System Prompt Case Studies
System Prompt Case Library
A collection of real system prompt designs from major AI companies. Study these cases to understand industry best practices, then apply them to your own AI Agent development.
Case Overview
| Company | Product | Core Feature | Learning Value |
|---|---|---|---|
| Anthropic | Claude.ai, Claude Code | RLHF safety design, multi-role | Agent safety boundary design |
| OpenAI | GPT-4o, Agent Mode | Function Calling, message channels | Tool calling conventions |
| Gemini CLI, Guided Learning | Project conventions, workflows | Code agent best practices | |
| xAI | Grok 3/4, Personas | Persona system, X platform integration | Personalized role design |
| Others | Perplexity, Kagi, Raycast | Search strategies, format specs | Vertical domain design |
Anthropic Claude
Claude.ai & Claude Code
Core design philosophy: Safe, helpful, honest
Key features:
- RLHF safety mechanisms: Constitutional AI design
- CLAUDE.md configuration: Users can customize agent behavior
- Tool specifications: Detailed Tool Use constraints
- Multi-role switching: Adapts to different scenarios via system prompt
Best suited for:
- Coding assistant development
- Long document processing
- Applications requiring high safety
Typical design pattern:
You are Claude, made by Anthropic...
You can use tools to complete tasks...
NEVER do X without explicit permission...
Deep dive into Claude System Prompts
OpenAI GPT
GPT-4o & Agent Mode
Core design philosophy: General-purpose, flexible, rich ecosystem
Key features:
- Function Calling: TypeScript Namespace-style tool definitions
- Message channel system: analysis / commentary / final separation
- Financial activity restrictions: Explicit Allowed / Not Allowed lists
- Safe browsing rules: Defense against Prompt Injection
Best suited for:
- Browser automation
- Complex multi-step tasks
- Applications requiring tool calling
Typical design pattern:
namespace tools {
type function_name = (_: { param: string }) => any;
}
Deep dive into GPT System Prompts
Google Gemini
Gemini CLI & Guided Learning
Core design philosophy: Project conventions first, workflow-driven
Key features:
- Project conventions first: NEVER assume — analyze existing code first
- Five-step workflow: Understand → Plan → Implement → Verify Tests → Verify Standards
- Guided learning: Socratic teaching method
- Self-verification loops: Emphasis on testing and standards checks
Best suited for:
- Code agent development
- Educational AI applications
- Scenarios requiring strict workflows
Typical design pattern:
## Software Engineering Tasks
1. Understand: Think about the user's request...
2. Plan: Build a coherent plan...
3. Implement: Use available tools...
4. Verify (Tests): Run project tests...
5. Verify (Standards): Run linting...
Deep dive into Gemini System Prompts
xAI Grok
Grok 3/4 & Persona System
Core design philosophy: Personalization, real-time information, X platform integration
Key features:
- Persona system: Multi-personality role switching
- X platform integration: Native X search and analysis support
- Render components: Dedicated citation and formatting system
- Real-time information: No strict knowledge cutoff date
Available Personas:
- Companion
- Unhinged Comedian
- Loyal Friend
- Homework Helper
- Not a Doctor / Not a Therapist
Best suited for:
- Social media applications
- Personalized chatbots
- Applications requiring real-time information
Typical design pattern:
You are Grok, a [personality] chatbot...
## Style Rules:
- match the user's vulgarity
- always write in lowercase
- use abbreviations like rn ur bc
Deep dive into Grok System Prompts
Other AI Products
Perplexity, Kagi, Raycast AI
Three products, three distinct approaches:
| Product | Positioning | Core Design |
|---|---|---|
| Perplexity | AI search engine | Real-time search strategy, voice interaction specs |
| Kagi Assistant | Premium search assistant | Most detailed format specification system |
| Raycast AI | Desktop productivity tool | User system preference injection |
Learning value:
- Perplexity: How to design search-oriented AI (re-search on every follow-up)
- Kagi: How to design detailed output format specifications
- Raycast: How to integrate user system preferences (language, timezone, units)
Deep dive into other AI product System Prompts
Design Pattern Comparison
From these cases, we can distill several core design patterns:
1. Identity Definition Patterns
| Company | Style |
|---|---|
| Claude | "You are Claude, made by Anthropic" |
| GPT | "You are ChatGPT, a large language model..." |
| Grok | "You are Grok, a [persona] chatbot..." |
2. Tool Definition Patterns
| Company | Style |
|---|---|
| OpenAI | TypeScript Namespace |
| Anthropic | XML-format Tool Definition |
| Grok | XML Function Call |
3. Safety Boundary Patterns
| Company | Approach |
|---|---|
| Claude | NEVER / ALWAYS keywords |
| GPT | Allowed / Not Allowed lists |
| Grok | Boundaries (Never Do) sections |
How to Apply These Cases
1. Choose a Reference Template
Pick the best reference based on your use case:
- Coding assistant → Reference Claude Code, Gemini CLI
- Search application → Reference Perplexity
- Chatbot → Reference Grok Personas
- Browser automation → Reference GPT Agent Mode
2. Reuse Design Patterns
Directly reuse proven design patterns:
SYSTEM_PROMPT = """
# Identity (reference: Claude)
You are [Your Agent Name], a [role] assistant.
