
Code Plan
- 1 installs
- 7 repo stars
- Updated August 5, 2026
- dmlguq456/claude_setting
Creates a detailed codebase-grounded implementation plan via a subagent, with QA review levels from quick to adversarial (Korean output).
About
Creates a detailed implementation plan grounded in the actual codebase via a plan-team subagent, with scalable QA review rounds. A developer uses it to plan a coding task before implementation.
- Delegates plan creation to a plan-team subagent reading real source files
- Scales QA review from quick to adversarial with an iterative review loop
Code Plan by the numbers
- 1 all-time installs (skills.sh)
- Ranked #2,479 of 3,282 Productivity & Planning skills by installs in the Skillselion catalog
- Data as of Aug 5, 2026 (Skillselion catalog sync)
npx skills add https://github.com/dmlguq456/claude_setting --skill code-planAdd your badge
Show developers this skill is listed on Skillselion. Paste this into your README.
| Installs | 1 |
|---|---|
| repo stars | ★ 7 |
| Last updated | August 5, 2026 |
| Repository | dmlguq456/claude_setting ↗ |
What it does
Creates a detailed codebase-grounded implementation plan via a subagent, with QA review levels from quick to adversarial (Korean output).
Files
Caller note: planning benefits fromhighorxhigheffort; lower effort may miss call sites in cross-file analysis.
Language Rule
- All user-facing output in natural Korean (no translationese — write Korean natively, don't translate from an English draft).
Pre-Check
Check if a similar plan already exists in .claude_reports/plans/. Behavior depends on plan status:
active: Always ask the user — "기존에 진행 중인 plan이 있습니다. 이어서 진행할까요, 새로 만들까요?" Do NOT proceed until confirmed.done/failed: Note it for reference and auto-proceed with new plan creation (no prompt).partial: Auto-create a new plan covering only the failed steps (readfailed_stepsfrom plan frontmatter). No prompt.
Record any user-facing pause for the pipeline_summary.md Decision Points table.
Delegate to 기획팀
Invoke the plan-team (기획팀) agent as a subagent with the following prompt:
Plan mode. Create a new implementation plan.
Task: {$ARGUMENTS}
Save English plan to: .claude_reports/plans/{YYYY-MM-DD}_{short-task-name}/plan/plan.md
Date: {YYYY-MM-DD}
{If a done/failed/partial plan exists: "Reference previous plan: [path], status: [status]"}
{If partial: "Failed steps from previous execution: [list from plan frontmatter failed_steps]"}
Read all relevant source files, analyze the current state, and create the plan.
Write the plan files directly. Return ONLY the file paths and a 3-5 line Korean summary. Do NOT return the plan content itself.The agent writes the plan file directly; the orchestrator only receives paths and a summary.
QA Scaling
If $ARGUMENTS contains --qa quick|light|standard|thorough|adversarial, use that level and strip the flag from the task description. Otherwise, auto-detect from the plan's scope. When qa_level is set in plan frontmatter, it overrides auto-detect.
| Level | Auto-detect condition | Action |
|---|---|---|
| Quick | (manual only — never auto-selected) | 1× 품질관리팀 (model: "sonnet"), single pass, max 1 review round (no iteration even if 🔴 found — 🔴 are recorded as 미해결 이슈 and loop exits) |
| Light | ≤3 steps, mechanical, single-variant | 1× 품질관리팀 (model: "sonnet") |
| Standard | 4-10 steps, logic changes, single module | 1× 품질관리팀 (default opus) |
| Thorough | >10 steps, cross-module/variant, architectural | 2-3× 품질관리팀 in parallel: Agent A correctness (opus), B completeness (sonnet), C risk (opus, optional, >15 steps); each writes round_{N}_{focus}.md; all 🔴 issues must be resolved |
| Adversarial | Cross-variant (SE+SS+CSS), shared modules (utils/, network.py), or >20 steps with architectural impact — AND Codex available | Thorough-level 품질관리팀 + 1× codex-review-team (adversarial-review) in parallel; Codex writes round_{N}_codex.md; all 🔴 from ANY agent (including Codex) must be resolved |
Codex availability check: Before selecting Adversarial, run codex --version (suppress stderr). If the command fails or Codex is not authenticated, fall back to Thorough silently. This check is skipped if --qa adversarial is explicitly specified (fail loudly instead).
