一、目标

Cursor 的针对性配置方案目标：

① 针对性修改agent
② 简单 format / 变量名不浪费 token
③ 减少 Cursor 自动吃上下文
④ 修改或增加自定义 API 接入

Cursor 的正确定位应该是：

👉 “受控的代码 diff 工具”，而不是自由发挥的 Copilot
能用 editor / formatter / LSP 解决的，绝不让 AI 插手

二、修改 Cursor 默认 agent行为

1️⃣ 关闭“自动扩展上下文”

Cursor setting

{
  "cursor.general.enableDocumentationSearch": false,  // 关闭文档搜索
  "cursor.chat.streaming": true,
  "cursor.chat.defaultModel": "claude-sonnet-4.5",   // 用 Sonnet 而非 Opus

  
}

{
  // 限制上下文窗口
  "cursor.cpp.disabledLanguages": ["plaintext", "markdown"],
  "cursor.chat.contextLength": "short",  // short | medium | long
  
  // 禁用不必要的自动索引
  "cursor.general.gitIgnore": true,      // 遵循 .gitignore
  
  // 限制 @Codebase 范围
  "cursor.codebase.indexing.excludePatterns": [
    "**/node_modules/**",
    "**/dist/**",
    "**/build/**",
    "**/.next/**",
    "**/coverage/**",
    "**/*.min.js",
    "**/*.bundle.js"
  ]
}

2️⃣ 固定系统级 Prompt

项目根目录建：

.cursor/instructions.md：

You are an AI coding assistant.

Rules:
- Do NOT reformat code; formatting is handled by Prettier/Black.
- Do NOT rename variables unless explicitly requested.
- Only modify the selected files.
- Output: unified diff only.

侧重行为权限和默认模型策略

.cursorrules：

# Core Principles
- Only modify the specified files/functions; do not rewrite the entire file.
- Simple formatting or renaming tasks should be executed directly without explanation.
- Prefer incremental changes in Git diff style.
- Avoid outputting unrelated code context.

# Response Format
**For simple changes** (formatting, renaming):
- Output only the modified code snippet.
- Do not include confirmation messages like "Understood."
- Do not explain why the change was made.

**For complex changes**:
- Briefly describe the modification in 1-2 sentences.
- Output only the changed code blocks.
- Use `// ...` to indicate unchanged parts.

侧重输出规范、token 优化、简洁 diff

三、把 format / 命名从 AI 手里“夺权”

1️⃣ Format：100% 交给工具，不给 AI

强烈建议

Prettier / ESLint / Black / gofmt
保存即 format（on save）

Cursor 设置里：

Editor
→ Format on Save = ON

然后在 AI prompt 里明确禁止：

Do NOT format code.

AI 的 format = token 粉碎机 Formatter 的 format = O(1)

2️⃣ 变量命名：只在“新代码”阶段允许 AI

你的工作流应是：

场景	是否允许 AI 命名
新函数 / 新模块	✅
既有代码维护	❌
refactor	❌（除非明确）

Cursor 使用时的固定指令

Keep all existing variable and function names unchanged.

如果你不写这句，Cursor 一定会自作聪明。

四、限制 Cursor 的 token 黑洞行为

1️⃣ 不用“继续修改”“再改一下”

这是 Cursor 最浪费 token 的操作。

正确姿势：

❌「再改一下，加个判断」
✅ 重新下达完整指令

Cursor 是无状态 diff 工具，不是对话机器人。

2️⃣ 强制 Diff 输出（避免全文回传）

你每次写需求时，都加一句：

Output:
- unified diff only

否则 Cursor 会：

重复输出原始代码
消耗大量 token

五、修改 / 增加接入 API

1️⃣ Cursor 支持自定义 API Provider

路径：

Cursor Settings
→ Models / Providers
→ Add Custom Provider

你可以接入：

OpenAI-compatible API
Azure OpenAI
自建 proxy
第三方 Claude / Gemini 兼容接口

关键字段（示例）

{
  "baseUrl": "https://api.your-endpoint.com/v1",
  "apiKey": "sk-xxxx",
  "models": [
    {
      "name": "fast",
      "model": "gpt-4o-mini"
    },
    {
      "name": "reasoning",
      "model": "o1"
    }
  ]
}

然后在 Cursor 中：

format / 简单改动 → fast
复杂逻辑 → reasoning

2️⃣ 一个高级用法：双模型分工

你可以在 Cursor 里设：

Default Model：fast / cheap
手动切换：reasoning / expensive

原则：

90% 的 Cursor 操作，不值得用高智模型

六、“低 token Cursor 工作流”

