Install any skill in seconds. Free to start, no credit card required.
Get Started Free →멀티 에이전트 디자인 워크플로우 supervisor. writer, locale adaptation, humanize, UI slop audit, designer review, final QA, image materialization을 routing한다. 2-round revision cap을 유지하며 다국어 문서·UI 개선·출간 준비처럼 여러 역할이 필요한 요청에 사용한다.
.claude/skills/kwakseongjae-omd-orchestrator/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | 81% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 76% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 145% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 36% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 146% | 0% |
<!-- omd:installed-skill — managed by omd install-skills. Do not edit; rerun the command to refresh. -->
omd v0.2 agent layer의 supervisor. 한 글/한 컴포넌트가 여러 specialist를 거쳐야 할 때 routing을 책임진다.
채택 패턴: Anthropic orchestrator-workers (Building effective agents, 2024-12) + LangGraph supervisor의 revision-cap. 자세한 비교는 data/research/2026-05-18-agent-landscape.md §1.
> 런타임 dependency 없음. Claude Code의 subagent 호출 메커니즘이 그대로 orchestrator-workers 토폴로지를 구현한다.
| 역할 | subagent | 용도 | |---|---|---| | Writer | omd-kr-writer | 한국어 본문 작성. preset_id 인자로 voice 결정. | | Localizer | omd-locale-adapter | KR → EN/JA/ZH-CN/ZH-TW adaptation (번역 아님) | | Copy finalizer | omd-humanizer | locale별 번역투·기계적 구조를 국소 수정하고 보호 구간 대조 | | Slop auditor | omd-slop-auditor | 실제 route의 context-free UI/copy cluster를 품질·취향과 분리해 감사 | | Visual reviewer | omd-designer-review | DESIGN.md 대비 typo/색/spacing/state 검수 | | UX engineer | omd-ux-engineer | 실제 route의 interaction·responsive·focus·perceived performance 감사 | | UX writer | omd-ux-writer | 섹션별 copy·정보 순서·CTA 계약 감사 | | A11y auditor | omd-a11y-auditor | axe/lighthouse/키보드 기반 결정론 gate | | Critic | omd-final-qa | Read-only rubric verdict. 2-round cap 강제. | | Image materializer | omd-codex-image | <!-- omd:gen-image --> 블록을 채널별로 실체화 (Codex native gen / asset-curator fallback / OpenCode user-queue) |
사용자 요청. 예:
사용자 요청 도착
├─ "글 작성" 키워드 → Stage 1: omd-kr-writer
├─ "AI 티/번역투/문장이 기계적" → omd-humanizer
├─ 다국어 요청 ("EN", "영문", "JA", "간체", "대만어") → Stage 3 + locale별 humanize 추가
├─ "AI slop/템플릿 같음" + audit만 → omd-slop-auditor
├─ 기존 UI "고쳐/개선/구현" → UI delivery lane → 수정은 caller의 omd:apply
├─ artifact 첨부 + "리뷰" → Stage 2부터 진입
└─ "출간 ready" → Stage 4 final-qa로 직행기존 UI의 변경 요청에서는 orchestrator가 제품 파일을 직접 편집하지 않는다. 다음 work packet을 먼저 고정하고 필요한 specialist를 최대 3개까지만 선택해 read-only로 실행한다.
yamlintent: audit | implement task: <user outcome> consumer_route: <real user route> acceptance: [] protected_behaviors: [] evidence: [] unknowns: [] implementation_owner: caller-main-agent verification: routes: [] viewports: [] states: [] commands: []
각 specialist에게는 packet과 필요한 artifact만 전달한다. 응답은 finding / evidence / smallest_useful_change / acceptance_check / unresolved로 정규화한다. orchestrator는 중복·충돌을 합쳐 우선순위가 있는 하나의 implementation handoff를 caller에게 반환한다.
intent: audit이면 report로 종료 가능.intent: implement이면 status: advice-ready, implementation_owner: caller-main-agent로 반환한다. 완료·수정됨·검증됨이라고 표현하지 않는다.omd:apply로 실제 편집한 뒤 packet의 같은 consumer route·viewport·state를 재검증한다.Stage 1 WRITE omd-kr-writer (preset=toss-tech-design)
Stage 1h HUMANIZE omd-humanizer (KO 보호 구간 + 자연스러움 검증)
Stage 2 REVIEW omd-designer-review (artifact + brand DESIGN.md)
Stage 2r REVISION omd-kr-writer (review feedback 반영) ← max 2 round
Stage 3 LOCALIZE omd-locale-adapter (KR → EN/JA/ZH-CN/ZH-TW)
Stage 3h HUMANIZE omd-humanizer (각 locale 독립 검증)
Stage 4 CRITIC omd-final-qa (rubric, read-only)
Stage 4r REVISION omd-kr-writer (final-qa feedback) ← max 2 round (Stage 2와 별도 카운트)
Stage 4i IMAGES omd-codex-image (`<!-- omd:gen-image -->` 블록 처리, 채널별 분기)
Stage 5 HANDOFF 사용자에게 최종 artifact + revision log + image manifest각 critic gate (designer-review, final-qa)는 최대 2 round.