Post-Plan Review Loop (max 3 revision rounds; quick = 1 round)
The log directory is the task root folder (parent of plan/). Example: .claude_reports/plans/2026-03-18_task/plan/plan.md → log dir is .claude_reports/plans/2026-03-18_task/. Run mkdir -p {log_dir}/_internal/plan_reviews before invoking QA.
Round counting: Initialize round = 0. A round = one plan-team fix → QA review cycle; all parallel Thorough agents count as one round. Increment round only when QA is re-invoked after a revision. "max 3 rounds" means 기획팀 is invoked at most 3 times to fix issues. `quick` mode: max rounds = 1 — after the single review pass, exit regardless of 🔴 (record residuals as 미해결 이슈 and skip the fix-round).
QA level lock: QA level is determined once at loop start; only upward escalation allowed (no downgrade). If --qa was NOT specified, the orchestrator MAY upgrade once (starting round 2) when 🔴 count ≥3 in the one-line verdict (no review file reading needed); round counter does NOT reset. If --qa was manually specified, no change allowed. quick is never auto-upgraded (user opted in for fastest path).
After the 기획팀 agent returns: 1. Assess QA level from plan scope per the QA Scaling table above. 2. Invoke 품질관리팀: Prompt: "Review this plan in plan review mode for feasibility. Plan file: [plan_path]. Write review results to: [log_dir]/_internal/plan_reviews/round_{N}.md. Return ONLY the file path and a one-line verdict."
- Light: pass
model: 'sonnet'. Thorough: 2-3 parallel agents with focus suffix and separate output files; passmodel: 'sonnet'for the B (completeness) agent, default opus for A (correctness) and C (risk). Do NOT read the review file unless relaying verdict to user.
3. Check one-line verdict:
- No 🔴: Loop ends → proceed to Korean Version Generation.
- 🔴 found AND qa_level == quick: Loop ends (no fix-round). Invoke 기획팀: "Refine mode. Add 🔴 issues from {log_dir}/_internal/plan_reviews/round_1.md to the plan's 리스크 section under ## 미해결 이슈. Return brief Korean summary." Then proceed to Korean Version Generation.
- 🔴 found: Re-invoke 기획팀: "Refine mode. Plan file: {plan_path}. QA review: {log_dir}/_internal/plan_reviews/round_{N}.md. Fix 🔴 issues. Return only changed steps + brief Korean summary." Increment
round, re-invoke 품질관리팀. Repeat until no 🔴 orround >= 3.
4. If 🔴 remain after `round >= 3`: Auto-proceed — invoke 기획팀: "Refine mode. Add remaining 🔴 issues to the plan's 리스크 section under ## 미해결 이슈. Return brief Korean summary." Then report to user: plan path, resolved issues, and unresolved issues with reasons.
Record any user-facing pause (e.g., active-plan ambiguity) so the pipeline skill can surface it in pipeline_summary.md.
Mirror Generation (편집팀 — conditional)
코드 plan 은 _코드 식별자 + 단계 설명_ 묶음 — primary language 는 English (코드 자체가 영문 자연). 한국어 사용자 검토용 mirror 가 보통 필요. 사용자가 영문 plan 만 본다고 명시한 경우 mirror 생성 skip.
Skip condition: 사용자가 영문 plan 만 검토한다고 명시 또는 영문 사용자.
Trigger (default for 한국어 사용자): plan.md 영문 + 한국어 mirror 필요.
모드 A — 영문에서 국문으로 옮기기.
영문 plan 경로: {plan_path}
국문 출력 경로: {same directory}/plan_ko.md
~/.claude/agents/editorial-team.md 의 모드 A 절차를 따른다.
~/.claude/agents/editorial-team.md 의 판교체 회피 절을 강제 적용. 사용자 표기 선호는 `mem profile 02_paper_writing_style` 보조 참조.
코드 식별자·파일 경로·라이브러리 이름은 영어 그대로, 그 외 일반 표현은 한국어로.
section 제목 매핑: Goals → 목표, Current State → 현황 분석, Change Plan → 변경 계획, Risks → 리스크, Verification → 검증 방법.
완료 시 파일 경로 + 한국어 요약 3-5 줄 + 의도적으로 한 표기 결정 한두 개만 돌려준다.Then report to the user: plan path(s) + summary + QA verdict.
