implement

name	implement
description	Execute an approved design plan step by step with validation and checklist tracking. Use when user says '実装して', 'implement this plan', 'design-*.mdを実装', or has an approved plan file ready for implementation. Do NOT use for planning, research, or implementing without a plan file.
allowed-tools	Agent, AskUserQuestion, Bash, Read, Edit, Write, Glob, Grep
argument-hint	<design file path> [--steps '1,3,5' to run specific steps only]
disable-model-invocation	true
metadata	{"author":"ysk1031","version":"1.0.0"}

Implementation Skill

Execute an approved implementation plan from start to finish. Read the plan, implement each step, update the checklist, and run validation continuously — without stopping until all steps are complete.

Instructions

Phase 1: Plan Loading & Tooling Detection (use Agent with subagent)

Call the Agent tool with:

• subagent_type: "plan-reader"
• description: "load plan and detect tooling"
• prompt: Replace $ARGUMENTS in the agent's loaded prompt with the actual user arguments and execute.

---

Phase 2: Scope Confirmation (main agent)

If Phase 1 returned "status": "NO_PLAN", display the message field and stop.

1. Read the full plan file with the Read tool.
2. Extract the implementation steps and checklist.

Display the implementation scope:

## 実装計画

**ファイル**: <plan file>
**ステップ数**: 残り <remaining> / 全 <total>
**検出ツール**: type check: <tooling.typecheck != "NONE" ? yes : no>, lint: <tooling.lint != "NONE" ? yes : no>, test: <tooling.test != "NONE" ? yes : no>

### チェックリスト
<display checklist with current status>

If selected_steps is not "ALL", show only the selected steps.

Use AskUserQuestion:

• question: "実装を開始しますか？"
• header: "Implement"
• options:
  1. label: "すべて実装", description: "残りの全ステップを実装します"
  2. label: "ステップを選択", description: "実装するステップを指定します（Otherで番号をカンマ区切り入力）"
  3. label: "キャンセル", description: "実装せずに終了します"

If "すべて実装": Proceed to Phase 3 with all remaining (unchecked) steps. If "ステップを選択": User provides step numbers via "Other". Proceed to Phase 3 with selected steps only. If "キャンセル": Print "実装を終了しました。" and stop.

---

Phase 3: Implementation Loop (main agent)

This is the core execution phase. Do NOT stop until all selected steps are complete or a critical error occurs.

Read the CLAUDE.md if it exists to follow project conventions during implementation.

For each remaining step (in order):

Step 3.1: Read Step Details

Re-read the plan file to get the detailed instructions for the current step. Each step should have:

• Target file and action (create / modify)
• Specific changes to make
• Reasoning and details

Step 3.2: Implement

Based on the step action:

modify:

1. Read the target file with the Read tool
2. Apply changes using the Edit tool
3. If the edit is too large or complex, use Write to rewrite the file

create:

1. Verify the parent directory exists: ls <parent-dir>
2. Write the new file with the Write tool

delete (if specified):

1. Confirm with the user before deleting
2. Use Bash rm to delete

For each implementation:

• Follow existing code patterns identified in the plan
• Do NOT add unnecessary comments, JSDoc, or type annotations beyond what the plan specifies
• Do NOT add features, error handling, or abstractions not in the plan
• Keep changes minimal and focused on what the step requires

Step 3.3: Validate

After each step, run available validation commands:

1. Type check (if available): Run the detected type check command via Bash
2. Lint (if available): Run the detected lint command via Bash

If validation fails, follow the detailed failure handling flow in references/validation-flow.md (relative to this skill's directory). Read it with the Read tool and follow the instructions.

If validation passes, continue to the next step without pausing.

IMPORTANT: Do NOT run the full test suite after every step — only run type check and lint. Tests are run once in Phase 4.

Step 3.4: Update Checklist

After each step completes successfully (including validation):

1. Read the plan file
2. Update the checklist: change - [ ] Step N: ... to - [x] Step N: ...
3. Write the updated plan file using Edit

Then immediately proceed to the next step. Do NOT ask for user confirmation between steps.

