moai-coworker | Skill creation based on the revfactory/harness methodology
Overview
This skill adapts the 6-Phase skill creation workflow from the harness open-source project to MoAI's cowork context. It reduces quality variance in manual skill authoring and produces consistently high-quality skills through a systematic creation process.
Name History: The skill provided as
skill-forgeup to v1.3.x was renamed tometa-skill-builderin v1.5.x. Aliases are not maintained, and all references must be updated tometa-skill-builder.
Trigger Keywords
skill creation skill authoring meta-skill-builder create new skill add skill new skill harness workflow
Workflow
Phase 1: Requirements → 사용자 의도 분석, 트리거 키워드 정의
Phase 1.5: Research → 공식 자료·베스트 프랙티스 외부 조사 (WebSearch + Context7) ★ v1.6.0
Phase 2: Architecture → 에이전트 패턴 선택 (6종 중 매핑)
Phase 3: Skill Draft → meta-skill-template 기반 SKILL.md 초안 + ## 출처 섹션
Phase 4: Test Gen → 테스트 프롬프트 2-3개 + 기대 출력 정의
Phase 5: Validation → 4차원 루브릭 스코어링 (meta-skill-tester 본문 §스코어링 루브릭)
Phase 6: Review → 품질 게이트 통과 확인, 파일 배치
Execution Rules
Phase 1: Requirements (requirements analysis)
Extract skill requirements from the user's natural-language request.
Required items to collect:
| Item | Question | Example |
|---|---|---|
| Purpose | "What problem will this skill solve?" | "I want to automatically generate sales proposals" |
| Target plugin | "Which plugin should it be placed in?" | moai-coworker |
| Input | "What will the user input?" | Company name, target industry, product information |
| Output | "What deliverable is expected?" | Proposal DOCX file |
| Complexity | "How complex is the skill?" | Standard (50-150줄) |
Deliverable: Requirements document (in-memory, confirmed via AskUserQuestion)
Phase 1.5: Research (external material research) ★ New in v1.6.0
Research official materials, standard formats, and best practices in the skill domain to ensure the accuracy and authority of SKILL.md.
Research triggers (Phase 1.5 is mandatory if any of the following applies):
- Related to Korean laws/regulations or standard formats (tax, labor, contracts, hiring, medical, etc.)
- Use of external libraries, SDKs, APIs, or CLIs
- Domains requiring citation of global best practices (B2B SaaS, HR, BI, DevOps, etc.)
- Citation of quantitative figures, statistical data, or market size
Research procedure:
-
WebSearch 2-4 times (parallel recommended)
- 1-2 Korean queries:
"<도메인> <핵심 키워드> <연도>"format (e.g., "Korean Employment Procedure Act NCS 2026") - 1-2 English queries (when citing global best practices):
"<topic> best practice <year>"format - Check the recency of search results (prioritize materials within 2 years)
- 1-2 Korean queries:
-
Context7 MCP (for library, SDK, and CLI-related skills)
resolve-library-id→get-library-docsorder- Check latest API/CLI changes and deprecations
- Supplement training data cutoff
-
Official source whitelist by domain (preferred citations):
Domain Primary recommended sources Korean tax and finance National Tax Service nts.go.kr, Hometax hometax.go.kr, DART dart.fss.or.kr Korean labor and employment Ministry of Employment and Labor moel.go.kr, NCS ncs.go.kr, Korea Labor Institute kli.re.kr Korean medical and health MFDS mfds.go.kr, Ministry of Health and Welfare mohw.go.kr, NHIS nhis.go.kr Korean government support K-Startup k-startup.go.kr, BIZINFO bizinfo.go.kr Korean statistics and market KOSIS kosis.kr, Bank of Korea ECOS ecos.bok.or.kr Korean law National Law Information Center law.go.kr, Supreme Court Comprehensive Legal Information glaw.scourt.go.kr Global accounting standards K-IFRS, IFRS Foundation ifrs.org Global hiring regulations EEOC eeoc.gov, NYC Local Law 144, ISO 42001:2023 Global SaaS BP Official product documentation (HubSpot, Salesforce, Greenhouse, etc.) + industry-leading blogs -
Source preservation:
- Preserve URLs, publication dates, and key quoted sentences from research results in memory
- In Phase 3, SKILL.md
## 출처cite in the section in markdown hyperlink format - Quantitative figures must have inline source attribution or
[추정]tag enforcement
Deliverables: research notes (in memory, passed as Phase 3 input)
Prohibited:
- Citing quantitative figures without sources (must specify source or
[추정]tag) - Missing Korean official sources in Korean domain skills
- Not using Context7 MCP for library and SDK-related skills (when available)
- Citing outdated materials (if older than 3 years, re-search for the latest materials)
Skip conditions (Phase 1.5 may be omitted):
- Simple format conversion and text processing skills (no domain knowledge required)
- User explicitly
--skip-researchflag specified - Only minor improvements to existing cowork skills or trigger keyword adjustments
Phase 2: Architecture (architecture selection)
Select the optimal pattern among 6 agent patterns.
