kaizen

Continuous-improvement ritual — capture one small improvement, attribute the object that actually improved, and emit a recursion-evidence receipt. Use after shipping a fix, workflow upgrade, or harnes

SKILL.md

GitHub

kaizen — Continuous Improvement (RE1)

Small improvements compound only when they are captured and attributed. kaizen turns "we fixed X" into a durable receipt that says what object improved, whether the improvement is structural or effective recursion, who led it, and whether it changed the next improvement cycle.

One receipt per improvement. Small by design — a kaizen receipt for a renamed variable is correct; a missing receipt for a shipped improvement is the bug.

When to Use

  • After shipping any improvement — code, skill, workflow, harness, doc, process
  • During /aar or /retrospective when an improvement is named
  • When an RSI or compounding claim needs evidence instead of assertion
  • Before /end-session if the session shipped improvements

When to Skip

  • No improvement shipped — activity is not gain. Do not receipt effort.
  • The claim is "the model got smarter" — recursive_target: model claims need model-level evidence; capture the harness/workflow/skill that actually changed.

Execution

1. Name the improved object — recursive_target

Attribute the improvement to what actually changed:

model | prompt | memory | skill | workflow | harness |
evaluator | curriculum | exploration_policy | infrastructure | organization

Do not infer model when the real target is the harness, memory, exploration policy, or organization. Multiple targets are allowed when the change is genuinely compound.

2. Classify the recursion

  • structural — can the improvement loop now reach the thing that produces improvements? (the system can change its change-maker)
  • effective — is there evidence improvement n increased the ability to produce or verify improvement n+1?

Ordinary iteration does not qualify as effective recursion. unknown is a valid answer — do not manufacture certainty.

3. Record the work attribution

  • task_lead: human | agent | mixed | unknown
  • human_interventions: count of human interventions required
  • successor_cycle_effect: did this change the next valid improvement cycle? improved | unchanged | degraded | unknown

4. Fill evaluator provenance

Who or what admits this claim — identity, funder, selection/removal authority, access/evaluation/publication scope, conflicts, and whether the candidate system can modify its own evaluator or eval data. External ≠ independent.

5. Emit the receipt

Append one JSONL line to:

~/.agents/_state/rsi/recursion-evidence.jsonl

conforming to docs/schemas/evals/recursion-evidence-receipt.v1.md (schema_version: recursion-evidence-receipt.v1.0.0). Corrections are new lines with supersedes: <prior receipt_id> — never edit a written line.

6. Verify the write

tail -1 ~/.agents/_state/rsi/recursion-evidence.jsonl

A missing line means the write failed silently — re-emit before reporting.

Examples

# Improvement to a workflow, agent-led, effect on next cycle unknown
kaizen "index freshness check added to end-session" \
  --target workflow --lead agent --structural present --effective unknown

# Skill improvement with clear successor effect
kaizen "skill-creator now emits provider-neutral frontmatter" \
  --target skill --lead mixed --structural present --effective present \
  --successor improved --interventions 1

Constraints

  • unknown / underdetermined is an allowed terminal state — never upgrade a guess into a claim.
  • No automatic authority expansion, TierShift, deployment, merge, or release follows from any receipt.
  • Fleet activity (agent count, task volume, AI work-share) is NOT recursive gain — receipt the object that improved, not the motion.
  • Corrections supersede; they never rewrite history.

Skill Chains

Mandatory

  • None — kaizen is a leaf capture ritual.

Advisory

  • Before kaizen: /aar, /retrospective (improvement source)
  • After kaizen: /rsi-dashboard (verify encoding lands in metrics), /ledger (if the improvement reveals a durable pattern)

Authority

  • T1 (TRUSTED): May run without restriction
  • T2 (Active/High): May run without restriction
  • T3 (Medium): Operator approval before writing receipts
  • T4 (Probationary): BLOCKED — cannot write receipts
  • Operator: Override any restriction