plan

Plan — Feature Planning & Implementation Plans

Decompose a feature into logical units and generate implementation plans for each unit.

This skill's outputs feed directly into /impl, where pseudocode becomes inline code comments. The clearer the pseudocode, the easier the final code review. Pseudocode expresses intent (WHAT), code expresses mechanism (HOW) — reviewers check alignment between the two.

Critical: This Skill Produces Documents, Not Code

This skill's sole purpose is to produce plan documents (docs/FEATURE_PLAN.md, docs/IMPL_PLAN_*.md). The deliverable is markdown files on disk — not a conversational summary, not a mental plan, not inline analysis.

If you find yourself analyzing code and forming a plan without writing documents, you are not following this skill. Stop and write the documents.

Plan Mode Awareness

When this skill runs inside Claude Code's Plan Mode, the environment tends to ask "shall I implement this?" after planning completes. This is misleading — the plan phase is NOT complete until documents are written to disk.

At the end of planning, always ask:

"계획 문서를 작성하겠습니다. 진행할까요?"

Do NOT ask "이대로 구현할까요?" or "코드를 작성할까요?" — those questions skip the document creation step. Implementation happens later via /impl, only after plan documents exist and the user has reviewed them.

When to Apply

• Designing or planning a feature
• Reviewing approach before implementation
• User invokes /plan

Process (Two Phases)

Phase 1: Generate FEATURE_PLAN.md

Step 1: Clarify Requirements

If the description is vague, ask up to 3 targeted questions. Wait for response.

Step 2: Scan Codebase

Check in order:

1. CLAUDE.md — project conventions
2. docs/PRD.md, docs/SPEC.md, docs/FEATURE_LIST.md — prior documents
3. Top-level project structure (ls only)
4. Key config files (package.json, build.gradle, go.mod, etc.)
5. Up to 5 relevant source files (Glob with keywords)

Summarize as CODEBASE_CONTEXT.

Step 3: Launch Analysis Agents

Environment Detection:

• If you have access to the Agent tool → use Option A (parallel)
• If the Agent tool is NOT available → use Option B (sequential)

Option A — Parallel (Agent tool available):

Run two agents simultaneously using the Agent tool.

Agent A — Risk Analyst Read ~/.claude/plugins/marketplaces/alexgim961101-claude-plugins/plugins/smart/skills/plan/agents/risk-analyst.md. Fill in {{FEATURE_DESCRIPTION}} and {{CODEBASE_CONTEXT}}.

Agent B — Architecture Reviewer Read ~/.claude/plugins/marketplaces/alexgim961101-claude-plugins/plugins/smart/skills/plan/agents/arch-reviewer.md. Fill in {{FEATURE_DESCRIPTION}} and {{CODEBASE_CONTEXT}}.

Option B — Sequential (Agent tool NOT available):

Perform each analysis inline, one at a time:

1. Read ~/.claude/plugins/marketplaces/alexgim961101-claude-plugins/plugins/smart/skills/plan/agents/risk-analyst.md. Perform the Risk Analysis inline, filling in {{FEATURE_DESCRIPTION}} and {{CODEBASE_CONTEXT}} with the values collected above. Write the result.
2. Read ~/.claude/plugins/marketplaces/alexgim961101-claude-plugins/plugins/smart/skills/plan/agents/arch-reviewer.md. Perform the Architecture Review inline, filling in {{FEATURE_DESCRIPTION}} and {{CODEBASE_CONTEXT}} with the values collected above. Write the result.

---

Extract RECOMMENDED_APPROACH from the Architecture Reviewer's ### 추천: block. If no clear recommendation, use fallback: "추천 접근법 미확정 — 별도 아키텍처 결정 필요"

Step 4: Generate FEATURE_PLAN.md

Write docs/FEATURE_PLAN.md with this structure:

# {기능명} — Feature Plan

> 생성일: {YYYY-MM-DD}
> 상태: 초안 (검토 필요)

---

## 목표

한 문장으로 이 기능의 목적을 정의한다.

