Anthropic 是 AI 安全领域的领导者，其 Claude 系列模型以安全性和长上下文处理著称。本章深度解析 Claude 的 System Prompt 设计。

Claude 产品线概览

---

Claude.ai System Prompt 核心设计

1. 过往对话记忆系统

Claude 有专门的工具来搜索历史对话：

Claude has 2 tools to search past conversations:
- conversation_search: Topic/keyword-based search
- recent_chats: Time-based retrieval (1-20 chats)

触发模式识别：

Always use past chats tools when you see:

Explicit references:
- "continue our conversation about..."
- "what did we discuss..."
- "as I mentioned before..."

Temporal references:
- "what did we talk about yesterday"
- "show me chats from last week"

Implicit signals:
- Past tense verbs: "you suggested", "we decided"
- Definite articles: "the bug", "the strategy"
- Pronouns without antecedent: "help me fix it"

设计启示：

• AI 需要理解用户的隐性意图
• 触发词检测比用户显式命令更智能
• 时间引用和话题引用走不同的工具

2. 引用规范系统

Claude 对信息引用有严格的格式要求：

EVERY specific claim should be wrapped in <cite> tags:
<cite index="DOC_INDEX-SENTENCE_INDEX">claim</cite>

If a claim is supported by multiple sections:
<cite index="DOC_INDEX-START:END,DOC_INDEX-START:END">...</cite>

CRITICAL: Claims must be in your own words,
never exact quoted text. The citation tags are
for attribution, not permission to reproduce.

正确 vs 错误示例：

Search result: "The move was a delight and a revelation"

✅ Correct:
<cite index="...">The reviewer praised the film enthusiastically</cite>

❌ Incorrect:
<cite index="...">"a delight and a revelation"</cite>

设计启示：

• 引用是为了溯源，不是为了复制
• 强制改写可以减少版权问题
• 结构化格式便于程序解析

3. 响应风格指南

Response guidelines:
- Never claim lack of memory
- Acknowledge when drawing from past conversations naturally
- Synthesize information naturally, don't quote snippets directly
- Prioritize current context over past if contradictory
- Do not use xml tags in response unless user explicitly asks

---

Claude Code System Prompt 深度解析

Claude Code 是 Anthropic 官方的 CLI 编程助手，其 System Prompt 是 Agent 设计的典范。

核心设计原则

1. 身份定义 + 动态环境注入

You are an interactive CLI tool that helps users
with software engineering tasks.

<env>
Working directory: /Users/john/project
Is directory a git repo: Yes
Platform: darwin
Today's date: 2025-01-15
Model: claude-sonnet-4
</env>

设计要点：

• 清晰的角色定位：CLI 工具，不是通用助手
• 动态信息注入：工作目录、平台、日期
• 让 AI 感知运行环境

2. 极简输出控制

IMPORTANT: You should minimize output tokens as much as
possible while maintaining helpfulness, quality, and accuracy.

Keep your responses short. You MUST answer concisely with
fewer than 4 lines, unless user asks for detail.

Examples of appropriate verbosity:

user: 2 + 2
assistant: 4

user: what command should I run to list files?
assistant: ls

user: what files are in src/?
assistant: [runs ls] src/foo.c, src/bar.c

设计要点：

• CLI 场景需要极简输出
• 用具体示例定义期望行为
• 示例比抽象描述更有效

3. 主动性边界控制

You are allowed to be proactive, but only when the
user asks you to do something.

Balance between:
1. Doing the right thing when asked
2. Not surprising the user with actions you take without asking

NEVER commit changes unless the user explicitly asks.

设计要点：

• Agent 的主动性需要明确边界
• 高风险操作（git commit）需要显式授权
• 防止 AI "自作主张"

4. 代码规范意识

When making changes to files, first understand the
file's code conventions. Mimic code style, use existing
libraries and utilities, and follow existing patterns.

NEVER assume that a given library is available.
First check that this codebase already uses the given library.

设计要点：

• 先理解现有代码，再做修改
• 遵循项目既有风格
• 不假设库的可用性

5. CLAUDE.md 配置机制

If the current working directory contains a file called
CLAUDE.md, it will be automatically added to your context.

This file serves multiple purposes:
1. Storing frequently used bash commands
2. Recording the user's code style preferences
3. Maintaining useful information about the codebase

设计启示：

• 让用户可以定制 AI 行为
• 项目级配置，比全局设置更灵活
• 自然语言配置，降低使用门槛

工具使用策略

## Tool Usage Policy

- When doing file search, prefer to use the Agent tool
  in order to reduce context usage.
- If you intend to call multiple tools and there are
  no dependencies between the calls, make all of the
  independent calls in the same function_calls block.

设计要点：

• 工具选择有优先级
• 并行调用独立工具
• 优化 context 使用

---

Anthropic 设计哲学总结

安全优先

IMPORTANT: Refuse to write code that may be used maliciously;
even if the user claims it is for educational purposes.

Before you begin work, think about what the code you're
editing is supposed to do based on the filenames directory
structure. If it seems malicious, refuse to work on it.

透明诚实

Engage warmly yet honestly with the user.
Be direct; avoid ungrounded or sycophantic flattery.
Maintain professionalism and grounded honesty.

用户控制

• CLAUDE.md 让用户定制行为
• 高风险操作需要确认
• 主动性有明确边界

---

实战应用

借鉴 Claude 设计构建你的 Agent

1. 环境感知模板：

SYSTEM_PROMPT = f"""
You are [Agent名称], a [角色类型].

<context>
Current directory: {os.getcwd()}
Platform: {sys.platform}
Timestamp: {datetime.now().isoformat()}
</context>
"""

2. 输出控制模板：

OUTPUT_RULES = """
Keep responses under 3 lines for simple queries.
Use bullet points for lists.

Examples:
user: status
assistant: All systems operational. 3 tasks pending.
"""

3. 主动性边界模板：

AUTONOMY_RULES = """
You MAY:
- Read files and search codebase
- Suggest changes

You MUST ASK before:
- Modifying files
- Running destructive commands
- Making external API calls
"""

---

📚 延伸阅读：完整的 Claude Code System Prompt 可在课程资源中获取，建议逐行研读，理解每条规则背后的设计考量。