Skip to content

cgallic/kai-cmo-harness / harness/skills/kai-retro

kai-retro

Verified

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 into enforced gate checks with golden corpus cases. Use when "retro", "what have we learned", "triage lessons", "promote lessons", "why does this keep failing", "harness retrospective", or monthly / after any heavy content sprint.

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).

Then read the outcome verdict ledger:

python -m scripts.quality_gates.eco_gate ledger --since <date of last retro>
python -m scripts.quality_gates.eco_gate unregistered

Diagnose every didn't work and partial row in its "Misses" table the same way, using the subject id. Read the verdict column, never CLOSED, because a CLOSED record can be a miss. List final-inconclusive and unmeasurable tickets as measurement gaps, not as losers. A change that unregistered lists shipped with no ticket: register it, or record why it cannot be measured.

Step 3 — Triage every lesson

Go through memory/lessons.md:

VerdictCriteriaAction
PromoteFired 3+ times, or checkable by a regex/thresholdGraduate it (Step 4), mark (promoted)
KeepTrue, useful, not yet recurringUpgrade candidate → active if verified
MergeNear-duplicate of another lessonCombine into the more general one
RetireNo 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], ...)
**Outcome verdicts:** worked [N] · partial [N] · didn't work [N] · inconclusive [N] · unmeasurable [N]
**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.

slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 194kUpdated Claude CodeCodexVerifiedInstall