Composition Patterns
Organize combination patterns and design guidelines for MCP × Skill × Agent.
About This Document
While MCPs and Skills have value individually, there is value that emerges only through combination. This document defines 4 composition patterns and organizes the design guidelines, trigger contexts, and use cases for each.
For the MCP construction strategy, see mcp-roadmap.md. For the Skill construction strategy, see skill-roadmap.md.
4 Composition Patterns
| Pattern | Composition | Core Concept | Difficulty |
|---|---|---|---|
| MCP + Skill | 1 MCP + 1 Skill | Data retrieval + Judgment criteria | ★★☆ |
| MCPs | Multiple MCPs | Data integration & comparison | ★★☆ |
| Skills | Multiple Skills | Knowledge layering | ★☆☆ |
| MCPs + Skills | Multiple MCPs + Multiple Skills | Complete workflow | ★★★ |
Pattern 1: MCP + Skill
A composition where "raw data" retrieved by MCP from external sources is evaluated using "judgment criteria" held by Skills. The basic form of hybrid composition.
Design Principle
MCP returns "what is written," and Skill determines "whether it is good or bad." This separation is important.
Concrete Examples
Electronic Signature Act Compliance Check
| Element | Role |
|---|---|
hourei-mcp | Retrieve Electronic Signature Act article text (MCP) |
rfcxml-mcp | Retrieve RFC 3161 technical requirements (MCP) |
| E-Signature Act Skill | Mapping criteria between legal and technical requirements (Skill) |
OAuth/JWT Implementation Review
| Element | Role |
|---|---|
rfcxml-mcp | Retrieve RFC 6749 (OAuth 2.0) / RFC 7519 (JWT) specifications |
| OAuth/JWT Skill | Token management best practices, implementation patterns |
Design Guidelines
- Write Skills independent of MCP output: Avoid implementation-dependent descriptions like "look at the
article_textfield in hourei-mcp's JSON response." Instead, use abstract descriptions like "verify whether the legal text contains the following requirements" - Make Skill judgment criteria verifiable: Clearly define the 3 states: "compliant," "non-compliant," and "cannot determine (additional information needed)"
Candidate List
| Theme | MCP | Skill | Status |
|---|---|---|---|
| Electronic Signature Act | hourei-mcp + rfcxml-mcp | Implementation guidelines | ⚡ MCP side already available |
| Personal Information Protection Act | hourei-mcp | Compliance checklist | ⚡ MCP side already available |
| Electronic Books Preservation Act | hourei-mcp | Storage requirements check | ⚡ MCP side already available |
| GDPR | Regulatory text MCP (planned) | DPIA checklist | 🔲 Not started |
| OWASP | Vulnerability DB MCP (planned) | Security review criteria | 🔲 Not started |
| OAuth/JWT | rfcxml-mcp | Implementation patterns | ⚡ MCP side already available |
Pattern 2: MCPs (Multi-MCP Coordination)
A composition that integrates, compares, and verifies data retrieved from multiple MCPs. There are already proven examples in workflows/patterns.md.
Design Principle
Proven Combinations
| Combination | Use Case | Track Record |
|---|---|---|
| deepl-mcp + xcomet-mcp | Translation → Quality evaluation → Re-translation | 180-page technical document translated in 1 day / $12 |
| rfcxml-mcp + w3c-mcp | Integrated reference of RFC specs + Web API definitions | WebSocket specification verification workflow |
| rfcxml-mcp + hourei-mcp | Technical specification × Legal text mapping | Electronic Signature Act × RFC 3161 correspondence table |
| pdf-spec-mcp + pdf-reader-mcp | PDF specification reference + Actual file verification | PDF/UA conformance verification |
Design Guidelines
- Clarify whether order-dependent or parallelizable: deepl → xcomet is order-dependent; rfcxml + w3c is parallelizable
- The Agent layer handles data integration logic: MCPs only return raw data. Integration and judgment are the Agent's role
Pattern 3: Skills (Multi-Skill Combination)
A composition that simultaneously triggers multiple Skills according to the context (development phase).
Design Principle
Context-Based Skill Sets
Design Guidelines
- Define priority order among Skills: In the Code Review set, evaluate in order: SOLID (principles) → Clean Code (style) → Refactoring (improvements)
- "Primary source of definition" rule: For concepts with overlapping scope (e.g., SRP), designate one Skill as the primary source and the other as reference
- Entire sets can be meta-Skill-ized: It is possible to define "Code Review Set" itself as a single Skill that internally references the three Skills
Pattern 4: MCPs + Skills (Full Integration)
The most complex but most powerful composition where multiple MCPs and multiple Skills cooperate.
Design Principle
Concrete Examples
Security Audit Workflow
| Layer | Element | Role |
|---|---|---|
| MCP | rfcxml-mcp | Retrieve RFC 6455/6749 specification requirements |
| MCP | w3c-mcp | Retrieve WebSocket API definition |
| Skill | OWASP Skill | Vulnerability judgment criteria |
| Skill | OAuth Skill | Authentication implementation best practices |
Specification Compliance Quality Gate
Glossary-Integrated Translation Workflow (Production Example)
deepl-glossary-translation is a production example of Pattern 4. Three MCPs and a Skill collaborate to translate the PDF specification (ISO 32000-2) into Japanese with consistent terminology.
| Layer | Component | Role |
|---|---|---|
| MCP | pdf-spec-mcp | Extract 71 terms structurally from spec |
| MCP | deepl-mcp | Glossary management & translation |
| MCP | xcomet-mcp | Quality score & error detection |
| Skill | deepl-glossary-translation | 5-step workflow procedure definition |
What this Skill defines is the knowledge of "which MCP to call, in what order, with what parameters" — not the tools themselves. See Skill Showcase and Workflow Pattern 9 for details.
Design Guidelines
- Agent orchestration is key: The Agent layer determines which MCPs to call first and when to reference which Skills
- Parallelize MCP calls: Independent MCP calls should be executed in parallel to efficiently use the context window
- Skill trigger timing: Clarify whether to judge collectively after data retrieval or to judge incrementally
Pattern Selection Guide
Accumulating Results
Practical results from each pattern are recorded in outputs.md. Include the following information.
| Item | Content |
|---|---|
| Pattern Used | Which of the 4 patterns |
| Components | MCPs and Skills used |
| Input | What was targeted |
| Results | Quantitative outcomes (time, cost, quality scores, etc.) |
| Learnings | Improvements and suggestions for next time |
Related Documents
- mcp-roadmap.md — MCP Construction Roadmap
- skill-roadmap.md — Skill Construction Roadmap
- workflows/patterns.md — Existing Workflow Patterns (Pattern 2 track record)
- concepts/03-architecture.md — MCP / Skill / Agent Layer Structure
- mcp/catalog.md — Built MCP Catalog