Task
$ARGUMENTS
code-plan
본 README 는SKILL.md의 GitHub 표시용 mirror. 권위 있는 동작 명세는SKILL.md.
개요
실제 코드베이스를 기반으로 상세 구현 계획을 작성하는 skill. 기획팀에 위임하여 영어 plan.md 생성 + QA 루프 후 한국어 plan_ko.md 전체 번역.
호출 형식
/code-plan <task description> [--qa light|standard|thorough|adversarial] [--autonomy proactive|standard|passive]Caller note: 계획 수립은high/xhigheffort에서 이점. 낮은 effort에서는 cross-file 분석에서 호출 지점 누락 가능.
언어 규칙
- 사용자 출력은 자연스러운 한국어 (번역체 회피)
Pre-Check — 기존 plan 상태 게이팅
--autonomy 플래그 파싱 후 .claude_reports/plans/ 유사 plan 존재 확인:
| 기존 상태 | 처리 |
|---|---|
active (Critical — 항상 질문) | "기존에 진행 중인 plan이 있습니다. 이어서 진행할까요, 새로 만들까요?" 확인 전 진행 금지 |
done/failed (Significant) | proactive: 참고용 기록 후 자동 생성. standard/passive: "이전에 완료/실패한 plan이 있습니다. 새로 생성할까요?" |
partial (Significant) | proactive: 자동으로 실패 step만 커버하는 새 plan. standard/passive: 사용자에게 질문 |
위임 — 기획팀
Plan mode. Create a new implementation plan.
Task: {$ARGUMENTS}
Save English plan to: .claude_reports/plans/{YYYY-MM-DD}_{short-task-name}/plan/plan.md
Date: {YYYY-MM-DD}
{If done/failed/partial plan exists: "Reference previous plan: [path], status: [status]"}
{If partial: "Failed steps from previous execution: [list]"}기획팀이 plan 파일을 직접 씀. 오케스트레이터는 경로와 요약만 받음.
QA Scaling
| Level | 조건 | 행동 |
|---|---|---|
| Light | ≤3 steps, 기계적, 단일 variant | 1× 품질관리팀 (sonnet) |
| Standard | 4-10 steps, 로직 변경, 단일 모듈 | 1× 품질관리팀 (opus, 기본) |
| Thorough | >10 steps, cross-module/variant, 아키텍처 | 2-3× 품질관리팀 병렬 (opus): A 정합성 / B 완전성 / C 리스크 |
| Adversarial | Cross-variant + Codex 가용 | Thorough + 1× codex-review-team 병렬 |
Codex 가용성 체크: Adversarial 선택 전 codex --version 실행. 실패 시 Thorough로 silent fallback (--qa adversarial 명시는 fail loudly).
Post-Plan Review Loop (최대 3 리비전 라운드)
로그 디렉토리 = task root (plan/의 부모). mkdir -p {log_dir}/plan_reviews 먼저.
라운드 카운팅: round = 0. 한 라운드 = 기획팀 수정 → QA 리뷰. Thorough 병렬 에이전트도 한 라운드. QA 재호출 시 round 증가.
QA level lock: 루프 시작 시 확정, 상향만 허용. --qa 미지정 시 라운드 2부터 🔴 ≥3이면 한 번 상향 허용 (라운드 리셋 없음).
절차: 1. QA level 평가 2. 품질관리팀 호출 (Light: sonnet / Thorough: 2-3 병렬 + 다른 focus) 3. verdict 확인:
- 🔴 없음 → Korean Version Generation
- 🔴 있음 → 기획팀 refine → QA 재호출.
round >= 3까지 반복
4. 3 라운드 후 🔴 잔여:
- proactive: 자동 진행 — 기획팀이 남은 🔴을
## 미해결 이슈에 추가 - standard/passive: "QA 3라운드 후에도 🔴 N개. 리스크 섹션 추가할까요?"
Korean Version Generation
리뷰 루프 종료 후 기획팀 Translate 모드 최종 호출:
Translate mode. English plan file: {plan_path}. Save Korean version to: {same dir}/plan_ko.md.
Full Korean translation (NOT a summary). Section titles: 목표, 현황 분석, 변경 계획, 리스크, 검증 방법.
Code identifiers stay in English. Return ONLY the file path.사용자에게 영·한 plan 경로, 요약, QA verdict 보고.
--- 원본: `~/.claude/skills/code-plan/SKILL.md`