AgentSkillsCN

refine-plan

将 v1 项目计划细化为清晰明了的任务，为代理提供明确的上下文、实施步骤以及验证标准。可在 /plan-project 生成 v1 计划后使用此功能。

SKILL.md
--- frontmatter
name: refine-plan
description: >
  Refines a v1 project plan into agent-ready tasks with clear context,
  implementation steps, and validation criteria. Use after /plan-project
  has produced a v1 plan.
allowed-tools:
  - mcp__sequential-thinking__sequentialthinking
  - Read
  - Write
  - Glob
  - Grep
agent: plan-refiner
context: fork

Instructions

You are refining a v1 project plan into agent-ready tasks.

Input

  • •Read the v1 plan. The user will either provide the file path or you should look for the most recent plan file in the project root or docs/ directory.

Process

  1. •

    Analyze the plan using sequential-thinking MCP:

    • •Identify all distinct pieces of work
    • •Determine dependencies between them
    • •Identify the optimal execution order
    • •Flag any ambiguities or gaps in the v1 plan
  2. •

    For each task, produce:

    • •Title: Clear, action-oriented (e.g., "Implement user auth middleware")
    • •Context: Everything an agent needs to understand the task without reading the full plan. Include relevant architectural decisions, file paths, dependencies on other tasks, and constraints.
    • •Implementation steps: Numbered steps specific enough to follow, but not so prescriptive that they prevent smart problem-solving
    • •Validation criteria: How the agent (or a reviewer) confirms the task is complete. Should include at minimum:
      • •What to test / run
      • •Expected output or behavior
      • •Edge cases to verify
    • •Estimated complexity: S / M / L
    • •Dependencies: List of task IDs this depends on
  3. •

    Add a "Task 0" at the top of the plan:

    • •Title: "Generate Linear issues and update CLAUDE.md"
    • •Description: Instructions to run /create-issues followed by /setup-claude-md
    • •This ensures the first thing an engineer does is set up the project infrastructure

Output

  • •Write the refined plan to docs/project-plan-refined.md
  • •Use a consistent markdown structure (template below)

Output Template

code
# [Project Name] — Refined Plan

## Summary
[1-2 sentence summary of the project]

## Task Dependency Graph
[Mermaid diagram showing task dependencies]

## Tasks

### Task 0: Generate Linear Issues and Update CLAUDE.md
- **Context**: [...]
- **Steps**: Run `/create-issues docs/project-plan-refined.md`, then
  run `/setup-claude-md`
- **Validation**: All issues exist in Linear; CLAUDE.md passes best
  practices checklist

### Task 1: [Title]
- **Context**: [...]
- **Steps**: [...]
- **Validation**: [...]
- **Complexity**: [S/M/L]
- **Dependencies**: [None | Task IDs]

[...repeat for all tasks]

Validation Criteria

Read .claude/skills/_shared/validation-pattern.md and apply it. Specific criteria for this skill:

  • • Every task has all required fields filled in (title, context, steps, validation, complexity, dependencies)
  • • An agent reading any single task would have enough context to start working without asking clarifying questions
  • • Dependencies form a valid DAG (no circular dependencies)
  • • Task 0 is present and correctly references the next skills
  • • The refined plan covers all work items from the v1 plan (nothing dropped)