writing-plans（写作计划）

Writing Plans

Write comprehensive implementation plans assuming zero codebase context. Document: which files to touch, code, testing, docs, how to test. Bite-sized tasks. DRY. YAGNI. TDD. Frequent commits. Assume skilled developer, minimal domain knowledge, weak test design.

Announce at start: "I'm using the writing-plans skill to create the implementation plan." Context: Run in a dedicated worktree (created by brainstorming skill). Save plans to: docs/plans/YYYY-MM-DD-<feature-name>.md using the Write tool. NEVER write to ~/.claude/plans/ — ignore the plan mode default path. REQUIRED: Before ExitPlanMode, run Plan Verification Checklist, then rule-of-five-plans. Verify *what* before polishing *how* -- scope errors caught early save wasted polish on deleted tasks.

Mandatory Tasks (Enforcement)

Create these 7 native tasks at plan start (each blocked by previous via addBlockedBy). Cannot call ExitPlanMode with pending tasks:

1. Write draft plan -- Initial structure with all tasks, dependencies, file lists

- After saving plan — context check: Look at your model ID in the system prompt (e.g., claude-opus-4-6[1m]). - [1m] present (default): Skip compact. Announce "Plan written to {path}. Proceeding to verification." and continue to task 2. - No [1m] (200k): Show copy-pasteable /compact command (see references/announcements-protocol.md) and wait for user's follow-up before proceeding to task 2.

1. Plan Verification Checklist -- Complete/Accurate/Commands valid/YAGNI/Minimal/Not over-engineered
2. Rule-of-five-plans: Draft pass -- Shape and structure
3. Rule-of-five-plans: Feasibility pass -- Can every step be executed? Deps available? Paths valid?
4. Rule-of-five-plans: Completeness pass -- Every requirement traced to a task?
5. Rule-of-five-plans: Risk pass -- What could go wrong? Migration, breaking changes?
6. Rule-of-five-plans: Optimality pass -- Simplest approach? YAGNI?

Tasks 2-7: Sub-Agent Dispatch. After proceeding from task 1 (immediately on 1M, after user "continue" on 200k), dispatch each pass sequentially as a sonnet sub-agent using the template in references/verification-dispatch.md. Mark each native task in_progress before dispatch, completed after collecting verdict. If any verdict is BLOCKED/FAIL, stop and report to user. After all 6 verdicts collected, assemble Verification Record (see references/verification-footer.md) and append to plan file.

See references/task-enforcement-examples.md for full TaskCreate blocks and dispatch loop.

Bite-Sized Task Granularity

Each step is one action (2-5 min): write failing test, run it to verify failure, implement minimal code, run test to verify pass, commit.

Plan Document Header

Every plan MUST start with:

# [Feature Name] Implementation Plan

> **For Claude:** After human approval, use plan2beads to convert this plan to a beads epic, then use `superpowers-bd:subagent-driven-development` for parallel execution.

**Goal:** [One sentence]
**Architecture:** [2-3 sentences]
**Tech Stack:** [Key technologies]
**Key Decisions:**
- **[Area]:** [Choice] -- [Why over alternatives]
---

Key Decisions: 3-5 decisions implementers might question. WHAT was decided AND WHY. Focus where alternatives existed.

File Structure

Before defining tasks, map out which files will be created or modified and what each one is responsible for. This is where decomposition decisions get locked in.

| File | Responsibility | Action |
|------|---------------|--------|
| `exact/path/to/file.py` | One clear responsibility | Create |
| `exact/path/to/existing.py` | One clear responsibility | Modify |
| `tests/exact/path/to/test.py` | Tests for file.py | Create |

• Design units with clear boundaries and well-defined interfaces. Each file should have one clear responsibility.
• Prefer smaller, focused files. Files that can be held in context at once produce more reliable edits.
• Files that change together should live together. Split by responsibility, not by technical layer.
• In existing codebases, follow established patterns. If a file has grown unwieldy, including a split in the plan is reasonable.

This structure informs task decomposition. Each task's Files: section must reference entries from this table — no task should introduce files not listed here.

Task Structure

CRITICAL: Every task MUST include Depends on:, Complexity:, and Files: sections. These enable safe parallel execution, model selection, and file conflict detection. See references/dependency-analysis.md and references/file-lists.md.

Recommended context: Purpose: (why), Not In Scope: (prevents overbuilding), Gotchas: (quirks).

### Task N: [Component Name]
**Depends on:** Task M, Task P | None
**Complexity:** simple | standard | complex
**Files:**
- Create: `exact/path/to/file.py`
- Modify: `exact/path/to/existing.py:123-145`
- Test: `tests/exact/path/to/test.py`

**Purpose:** [One sentence]

**Step 1: Write the failing test**
[code block]
**Step 2: Run test to verify it fails**
Run: `pytest tests/path/test.py::test_name -v`
Expected: FAIL
**Step 3: Write minimal implementation**
[code block]
**Step 4: Run test to verify it passes**
Run: `pytest tests/path/test.py::test_name -v`
Expected: PASS
**Step 5: Commit**
`git add ...` then `git commit -m "feat: add specific feature"`

Plan Verification Checklist

Before rule-of-five-plans, verify scope and accuracy:

• Complete -- All requirements from brainstorming addressed?
• Accurate -- File paths verified? (existing files exist, new files in correct locations)
• Commands valid -- Test/build commands correct and runnable?
• YAGNI -- Every task directly serves a stated requirement?
• Minimal -- Could any task be removed/combined without losing functionality?
• Not over-engineered -- Simplest approach that works?
• Key Decisions documented -- 3-5 decisions with rationale?
• Context sections present -- Purpose for non-obvious tasks? Not In Scope for boundary tasks?
• File Structure complete -- Every file in task Files: sections appears in File Structure table? No undeclared files?

Remember

• Exact file paths always
• Complete code in plan (not "add validation")
• Exact commands with expected output
• Reference relevant skills by name
• Every task needs Depends on:, Complexity:, and Files:
• Include Purpose:, Not In Scope:, Gotchas: where needed
• File Structure table before tasks — tasks reference it, never introduce undeclared files
• Run Plan Verification Checklist before rule-of-five-plans
• Announce each verification phase (see references/announcements-protocol.md)
• Plan MUST end with Verification Record (see references/verification-footer.md)
• After approval, follow execution handoff (see references/execution-handoff.md)

Reference Files

• references/task-enforcement-examples.md: Full TaskCreate blocks with blocked-by relationships
• references/dependency-analysis.md: Identifying and expressing task dependencies
• references/file-lists.md: File list format, parallel execution rules, and complexity estimation
• references/announcements-protocol.md: Required announcement templates for verification phases
• references/verification-dispatch.md: Sub-agent prompt templates and dispatch flow for verification passes
• references/verification-footer.md: Plan Document Footer template (Verification Record)
• references/execution-handoff.md: Post-plan workflow: plan2beads -> subagent-driven-development