Claude Code の出力が長すぎるときの設定:短く、エンジニアリング報告らしくする

Claude Code の出力が長すぎるときの実践設定。短い一時プロンプト、CLAUDE.md、output style、custom slash command、出力フォーマット制約で、結果、検証、リスクを先に出すようにします。

Claude Code を使っていると、小さな悩みが出てきます。仕事はできるのに、返答が長すぎることです。

テストを一つ直してほしいだけなのに、テストフレームワークを説明し、依頼を言い直し、「これからやります」と書き、最後に変更ファイルを言います。コードは問題なくても読むコストが高く、長い会話では context を圧迫します。

解決策は「簡潔に」と言うだけではありません。一時プロンプト、CLAUDE.md、output style、custom slash command を分けて使うのが安定します。

目標は、Claude Code の出力を工程報告に近づけることです。結果を先に、前置きを減らし、task restatement を避け、必要な files、commands、risks、next steps だけを残します。

Quick Answer

レイヤー 場面 方法
一時プロンプト 今の turn が長い result、files、validation、risks だけ要求
CLAUDE.md project 長期 preference concise report rules を書く
Output style 全体の tone と format output style を定義
Custom slash command review、test fix、summary 短い template を command 化

おすすめ順は、一時プロンプト、CLAUDE.md、output style、slash command です。

今すぐ短くする

1
2
3
4
5
6
后续回复请压缩:
- 不复述我的需求。
- 不写礼貌铺垫。
- 先给结论。
- 只列改动文件、验证命令、结果、风险。
- 没有风险就写“未发现额外风险”。

コード変更では:

1
2
3
4
5
6
7
请用简短工程汇报格式输出:
1. 改了什么
2. 改了哪些文件
3. 跑了什么验证
4. 还有什么风险

不要解释通用背景,不要复述任务。

調査では:

1
2
3
4
5
6
只输出:
- 最可能原因
- 证据
- 下一步验证命令

不要写长篇分析过程。

これはすぐ効きますが、会話単位です。長期化するなら project rules に移します。

CLAUDE.md に書く

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
## Response Style

- Be concise and engineering-focused.
- Do not restate the user's request.
- Do not include generic explanations unless asked.
- Put the result first.
- For code changes, report only:
  1. changed files
  2. behavior changed
  3. validation run
  4. remaining risks
- If no validation was run, say so directly.
- If no extra risk is found, write "未发现额外风险".

中国語プロジェクトなら:

1
2
3
4
5
6
7
8
9
## 回复风格

- 默认用中文,简洁、直接、工程化。
- 不复述用户需求。
- 不写礼貌铺垫。
- 不解释通用背景,除非用户明确要求。
- 修改代码后只汇报:改动文件、行为变化、验证命令、剩余风险。
- 没有运行验证时必须明确说明。
- 没有额外风险时写“未发现额外风险”。

CLAUDE.md 自体が長くなると context の負担になります。圧縮方法は /ja/2026/07/10/ai-coding-prompt-compression-guide/。

Output Style を使う

output styles は Claude の response style、role、tone、format を変えます。project facts を教えるものではありません。

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
# Concise Engineering

Respond like a senior engineer giving a short handoff.

Rules:
- Lead with the result.
- Avoid restating the task.
- Avoid generic explanations.
- Use short bullets only when useful.
- For code work, include changed files, validation, and risks.
- If blocked, state the blocker and the next concrete action.
- Keep final answers short unless the user asks for detail.

