Atomic self-feedback refinement loop. Review one bounded work product, extract the highest-value improvement, run the smallest fix, and repeat until the explicit quality bar is satisfied or a real blocker stops progress.
Drive bounded self-critique and improvement loops toward a simpler, higher-quality result without drifting into over-engineering.
[stages: preflight>detect>analyze>review>reflect>plan>handoff>audit | scope: diff|repo|paths(glob,...) | policy: evidence,deterministic-output | lens: hickey-carmack | output: md(contract=v1)]
This skill uses hickey-carmack because it keeps the work aligned with: Data model first, explicit side effects, smallest high-value fix over decorative polish.
review-final-verify instead.MISSION_GOAL (string; required): One bounded mission that should improve through self-critique until done or blocked.TARGET_SCOPE (path|artifact|module|folder|repo|diff; required): Exact scope the loop is allowed to inspect or improve.DONE_CONDITION (list; required; shape: {CONDITION, PROOF_REQUIRED}): Explicit quality bar and proof required per condition.CURRENT_EVIDENCE (list; optional; shape: {TYPE, REF, WHY_RELEVANT}): Current evidence, review notes, tests, or docs from prior passes.COMPANION_SKILLS (list; optional; allowed: build-write-code|doc-write|tidy-simplify|test-write-guards|review-change|review-final-verify|plan-task-breakdown; shape: {SKILL}): Narrow companion skills allowed for the next improvement pass.MAX_PASSES (integer; optional): Safety ceiling for same-turn refinement passes. Default 3.CONSTRAINTS (list; optional; shape: {CONSTRAINT}): Non-goals, simplicity limits, or quality boundaries.MISSION_STATUS (continue|done|blocked; required): Whether the mission needs another pass, is complete, or is blocked.SELF_FEEDBACK (list; required; shape: {ISSUE, WHY_IT_MATTERS, SMALLEST_FIX}): Highest-value critique items from the current work.APPLIED_IMPROVEMENTS (list; required; shape: {ACTION, TARGET, EVIDENCE}): Improvements already applied during the current loop.DONE_CONDITION_STATUS (list; required; shape: {CONDITION, STATUS, EVIDENCE_OR_GAP}): Per-condition status against the done contract.NEXT_PASS (list; optional; required when MISSION_STATUS=continue; shape: {GOAL, RECOMMENDED_SKILL, PASS_CONDITION, WHY_THIS_FIRST}): Exactly one smallest next pass.BLOCKERS (list; optional; required when MISSION_STATUS=blocked; shape: {ISSUE, LOCATION, UNBLOCKING_CHECK}): Real blockers and cheapest check to unblock.LOOP_EXIT_REASON (string; required): Why the loop stopped.done only when every DONE_CONDITION row has proof and remaining improvements are non-essential.continue only when the next pass is known but cannot be safely executed within the current turn.primary_lens: hickey-carmackwhy: Self-verification should keep only the smallest high-value improvement, reject decorative over-engineering, and stop when the essence-first quality bar is proven.artifacts_in: implementation-delta.v1, documentation-report.v1, review-report.v1artifacts_out: improve-loop-report.v1Think and operate in English, but deliver the final response in Korean. 쉽고 간결한 한국어로 답하라. 전문 용어 금지. 핵심만 간단하게.
루프 상태: 완료 / 계속 / 중단
완료라면: 각 완료 기준과 근거 목록. 끝.
계속이라면: 다음 패스 하나 — 목표, 쓸 스킬, 통과 조건. 왜 이것인지.
중단이라면: 막힌 것과 가장 빠른 해결 방법.
자기 피드백: 이슈 → 왜 중요한지 → 가장 작은 수정.
MISSION_STATUS=done while any DONE_CONDITION row lacks proof.MISSION_STATUS=continue, NEXT_PASS must contain exactly one row.$compose + $control-improve-loop + $build-write-code + $review-final-verify + @src/auth + [Keep reviewing and tightening the refresh flow until the simplest correct shape is proven.]