# Tools (reference: OpenAI)
You have access to the following tools...
# Constraints (reference: Grok)
## Boundaries (Never Do):
- Never do X without permission
- Never share Y information
# Output (reference: Kagi)
Format your response with proper markdown...
"""
3. Iterate and Optimize
- Start with a simple version
- Observe agent behavior
- Gradually add constraints and examples
- Test edge cases
- Refine safety boundaries
Further Reading
- System Prompt Design in Practice - 10 reusable design patterns
- Prompt Master Course - Systematic Prompt Engineering training
- AI Agent Development - Building autonomous AI agents
These cases all come from real products, showing how top companies approach system prompt design. Study them, then develop your own style.
📚 相关资源
❓ 常见问题
关于本章主题最常被搜索的问题，点击展开答案
Anthropic、OpenAI、Google 三家在 system prompt 风格上最大的差异是什么？
三家走三条路。Anthropic Claude 走 "安全优先"：RLHF + Constitutional AI + 用 NEVER / ALWAYS 关键词划红线 + CLAUDE.md 让用户定制 Agent 行为。OpenAI GPT 走 "工具优先"：TypeScript Namespace 定义工具、Function Calling、明确的 Allowed / Not Allowed 列表（如 banking transfers 禁止）。Google Gemini 走 "工作流优先"：Understand → Plan → Implement → Verify Tests → Verify Standards 五步流程，强调 "NEVER assume"，先看现有代码再动手。
xAI Grok 的 Persona 系统是怎么设计极端个性化的？
Grok 提供多套 persona 切换：Companion、Unhinged Comedian、Loyal Friend、Homework Helper、Not a Doctor / Not a Therapist。每个 persona 独立的语言风格规则 —— 比如 Loyal Friend 要求：lowercase 写作（除强调外）、用缩写如 rn / ur / bc、逗号少、不假设朋友性别、match 用户的脏话级别（用户骂才骂）。这种细粒度 style rules 是 "用 system prompt 创造一致人格" 的极端样本，也是社交类 AI 应用的参考模板。
Perplexity 怎么用 system prompt 保证搜索结果的时效性？
Perplexity 在 prompt 里明确写：每次用户追问 "that might also require fresh details" 都必须重新调 search_web，不能假设之前的搜索结果还能用；遇到任何不确定都重搜一次。原文是 "Always verify with a new search to ensure accuracy if there's any uncertainty." 这条规则把 "缓存即过时" 的判断权从工程层下放到 prompt 层 —— 模型每一轮都自己决定要不要重搜，比写硬规则更灵活。这是搜索型 AI 应用的核心设计模式。
OpenAI 的 TypeScript Namespace 和 Anthropic 的 XML 工具定义哪种更好？
没有绝对更好，看场景。OpenAI 的 `namespace file_search { type msearch = (_: { queries?: string[]; }) => any; }` 用 TS 类型给参数划约束，对前端工程师来说极其熟悉，IDE 都能补全。Anthropic 的 XML 格式（`<tool_definition>...</tool_definition>`）更适合需要嵌套结构和说明文本的场景。Grok 也用 XML Function Call。规则：JS / TS 团队优先 namespace，Python / 多语言混合环境优先 XML —— 但模型对两种格式都训练过，差距远小于 description 写得清不清楚的影响。
想给自己的 AI 应用挑一个参考 system prompt，该照哪家学？
按场景对号入座：代码助手抄 Claude Code + Gemini CLI（一个学输出极简，一个学验证流程）；搜索类应用抄 Perplexity（学时效性策略）；聊天机器人抄 Grok Persona（学个性化风格规则）；浏览器自动化 / Agent Mode 抄 OpenAI（学工具定义和消息通道）；高安全性企业应用抄 Claude（学 NEVER / ALWAYS + 安全边界）。复用模式时按 Identity → Tools → Constraints → Output 四段拼装，每段都有现成模板可直接改。