---

## 배경 & 컨텍스트

코드베이스 컨텍스트 요약 (기존 관련 모듈, 기술 스택 등)

---

## 범위

**In Scope**
- ...

**Out of Scope**
- ...

---

{Agent A 리스크 분석 결과 — 그대로 삽입}

---

{Agent B 대안 검토 결과 — 그대로 삽입}

---

## 기능 분해

### 큰 기능 1: {이름}
> {한 줄 설명}

- **1-1. {작은 기능명}**: {설명}
- **1-2. {작은 기능명}**: {설명}
- **1-3. {작은 기능명}**: {설명}

### 큰 기능 2: {이름}
> {한 줄 설명}

- **2-1. {작은 기능명}**: {설명}
- **2-2. {작은 기능명}**: {설명}

---

## 구현 순서

의존 관계를 고려한 순서를 명시한다.
예: 1-1 → 1-2 → 2-1 → 1-3 → 2-2

---

## 미결 사항 (Open Questions)

- ...

Feature decomposition rules:

• Big features: logical groups by domain/concern
• Small features: independently implementable, reviewable in one sitting (1-2 day scope)
• No code — descriptions only
• Each small feature maps 1:1 to one IMPL_PLAN document

After saving, output:

• Save path
• Count of big features and small features
• Top 2-3 risks
• Recommended approach (one line)

Then ask: "기능 분해를 검토해 주세요. 수정할 부분이 있으면 말씀해 주세요. 확정되면 각 기능 단위별 구현 계획(IMPL_PLAN) 문서를 생성합니다."

Do not proceed to Phase 2 until user confirms.

Plan Mode 주의: 이 시점에서 Plan Mode를 종료(ExitPlanMode)하더라도, 코드 구현이 아닌 IMPL_PLAN 문서 작성이 다음 단계이다. "코드를 작성하겠습니다"가 아니라 "구현 계획 문서를 생성하겠습니다"로 안내한다.

---

Phase 2: Generate IMPL_PLAN Documents (After User Confirmation)

Once the user confirms the decomposition, generate one IMPL_PLAN per small feature unit.

Step 5: IMPL_PLAN Generation

Environment Detection:

• If you have access to the Agent tool → use Option A (parallel)
• If the Agent tool is NOT available → use Option B (sequential)

Context passed to each agent/analysis:

• CODEBASE_CONTEXT (shared — all agents receive the same)
• FEATURE_PLAN content (shared — for understanding relationships between units)
• RECOMMENDED_APPROACH (shared)
• Specific small feature unit description (individual)

Option A — Parallel (Agent tool available):

Spawn one Agent per small feature unit using the Agent tool. Independent units run in parallel. Agent instructions: read ~/.claude/plugins/marketplaces/alexgim961101-claude-plugins/plugins/smart/skills/plan/agents/impl-planner.md.

Option B — Sequential (Agent tool NOT available):

For each small feature unit, one at a time:

1. Read ~/.claude/plugins/marketplaces/alexgim961101-claude-plugins/plugins/smart/skills/plan/agents/impl-planner.md.
2. Perform the IMPL_PLAN generation inline with the context above.
3. Save the result before proceeding to the next unit.

Each agent generates: docs/IMPL_PLAN_{feature-id}.md

Examples:

• docs/IMPL_PLAN_1-1.md
• docs/IMPL_PLAN_1-2.md
• docs/IMPL_PLAN_2-1.md

Step 6: Consistency Check

After all agents complete, the parent verifies:

• No overlapping responsibilities between units
• Dependency ordering matches FEATURE_PLAN
• Naming conventions are consistent

Fix any issues found.

Step 7: Request Review

Output:

• List of generated IMPL_PLAN files
• One-line summary for each

Then ask: "구현 계획 문서가 생성되었습니다. 각 IMPL_PLAN 문서를 검토해 주세요. 확정되면 /impl로 구현을 시작할 수 있습니다."

이 시점이 /plan 스킬의 완료 지점이다. 사용자가 명시적으로 /impl을 호출하기 전까지 코드를 작성하지 않는다.

Quality Rules

• Be critical — challenge the user's initial assumptions
• Each small feature unit must be independently reviewable in one sitting
• IMPL_PLAN 문서에 코드 블록(```) 사용 금지 — 추천 커밋 메시지만 예외. 함수 시그니처, import문, 변수 선언 등 프로그래밍 언어 문법을 포함하지 않는다. 동작 흐름은 번호 매긴 한글 문장으로만 표현한다.
• Every IMPL_PLAN must include a recommended English commit message
• Pseudocode will become inline code comments during /impl, so write it clear and concise
• 실제 코드는 /impl 실행 시에만 코드베이스에 작성된다. 계획 문서는 순수하게 한글 설명만 포함한다.