| Pattern | Application scenario | Skill example |
|---|---|---|
| Pipeline | Sequential processing, where each step's output becomes the next input | doc-docx, content-blog |
| Fan-out/Fan-in | Integrate results after parallel analysis | collab-ux-design, market-analysis |
| Expert Pool | Combine expertise from multiple domains | consult-brief, consult-startup |
| Producer-Reviewer | Generation + quality review cycle | collab-sales-playbook, content-copywriting |
| Supervisor | Orchestration of multiple specialized skills | moai-coworker:ai-diagnostic, project |
| Hierarchical Delegation | Hierarchical decomposition of complex tasks | consult-startup |
Selection criteria: Judge based on the number of workflow steps in the skill, whether there is parallelism, and the need for quality verification.
Output: Selected pattern + rationale (1-2 sentences)
Phase 3: Skill Draft (skill draft generation)
meta-skill-template Generate a SKILL.md draft based on the template.
Required application:
- Use the required section structure of meta-skill-template
- Frontmatter는
name,descriptionInclude only (metadata prohibited) - Check that trigger keywords do not duplicate the existing 73 skills
- Include at least 2 usage examples
- Specify before/after/alternative relationships in the related skills section
## 출처Section mandatory: Specify URLs and citations collected in Phase 1.5 as markdown hyperlinks (skills with Phase 1.5 applied)- Quantitative figure citation rule: All figures must have inline source attribution or
[추정]tag
Reference: moai-coworker:meta-skill-tester (plugins/moai-coworker/skills/meta-skill-tester/SKILL.md §Scoring rubric (quality criteria single source)
Reference: moai-coworker:meta-skill-template (plugins/moai-coworker/skills/meta-skill-template/SKILL.md (template)
Output: <target-plugin>/skills/<skill-name>/SKILL.md draft
Phase 4: Test Generation (test generation)
Generate test cases for skill validation.
Test composition:
tests:
- name: "happy-path"
prompt: "<대표적인 사용 프롬프트>"
expected_output:
format: "<출력 형식>"
contains: ["<필수 포함 내용>"]
not_contains: ["<금지 내용>"]
- name: "edge-case"
prompt: "<경계 조건 프롬프트>"
expected_output:
handles_gracefully: true
fallback_behavior: "<설명>"
- name: "complex-input"
prompt: "<복잡한 입력 프롬프트>"
expected_output:
completeness: "모든 요구사항 충족"
Output: <skill-dir>/tests/test-cases.yaml
Phase 5: Validation (quality validation)
Evaluate the skill with a 4-dimensional rubric.
| Dimension | Weight | Evaluation |
|---|---|---|
| Correctness | 30% | Goal achievement |
| Completeness | 25% | Edge case coverage |
| Clarity | 25% | User comprehensibility |
| Efficiency | 20% | Quality relative to tokens |
Pass criteria: Weighted average >= 0.70, all dimensions >= 0.50
If below threshold: Return to Phase 3 and supplement the deficient dimensions. Repeat up to 3 times.