revision_state = {
"designer_review": { "round": 0, "max": 2 },
"final_qa": { "round": 0, "max": 2 }
}Round 2 후에도 BLOCK이면:
각 stage 전후로 <work_dir>/.orchestrator.log에 append:
[2026-05-18T10:23:11] STAGE=write agent=omd-kr-writer status=ok artifact=content/posts/karrot/index.ko.md
[2026-05-18T10:25:02] STAGE=review agent=omd-designer-review status=WARN issues=2
[2026-05-18T10:25:03] STAGE=write agent=omd-kr-writer round=2 reason="색 budget 위반 fix"
...Orchestrator 자체는 stateless. 모든 state는 파일 시스템:
content/posts/<slug>/index.<locale>.mdcontent/posts/<slug>/.reviews/round-<N>.mdcontent/posts/<slug>/.reviews/final-qa.mdClaude Code subagent 호출 시 다음 envelope:
yamlagent: omd-kr-writer inputs: work_packet: intent: publish task: "당근 디자인 분석 글 작성" consumer_route: null acceptance: ["KO article passes final QA"] protected_behaviors: ["facts and source URLs remain unchanged"] evidence: ["references/karrot/DESIGN.md"] unknowns: [] implementation_owner: omd-kr-writer verification: { routes: [], viewports: [], states: [], commands: [] } task: "당근 디자인 분석 글 작성" preset_id: toss-tech-design brand_design_md: references/karrot/DESIGN.md target_length: 6000 output_path: content/posts/karrot/index.ko.md revision_round: 0 prior_review: null # 또는 review report 경로
응답을 받으면 .orchestrator.log에 기록 후 다음 stage 결정.
done 금지..orchestrator.log에 ABORT 기록 후 종료| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 29,119 | 12,198 | -58% | 1 | 1 | 0% | 5,756 | 3,355 | -42% | 0 | 0 | — |
case-02 | fail→fail | 33,884 | 6,126 | -82% | 1 | 1 | 0% | 7,270 | 2,793 | -62% | 0 | 0 | — |
case-03 | fail→fail | 11,175 | 7,779 | -30% | 1 | 1 | 0% | 1,994 | 2,899 | +45% | 0 | 0 | — |
case-04 | pass→pass | 10,591 | 6,661 | -37% | 1 | 1 | 0% | 2,458 | 3,822 | +55% | 0 | 0 | — |
case-05 | pass→pass | 12,512 | 9,884 | -21% | 1 | 1 | 0% | 2,415 | 4,315 | +79% | 0 | 0 | — |
case-06 | pass→pass | 12,815 | 8,566 | -33% | 1 | 1 | 0% | 2,295 | 3,844 | +67% | 0 | 0 | — |
case-07 | fail→pass | 11,374 | 5,633 | -50% | 1 | 1 | 0% | 1,847 | 3,340 | +81% | 0 | 0 | — |
case-08 | pass→pass | 15,442 | 6,772 | -56% | 1 | 1 | 0% | 2,603 | 3,601 | +38% | 0 | 0 | — |
case-09 | fail→pass | 11,986 | 6,202 | -48% | 1 | 1 | 0% | 2,053 | 3,604 | +76% | 0 | 0 | — |
case-10 | fail→pass | 7,984 | 6,005 | -25% | 1 | 1 | 0% | 1,410 | 3,452 | +145% | 0 | 0 | — |
case-11 | fail→pass | 18,974 | 10,614 | -44% | 1 | 1 | 0% | 3,351 | 4,574 | +36% | 0 | 0 | — |
case-12 | pass→pass | 11,112 | 4,824 | -57% | 1 | 1 | 0% | 1,798 | 3,263 | +81% | 0 | 0 | — |
case-13 | fail→pass | 10,947 | 9,623 | -12% | 1 | 1 | 0% | 1,709 | 4,206 | +146% | 0 | 0 | — |
case-14 | pass→pass | 12,754 | 6,447 | -49% | 1 | 1 | 0% | 2,134 | 3,448 | +62% | 0 | 0 | — |
case-15 | fail→pass | 12,642 | 4,218 | -67% | 1 | 1 | 0% | 2,112 | 3,057 | +45% | 0 | 0 | — |
case-16 | fail→pass | 12,937 | 4,653 | -64% | 1 | 1 | 0% | 2,398 | 3,302 | +38% | 0 | 0 | — |
case-17 | pass→pass | 14,190 | 10,843 | -24% | 1 | 1 | 0% | 2,812 | 4,603 | +64% | 0 | 0 | — |
case-18 | pass→pass | 11,979 | 3,691 | -69% | 1 | 1 | 0% | 2,137 | 3,062 | +43% | 0 | 0 | — |
case-19 | fail→pass | 8,144 | 4,002 | -51% | 1 | 1 | 0% | 1,480 | 3,091 | +109% | 0 | 0 | — |
case-20 | fail→pass | 14,436 | 9,815 | -32% | 1 | 1 | 0% | 2,590 | 4,251 | +64% | 0 | 0 | — |
case-21 | fail→pass | 9,513 | 4,082 | -57% | 1 | 1 | 0% | 1,493 | 3,169 | +112% | 0 | 0 | — |
case-22 | fail→pass | 8,083 | 2,793 | -65% | 1 | 1 | 0% | 1,415 | 2,967 | +110% | 0 | 0 | — |
case-23 | fail→pass | 12,949 | 5,180 | -60% | 1 | 1 | 0% | 2,568 | 3,286 | +28% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted, and 20 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +52 percentage points is the difference between those two pass rates over the 20 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 8/21/2026 | +64% |
Other measured skills in the registry, with their headline benchmark lift.