1. Formatter 自动处理格式
2. 人判断是否需要 AI
3. Cursor：
   - 限文件
   - 禁 format
   - 禁命名
   - 出 diff
4. 复杂判断 → API / Web

当你发现：

prompt 越来越短
Cursor 很少“发挥”
修改一次成功率 > 80%

说明你已经把 Cursor 驯化成功了。

Cursor 不是“更聪明的编辑器”，而是“被严格约束的 AI 外包程序员”。

单文件临时场景：

1️⃣ 打开单文件
2️⃣ Cursor → Context → Limit to current file
3️⃣ 弹出临时 instructions：
     - Only this file
     - Do not format
     - Keep variable names
     - Unified diff
4️⃣ 调用 AI 修改
5️⃣ 用 formatter / lint 做格式化
6️⃣ Review diff → commit

效果：

AI 只用在增量修改 / 逻辑生成
token 最小化
单文件快速处理
不污染全 repo

针对 Cursor Agent 架构，把 Rules、Skills、Subagents、Command 的完整英文配置模板

1️⃣ Rules（行为约束 + 输出规范）

# Rules: Core Principles
- Only modify the selected files/functions; do not rewrite entire files.
- Simple formatting or renaming tasks should be executed directly without explanation.
- Prefer incremental changes in Git diff style.
- Avoid outputting unrelated code context.

# Response Format
**For simple changes** (formatting, renaming):
- Output only the modified code snippet.
- Do not include confirmation messages like "Understood."
- Do not explain why the change was made.

**For complex changes**:
- Briefly describe the modification in 1-2 sentences.
- Output only the changed code blocks.
- Use `// ...` to indicate unchanged parts.

2️⃣ Skills（任务模板 / 能力模块）

{
  "skills": [
    {
      "name": "refactor",
      "description": "Perform cross-file refactoring, keeping all unchanged code intact.",
      "rules": ".cursorrules.md",
      "instructions": [
        "Only modify selected functions or files.",
        "Output in unified git diff style.",
        "Do not change unrelated code."
      ]
    },
    {
      "name": "format",
      "description": "Perform simple formatting or variable renaming.",
      "rules": ".cursorrules.md",
      "instructions": [
        "Apply formatting or renaming directly without explanation.",
        "Do not output any confirmation text.",
        "Keep changes minimal."
      ]
    },
    {
      "name": "docgen",
      "description": "Generate documentation or comments for specified functions.",
      "rules": ".cursorrules.md",
      "instructions": [
        "Generate concise documentation for selected code only.",
        "Do not modify code structure.",
        "Use // ... for unchanged parts if including code examples."
      ]
    }
  ]
}

3️⃣ Subagents（模型/API 指定）

{
  "subagents": [
    {
      "name": "fast_code_agent",
      "model": {
        "type": "openai",
        "url": "https://api.openai.com/v1/chat/completions",
        "apiKey": "sk-xxxx",
        "modelName": "gpt-4o-mini"
      },
      "skills": ["format", "refactor"],
      "rules": ".cursorrules.md"
    },
    {
      "name": "reasoning_agent",
      "model": {
        "type": "openai",
        "url": "https://api.openai.com/v1/chat/completions",
        "apiKey": "sk-xxxx",
        "modelName": "gpt-4o",
        "temperature": 0
      },
      "skills": ["refactor", "docgen"],
      "rules": ".cursorrules.md"
    }
  ]
}

4️⃣ Command（实际执行指令模板）

{
  "commands": [
    {
      "name": "refactor_selected",
      "description": "Refactor selected functions/files with diff output only",
      "agent": "fast_code_agent",
      "skill": "refactor",
      "instructions": "Modify only the selected functions/files. Output unified git diff. Do not change unrelated code."
    },
    {
      "name": "format_code",
      "description": "Apply formatting or variable renaming directly",
      "agent": "fast_code_agent",
      "skill": "format",
      "instructions": "Apply formatting or renaming without explanation. Output only the modified code snippet."
    },
    {
      "name": "generate_doc",
      "description": "Generate documentation for selected code blocks",
      "agent": "reasoning_agent",
      "skill": "docgen",
      "instructions": "Generate concise documentation for selected code only. Use // ... for unchanged parts."
    }
  ]
}

✅ 特点与说明

Rules：约束行为和输出，保证 token 高效
Skills：能力模块化，便于不同任务调用
Subagents：绑定不同模型/API，区分 fast/ reasoning
Commands：直接映射到操作指令，便于 GUI 或快捷键执行

Cursor针对性配置