Reference: moai-coworker:meta-skill-tester (plugins/moai-coworker/skills/meta-skill-tester/SKILL.md §Scoring rubric (anchor score table + evaluation procedure)
Phase 6: Review (final review)
Run the final checklist before deployment.
Checklist:
- Frontmatter format compliance (name, description only, no metadata)
- All required sections present (overview, triggers, workflow, examples, output, precautions, related skills)
- Trigger keywords do not duplicate existing skills
- Guide if plugin.json version needs updating
- Skill chain relationships match the project's skill chain definition (if applicable)
- Test cases are generated
- Rubric score passes at 0.70 or above
-
## 출처Section present (skills with Phase 1.5 applied,--skip-researchif unspecified) - All quantitative figures have sources or
[추정]tag (skills with Phase 1.5 applied)
Placement: <target-plugin>/skills/<skill-name>/SKILL.md Place the final file in .
Usage examples
Example 1: Creating a new skill
"Create a sales proposal auto-generation skill in moai-coworker. When a company name and target company are entered, a proposal DOCX should be produced"
Example 2: Creating a Vibe gap skill
"Use meta-skill-builder to create a UX design analysis skill in moai-coworker. Referencing the Vibe Designing feature"
Output format
| Deliverables | Location | Description |
|---|---|---|
| SKILL.md | <plugin>/skills/<name>/SKILL.md |
Completed skill definition |
| test-cases.yaml | <plugin>/skills/<name>/tests/ |
Test cases |
| score-report | In-memory | Rubric score report |
Precautions
- Not for modifying existing skills, creating new skillsused only for
- Plugin directory (
moai-*) must be checked in advance for existence - After skill creation, plugin.json version bump is handled collectively in the release process
- If trigger keywords overlap, the existing skill's description must be checked to clearly distinguish them
Rules for writing Korean example copy (slop prevention)
When adding Korean example copy to a new skill, the following 3 types are prohibited so that AI slop patterns are not transmitted through examples. This rule ai-slop-reviewer·korean-humanize fundamentally blocks patterns classified as slop by the gate from being registered as "good examples."
| # | Prohibited patterns | Description | [Bad example] (slop) | Recommended examples |
|---|---|---|---|---|
| 1 | Dash-contrast headline | Dash (—"X — Y" structure that splits a sentence with a dash (dash-contrast headline) |
[Bad example] "From copy-paste to delegation — just give the goal" | Write as complete sentences |
| 2 | Particle/noun-ending fragments | Fragment sentences ending with a particle (~은/에/로) or a noun (particle/noun ending) | [Bad example] "The key to success — automation" | Complete sentence with a predicate |
| 3 | "From A to B" transition formula | Introducing the "from X to Y" transition formula (slop cliché) | [Bad example] "From Excel to Notion" | Start with a concrete example |
Do not include this slop pattern in example copy — ensure the new skill does not become a source that teaches slop.
Related skills
| Skill | Relationship | Description |
|---|---|---|
| meta-skill-template | before | Base skill that provides the template structure |
| meta-skill-tester | after | Test execution + rubric scoring of the generated skill |
| ai-slop-reviewer | after | AI pattern review of the skill body (optional) |
Related commands
| Command | Description |
|---|---|
/harness |
Automatic new→test→review chain (orchestration of this skill + meta-skill-tester + ai-slop-reviewer) |
Source: revfactory/harness 6-Phase workflow (Apache 2.0) + MoAI adaptation
Changelog
- 1.6.0 (2026-05-01): Phase 1.5 Research newly established — external source research (WebSearch·Context7) made mandatory, domain-specific official source whitelist added,
## 출처section made mandatory + quantitative figure sources/[추정]tag enforced - 1.5.x: skill-forge → meta-skill-builder renamed
- Previous: absorbed harness 6-Phase workflow