---

Phase 4: Verification & Summary (main agent)

After all selected steps are complete:

1. Run full test suite (if available):

<detected test command>

If tests fail:

• Display the failure output

• Use AskUserQuestion:

  • question: "テストが失敗しました。修正を試みますか？"
  • header: "Test Failure"
  • options:
    1. label: "修正する", description: "失敗したテストを分析して修正します"
    2. label: "スキップ", description: "テスト失敗を無視して完了とします"

  If "修正する": Analyze the test failure, fix the issue, re-run tests. Repeat until tests pass or user chooses to skip. If "スキップ": Continue to summary.

2. Show summary:

git diff --stat

3. Display completion message:

## 実装完了

### 変更されたファイル
<git diff --stat output>

### チェックリスト
<display updated checklist — completed steps show [x], skipped steps show [ ] with [SKIPPED] annotation, not-attempted steps show [ ]>

### 検証結果
- Type check: <PASS / FAIL / N/A>
- Lint: <PASS / FAIL / N/A>
- Test: <PASS / FAIL / SKIPPED / N/A>

<if any steps were skipped>
### スキップされたステップ
<list skipped steps with their validation error summary>
</if>

### 次のステップ
- 動作確認を行ってください
- 問題なければ `/commit` でコミットできます

---

Rules

• ALWAYS display messages in Japanese — the user is a Japanese speaker and needs to review progress and errors in Japanese
• NEVER stop between steps unless a critical error occurs — keep going until all steps are done — stopping between steps increases user wait time and significantly slows implementation
• NEVER commit or push changes — only apply file modifications — commits should only happen after user review; unintended commits are difficult to undo
• NEVER add comments, JSDoc, or type annotations not specified in the plan — additions not in the plan create noise in code review and diverge from the approved plan
• NEVER add features or abstractions beyond what the plan specifies — out-of-scope changes can introduce unexpected bugs and architectural distortions
• ALWAYS update the checklist after each step — without an up-to-date checklist, the resume point is unknown if implementation is interrupted
• When a step is skipped due to validation failure, NEVER mark it as completed — leave it unchecked with a [SKIPPED] annotation — marking a skipped step as complete misleads the user about implementation status
• When implementation is aborted, ALWAYS show a partial completion summary before stopping — the user needs to know which steps succeeded and which remain to plan their next action
• ALWAYS run type check and lint after each step (if available) — validating after each step makes it easier to identify the cause of issues and reduces fix cost
• Run the full test suite only ONCE at the end (Phase 4), not after every step — running tests is time-consuming; running after every step significantly increases total implementation time
• If a step's target file doesn't exist for a "modify" action, flag it and ask the user — forcing changes to a non-existent file would create an unintended new file
• If the plan file changes during implementation (e.g., user edits it), re-read it before each step — implementing based on an outdated plan fails to reflect the user's revisions
• Respect CLAUDE.md conventions if the file exists — code violating project conventions will be rejected during lint/review
• Keep the implementation faithful to the plan — do not deviate or "improve" beyond what's specified — deviating from the plan bypasses the approval process and introduces unintended changes

---

Examples

Example 1: 計画全体の実装

User says: "design-auth.md を実装して" Actions:

1. plan-reader が計画ファイルを読み込み、ツーリングを検出
2. チェックリストと実装スコープを表示
3. ユーザー確認後、全ステップを順番に実装
4. 各ステップ後にtype check/lintを実行
5. 全ステップ完了後にテストを実行し、サマリーを表示 Result: 計画に沿った実装が完了し、チェックリストが更新される

---

Troubleshooting

"NO_PLAN"

Cause: 指定された計画ファイルが見つからない Solution: ファイルパスを確認し、design-*.md 形式のファイルを指定

バリデーションが繰り返し失敗

Cause: 自動修正では解決できない問題がある Solution: 「手動で修正して続行」を選択して自分で修正するか、「スキップして次へ」で一旦進む