new hard block), flag it for human approval rather than applying silently.