AgentSkillsCN

writing-plans

当一项多步骤任务已具备经验证的设计,且在接触代码之前,您需要一份详尽且可评审的实施计划时,运用此技能。

SKILL.md
--- frontmatter
name: writing-plans
description: >
  Use this skill when a multi-step task has a validated design and you need a
  detailed, reviewable implementation plan before touching code.

Writing Plans

Intent

Create a detailed, test-driven implementation plan that is easy to execute, review, and recover after interruptions.


When to Use

  • After a design/spec has been approved.
  • Before any multi-step implementation or refactor.
  • When work will be executed by multiple agents or sessions.

Precondition Failure Signal

  • Work starts without a plan in docs/plans/.
  • Tasks are too large to complete in 2-5 minutes.
  • Plan lacks exact file paths, commands, or verification steps.

Postcondition Success Signal

  • Plan exists in docs/plans/YYYY-MM-DD-<name>.md using skills/writing-plans/templates/plan_template.md.
  • Tasks are atomic and follow RED -> GREEN -> REFACTOR -> COMMIT.
  • Each task includes exact file paths and verification commands.
  • Tasks are only marked Done after verification steps are executed and evidence is recorded.
  • The user approves the plan before execution begins.
  • Required review evidence is recorded before marking the plan Delivered.
  • Plan is only marked Delivered with explicit user permission.

Process

  1. Source Review: Read the design and identify impacted components.
  2. Structure: Use skills/writing-plans/templates/plan_template.md.
  3. Decompose: Split into 2-5 minute tasks.
  4. Specify: For each task, include:
    • Exact file paths (create/modify/test).
    • RED/GREEN/REFACTOR steps.
    • Commands to run and expected outcomes.
    • Evidence capture requirements for verification output.
    • Commit guidance.
  5. Verification: Ensure plan tasks include explicit verification steps.
  6. Approval: Present plan and wait for user approval.
  7. Review: Request review from relevant review personas for plan quality and boundary alignment, and record their collated feedback, agreed next steps, and any follow-up plan link in the plan document.
  8. Deliver Gate: Ask for explicit user permission before marking the plan Delivered.
  9. Handover: Offer execution via executing-plans or subagent-driven-development (if tasks are independent).

Example Test / Validation

  • Plan includes at least 3 tasks, each with exact file paths and test commands.

Common Red Flags / Guardrail Violations

  • Tasks labelled "Implement X" without steps or commands.
  • Missing verification steps or expected outputs.
  • Plans that skip RED-GREEN-REFACTOR.
  • Proceeding to execution without approval.
  • Marking a plan Delivered before required review evidence is recorded.
  • Capturing review feedback outside the plan without summarizing it there.
  • Marking Delivered without explicit user permission.

Recommended Review Personas

  • Tech Lead - validates architecture alignment and scope.
  • Software Engineer - validates task granularity and feasibility.
  • Lead Developer - validates plan quality and sequencing.
  • Architect - validates boundaries, colocation, and reuse decisions.

Skill Priority

P2 - Consistency & Governance


Conflict Resolution Rules

  • Execution skills must not start until the plan is approved.
  • If requirements change, return to planning.

Conceptual Dependencies

  • test-driven-development
  • brainstorming

Classification

Governance
Core


Notes

Plans are the recovery point for interruptions. Make them explicit and complete.