中国語なら:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
# 简洁工程汇报

默认用中文回答,风格像资深工程师交接工作。

规则:
- 先说结果。
- 不复述用户需求。
- 不写通用背景。
- 不写“我将会”这类过程铺垫。
- 代码修改后只汇报改动、验证、风险。
- 遇到阻塞时只说明阻塞点和下一步。
- 用户没有要求长解释时,回答保持短。

project paths、test commands、forbidden files は output style ではなく CLAUDE.md に置きます。

/config で設定する

Claude Code で:

1
/config

output style の項目を確認します。バージョンによっては次のような形式を使える場合があります。

1
/config outputStyle=your-style-name

key 名は変わる可能性があるため、まず /config で確認するのが安全です。

Custom Slash Command

固定タスクには command を作ります。

1
.claude/commands/short-review.md
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
只读审查当前改动,不要修改文件。

输出格式:

1. 高风险问题
2. 中低风险问题
3. 缺失测试
4. 结论

规则:
- 每个问题最多 2 行。
- 必须给文件路径。
- 没有问题时写“未发现高风险问题”。
- 不要复述任务。
- 不要写通用审查方法。

実行:

1
/short-review

同様に short-fixshort-testshort-summary を作れます。

コード変更後の出力形式

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
结果:
- ...

改动文件:
- path/to/file

验证:
- 命令:...
- 结果:...

风险:
- ...

さらに短く:

1
2
3
4
改动:...
文件:...
验证:...
风险:...

検証していない場合:

1
验证:未运行,原因是 ...

Review で短くする

1
2
3
4
5
6
7
只读审查当前 diff。

只输出真实问题,不输出优点。
按严重程度排序。
每个问题包含:文件路径、问题、建议。
没有问题就写“未发现高风险问题”。
不要写总结性夸奖。

Codex を独立 review に使う場合、Claude Code は実装、Codex は read-only review にできます。分担: /ja/2026/07/10/cursor-codex-claude-code-workflow-division/。

Test fix で短くする

1
2
3
4
5
6
7
请修复当前 failing test。

输出限制:
- 不解释测试框架。
- 不复述错误日志。
- 只保留失败原因、改动文件、验证结果。
- 连续失败两次后停止修改,先总结共同原因。

token usage が高いときは log output も制御します: /ja/2026/07/08/claude-code-token-usage-spike-troubleshooting-faq/。

すべてを極短にしない

短い出力が向くのは bug fix、failing tests、commit summary、code review、simple command explanation、明確な実装タスクです。

短すぎないほうがよいのは onboarding、architecture tradeoff、incident review、security risk、database migration、曖昧な requirements です。

実行フェーズは短く、設計フェーズは少し展開してよいと考えます。

caveman との違い

caveman は Claude Code や Codex の出力を短くする skill/plugin として紹介しました: /ja/2026/07/03/caveman-agent-output-compression-guide/。

この記事の方法は基礎です。一時プロンプト、CLAUDE.md、output style、slash command で自分の rule を作ります。

よくある間違い

“Be concise” だけ

曖昧です。

1
只保留:结果、文件、验证、风险。

毎回 style を繰り返す

安定ルールは CLAUDE.md や output style に移します。

output style に project facts を入れる

output style は tone と format、project facts は CLAUDE.md です。

safety を削る

危険操作、database、permissions、production config、file deletion は省かないでください。

説明を完全禁止する

良い書き方:

1
默认短输出;只有遇到设计取舍、风险或阻塞时再解释。

Copy-Paste Config

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
## Response Style

- Default language: Chinese.
- Be concise and engineering-focused.
- Do not restate the user's request.
- Do not include generic background unless asked.
- For code changes, report only:
  - result
  - changed files
  - validation
  - risks
- For reviews, list only actionable issues.
- For test failures, summarize the failing cause and next command only.
- If validation was not run, say that explicitly.
- If no risk is found, write "未发现额外风险".

一時プロンプト:

1
2
3
本轮请使用短汇报格式:
结果 / 文件 / 验证 / 风险。
不要复述任务,不要解释通用背景。

まとめ

Claude Code の出力が長すぎるとき、すぐ道具を変える必要はありません。現在の会話は短い prompt、project は CLAUDE.md、全体表現は output style、固定タスクは slash command で制御します。

大事なのは「できるだけ短く」ではなく、場面に合う長さです。実装とテスト報告は短く、architecture と risk discussion は必要に応じて展開します。

参考資料

记录并分享
Hugo で構築されています。
テーマ StackJimmy によって設計されています。