사용 방법
- 설치 명령을 복사하고 자신의 에이전트 환경에 설치하기 전에 소스 패키지를 확인하세요.
- 필요한 도구, 인증 정보, 지원 환경은 원본 문서에서 확인하세요.
- 입력과 원하는 산출물을 설명하고 출처와 결과를 대조하세요.
npx skills add https://github.com/mblode/agent-skills --skill planning시작 전 확인
- 설치에는 모델 구독, API 한도, 파일 접근 권한이 포함되지 않습니다.
- 별점 수와 설치 수는 활동 지표이며 독립적인 품질·보안 검증이 아닙니다.
원본 저장소의 제목, 예시, 코드, 표, 링크, 이미지를 유지해 표시합니다.
Planning
Produce an executable plan, or strengthen an existing one. Planning alone produces no implementation edits. When the user asks to plan and implement, finish the plan and continue within the host's active mode and authorization.
- Create: no plan exists; ground the approach in the repository and write the plan.
- Review: a plan exists; verify its consequential claims and resolve concrete gaps in the file.
- Interview: the user asks to be grilled or interviewed; explore decisions interactively.
- Split: multiple independently deliverable outcomes need tickets with native dependency links.
For architecture contracts use codebase-architecture; for code findings use tidy.
References
| File | Read when |
|---|---|
references/decision-briefs.md | Presenting a consequential decision to the human: measuring options, previews, recording the answer |
references/interrogation-protocol.md | Choosing which question to ask about an unresolved choice, or the user requested an interview |
references/doc-grounding.md | ADRs, specifications, or library docs constrain the approach |
references/handoff-plans.md | Another session or person will execute the plan |
references/plan-quality-rubric.md | Reviewing completeness, feasibility, scope, testability, risk, and assumptions |
references/questioning-framework.md | A review gap needs a focused user question |
references/claim-verification.md | A plan claim can be checked against code or documentation |
references/splitting.md | Decomposing work into executable tickets |
Workflow
- Identify the requested outcome and authoritative plan path. Use the host's plan file where one exists. A durable handoff goes at the project path, default
docs/plans/<slug>.md; name which copy is authoritative. - Inspect the modules, tests, and decisions that constrain this change. Resolve questions the repository answers yourself. Ask only when an unresolved choice materially changes scope, behavior, or a hard-to-reverse action. Routine assumptions belong in the draft. Before asking, measure the options and name what would flip your recommendation; when that hinge is a fact only the human has, ask about it directly (
references/decision-briefs.md). - Choose the smallest vertical slice that exercises the real boundary. Name the existing code or platform capability it extends. A new dependency or abstraction needs a current requirement the existing mechanism cannot satisfy.
- Write the plan with the contract below. Review the consequential claims once against the rubric; fix evidenced gaps directly. Interview mode can explore competing approaches, but has no minimum question count.
- Return the plan path, unresolved decisions, and verification limits. Use the host's approval mechanism when its mode requires it. Do not add a second approval question for the plan's own review.
Plan contract
Include only sections this change needs:
- Outcome: triggering problem, intended behavior, and acceptance criteria.
- Approach: chosen slice, affected files or interfaces, and migration order where applicable.
- Decisions: a decision log: question, choice, who decided and when, measured evidence per option, and what would flip it; assumptions that remain unverified.
- Boundaries: exclusions only where an adjacent change would plausibly be mistaken for scope.
- Verification: a command, test scenario, or observation tied to each material acceptance criterion, including expected failure behavior.
- Docs: which current docs (those the repository marks as kept true, such as a "Current" entry in its docs index) this change makes wrong, updated in the same PR.
- Recovery: rollback or recovery for migrations and irreversible writes.
A handoff is self-contained. Replace "as discussed" with the decision. Preserve user corrections in the file, not just the chat. Split tickets by shippable outcome, not database/backend/frontend layers.
Review completion
Resolve gaps supported by code, the task, or operational constraints. Do not add speculative requirements to improve a self-score. Scores are optional unless requested; when used, mark unverified claims and explain residual gaps rather than iterating until every cell says 5/5.
Repeat review only after a substantive edit or new evidence. A user decision that remains unanswered is recorded at the affected step; continue independent work. If the user says to skip questions, draft from available evidence and label the assumptions.
Gotchas
- A plan in
~/.claude/plans/is not available to other checkouts or CI. Durable handoffs need a project artifact. - A bare "run tests" step does not establish the changed behavior. Name the acceptance scenario and expected result.
- Publishing slices without native blocker relations leaves the execution queue unaware of dependencies.
- A plan written for a prior revision can name moved files. Verify consequential paths and interfaces against the current checkout.
Maintenance only: evals/evals.json contains regression scenarios for changes to this skill; it does not load during a user task.




