AgentSkillsCN

plan

在正式编写代码之前，先制定具体的实施步骤并提交评审。当研究工作告一段落后，即可开始编写代码。

SKILL.md
--- frontmatter
name: plan
version: 1.0.0
changelog: Initial planning skill with handoff format
description: Create specific implementation steps for review. Use after research is complete and before writing code.
allowed-tools: Read, Write, Glob, Grep
temperature: 0.2  # Mostly deterministic, slight flexibility

Planning Phase

Create a specific implementation plan for review.

Task

$ARGUMENTS

Instructions

Output: ## Planning Phase

State: At phase start, update STATE.md:

  • •Set task: from $ARGUMENTS (if STATE.md is idle/complete or task is "None")
  • •Set phase: plan
  • •Set status: in_progress

1. Consume Research Handoff

Read ## Research Findings from STATE.md:

  • •Use Key Files as starting points
  • •Honor Constraints in plan design
  • •Address Remaining items from research

2. Gather Context

  • •Read files identified in research handoff
  • •Verify understanding is current
  • •Spawn parallel research tasks if gaps exist

3. Define Success Criteria

Automated verification:

  • •Build commands that must pass
  • •Test commands that must pass
  • •Type check commands

Manual verification:

  • •UI/UX checks requiring human review
  • •Performance validation
  • •Edge case review

4. Scope Boundaries

What we're doing:

  • •Specific deliverables

What we're NOT doing:

  • •Explicit exclusions (prevents scope creep)

5. Create Step-by-Step Plan

Use the plan format template:

markdown
## Phase 1: [Name]

- [ ] Step 1: Description
      File: `path/to/file.ts`
      Verification: how to verify this step

- [ ] Step 2: Description
      File: `path/to/other.ts`
      Verification: run `npm test`

## Phase 2: [Name]

- [ ] Step 3: ...

Task schema (from style.md): Each task should have:

  • •Action: What to do and why
  • •Files: Specific paths (src/auth.ts:42)
  • •Verify: Testable command
  • •Done: Measurable acceptance criteria

6. Identify Decision Points

  • •Where are there meaningful alternatives?
  • •What tradeoffs exist?
  • •What needs human input before proceeding?

7. Checkpoint

Update STATE.md incrementally:

  • •Set phase: plan
  • •Add plan decisions to ## Decisions
  • •Update ## Next Steps with plan summary and approval status

Run /checkpoint if context is heavy or taking a break.

8. Produce Handoff

Before exit gate, append handoff to STATE.md under ## Plan:

markdown
## Plan

### Completed
- [x] Implementation plan created
- [x] Success criteria defined

### Context
**Plan Summary**:
1. [Step 1 brief]
2. [Step 2 brief]
...

**Verification Commands**:
- `[test command]`
- `[build command]`

**Risks Identified**:
- [Risk]: [mitigation]

### Remaining
- [ ] Execute plan steps
- [ ] Verify each step before proceeding

See handoff template for full format.

Constraints

  • •Do not implement yet
  • •Plan should be specific enough that implementation is mechanical
  • •Each phase should be independently verifiable
  • •Keep total steps reasonable (3-10 per phase, 20 max overall)

Exit Criteria

Plan is approved by human before proceeding to implementation.

Gate: "Here's the plan. Approve to proceed?"