| name | blueprint |
| description | Agentic system design blueprint generator. Interviews the user to understand a task they want to automate, then produces a comprehensive integrated design document (.md file) that serves as a concrete implementation plan for Claude Code. The design document includes task context, workflow definition with LLM vs code boundaries, and implementation spec (folder structure, agent architecture, skills/scripts list, sub-agent design). Trigger on "/blueprint", "blueprint", "에이전트 설계", "설계서 만들어", "agentic workflow design". Use this skill when the user mentions automation design, workflow planning, agent architecture, system blueprinting, or multi-step AI workflow planning. |
Blueprint
When NOT to use this skill
Simple task planning or to-do organizing — where a design document is not the deliverable — does not need this skill. Use blueprint only when the user wants an agentic system design document as the output.
Overview
Conduct a structured interview to understand the user's automation task, then generate a complete agentic system design document. The deliverable is a single .md file ready for use as an implementation reference in Claude Code.
Before Starting
Read both reference files before doing anything else. They contain the document structure and design rules you'll apply in Phase 2. Skip them and the output will be incomplete.
- Read
references/document-template.md — the full section-by-section template for the output file
- Read
references/design-principles.md — design rules for agent structure, validation, data transfer, skill vs sub-agent
- (Optional) Skim
references/example-blueprint.md — a fully annotated sample blueprint document for calibration
Workflow
Phase 1: Assess & Interview
Evaluate user input against these three areas. Ask only about gaps. If all sufficiency criteria are met, skip directly to Phase 2.
| Area | Sufficiency criteria (all must be answerable) | Example question |
|---|
| Goal & success criteria | (1) 완료 상태를 판별하는 구체적 기준이 1개 이상 (2) 실패 상태를 판별할 수 있는가 | "어떤 결과가 나와야 이 에이전트가 성공했다고 볼 수 있나요?" |
| Task procedure | (1) 입력 형식과 출처가 특정됨 (2) 출력 형식과 저장 위치가 특정됨 (3) 분기 조건이 있다면 판단 기준이 명확함 | "A 이후 B로 갈지 C로 갈지는 어떤 기준으로 판단하나요?" |
| Tools & constraints | (1) 사용할 도구/API가 정해졌거나 추천 가능 (2) 기술적 제약(API 한도, 포맷 제한 등)이 파악됨 | "지금 쓰는 도구가 있나요? 없다면 이런 방식들이 가능한데 어떤 게 맞을까요?" |
Agent organization(단일 vs 멀티 에이전트)은 인터뷰 대상이 아니다. design-principles.md의 "Agent Structure Choice" 기준에 따라 에이전트가 Phase 2에서 자체 결정하고, Phase 3에서 근거를 설명한다.
Interview rules:
- Questions must be specific and probing, never generic or formulaic
- If user says "모르겠다" or "알아서 해줘": apply reasonable defaults, state your choice and reasoning, ask only for unavoidable decisions
- Group related questions — never ask more than 3 questions per turn
- Maximum 4 turns total. If gaps remain after 4 turns, apply reasonable defaults, document them as assumptions in the design document, and proceed to Phase 2
Phase 2: Generate Design Document
Once requirements are clear, map interview findings to document sections before writing:
| Interview finding | → Document section |
|---|
| Why this is needed, what problem it solves | § 1. 작업 컨텍스트 › 배경 및 목적 |
| What's in scope / out of scope | § 1. 작업 컨텍스트 › 범위 |
| Input format, output format, trigger | § 1. 작업 컨텍스트 › 입출력 정의 |
| Technical constraints, API limits | § 1. 작업 컨텍스트 › 제약조건 |
| Domain-specific terminology | § 1. 작업 컨텍스트 › 용어 정의 |
| Step-by-step process, branching logic | § 2. 전체 흐름도 + 단계별 상세 |
| What the agent decides vs what code handles | § 2. LLM 판단 vs 코드 처리 구분 |
| State transitions between steps | § 2. 상태 전이 |
| What tools/APIs are used | § 3. 스킬/스크립트 목록 |
| Failure conditions, retry expectations | § 2. 단계별 상세 › 실패 시 처리 |
에이전트가 자체 결정하는 섹션 (인터뷰 불필요): § 3. 에이전트 구조 (design-principles.md 기준 적용), § 3. 폴더 구조, § 3. CLAUDE.md 핵심 섹션 목록, § 3. 주요 산출물 파일. 이 섹션들은 인터뷰 결과를 종합하여 에이전트가 추론으로 채운다.
Fill every section using the template in references/document-template.md. Apply design rules from references/design-principles.md.
Save as blueprint-<task-name>.md in the current working directory.
Output rules:
- CLAUDE.md, AGENT.md, skill file contents are NOT written — only their names and roles
- Implementation spec covers structure and responsibilities, not code or prompts
- Every workflow step must have: success criteria, validation method, failure handling
- Always include a "CLAUDE.md 작성 원칙" section in the implementation spec. Apply the 4 principles (Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution) from
references/design-principles.md › "CLAUDE.md / AGENTS.md 작성 원칙". Each principle must include a self-verification test. Include tradeoff statement and success metrics.
- Always include a "스킬 생성 규칙" section in every blueprint document. All skills defined in this document must be created via
skill-creator at implementation time — direct manual authoring of SKILL.md is prohibited (see references/design-principles.md › Skill Creation Standards for the exact section text). The document must contain the literal string skill-creator — the structural validator checks for this and will fail without it.
Completeness check before saving — confirm each item is filled:
Phase 2.5: Validate Document
After saving, run the structural validation script before presenting the document to the user.
Do not assume a relative path from the target project — run the script from the currently installed blueprint skill path.
python ~/.claude/skills/blueprint/scripts/validate_blueprint_doc.py ./blueprint-<task-name>.md
python .claude/skills/blueprint/scripts/validate_blueprint_doc.py ./blueprint-<task-name>.md
If validation fails, fix the document and run again. This script checks structure only (required sections, step fields, and implementation/workflow section presence) — it does not check content quality.
If the script is not found (e.g., first install or missing file), skip script validation and manually verify the Phase 2 completeness checklist instead:
Phase 3: Review
After presenting the document, summarize the key design decisions:
- Agent structure choice (single vs multi) and the reason
- Any tradeoffs locked in (e.g., "LLM judges step X because rule-based detection was too fragile")
- Any constraints that shaped the design
- Any assumptions applied (from "모르겠다" responses or turn limit)
Then ask: "이 결정들이 의도에 맞는지 확인해 주세요. 변경할 부분이 있으면 말씀해 주세요."
If user says "괜찮아" / "잘 모르겠는데 진행해" — confirm the specific assumptions that will be locked in: "다음 가정들이 확정됩니다: [list]. 구현 후 변경하면 비용이 크니 지금 확인해 주세요." Then proceed.
Apply any requested changes and re-confirm.
References
references/document-template.md: Full template for the output design document (all sections, formats, tables)
references/design-principles.md: Design rules for folder structure, agent architecture, validation patterns, failure handling, data transfer, and skill vs sub-agent distinctions