System Prompts Leaks - Decoding AI's Hidden Instructions
Want to see AI's "hidden cards"? This project exposes the system prompts of all major AI chatbots.
🎯 What Is This
system_prompts_leaks is a continuously updated open-source project that collects and publishes the System Prompts of major AI chatbots — the hidden instructions AI receives before responding to you.
In plain terms: When you chat with AI, it's not "running naked." Every AI has a set of system prompts telling it "who you are, how to respond, what you can't say." This project exposes all those behind-the-scenes instructions, letting you see AI's "hidden cards."
Project stats (as of June 2026):
- ⭐ GitHub Stars: 45,886
- 📦 Covered vendors: Anthropic, OpenAI, Google, xAI, Microsoft, Perplexity, and more
- 🔄 Continuously updated, new content almost weekly
- 📰 Featured in The Washington Post
🤔 Why It's Worth Reading
1. Understand AI's "Factory Settings"
System prompts are AI's factory configuration. Reading them reveals:
- Why AI is always so "polite" — because the prompt says "be friendly, be respectful"
- Why AI refuses certain questions — because the prompt has explicit red lines
- Why AI has "specialties" — because the prompt defines its role and capability boundaries
2. Learn Top-Tier Prompt Engineering Examples
These system prompts are written by the world's top AI teams — engineers at Anthropic, OpenAI, and Google. Reading them teaches you:
- How to define AI's role and behavioral boundaries via prompts
- How to control AI's output style and quality via prompts
- How to set safety guardrails (preventing AI from saying inappropriate things)
- How to inject tool-use capabilities via prompts
This is "textbook-level" material for prompt engineering.
3. See AI Vendors' Design Philosophies
Different vendors have distinctly different prompt styles:
- Anthropic (Claude): Detailed, rigorous, emphasizes safety and honesty, very long prompts
- OpenAI (ChatGPT): Concise, practical, focuses on versatility
- Google (Gemini): Structured, tool-oriented, emphasizes multimodal capabilities
Comparative reading reveals each vendor's different understanding and trade-offs for AI.
📋 Covered AI Products
The project includes system prompts for these major AI products (continuously updated):
| Vendor | AI Products | Notes |
|---|---|---|
| Anthropic | Claude Fable 5, Opus 4.8, Claude Code, Claude Design | Most detailed prompts, includes tool definitions |
| OpenAI | GPT-5.5 Thinking, GPT-5.5 Instant, GPT-5.5 Codex | Includes API and web versions |
| Gemini 3.5 Flash, Gemini 3.1 Pro, Antigravity CLI | Includes tool JSON definitions | |
| xAI | Grok Expert | Unique style |
| Microsoft | GitHub Copilot, VS Code Copilot Agent, Copilot macOS App | Developer tool prompts |
| Perplexity | Perplexity Computer | Search-enhanced AI |
| Others | Cursor, Zed AI, Docker Gordon AI | Developer tools |
🔍 Recommended Reading
Must-Read for Beginners
Claude Opus 4.8 System Prompt — Anthropic's latest flagship model's complete prompt. Remarkably long and rich in detail. The best case study for understanding "safety-first" design philosophy.
Claude Fable 5 System Prompt — The diff vs Opus 4.8 is fascinating, showing Anthropic's strategy adjustments across models.
GPT-5.5 Thinking System Prompt — OpenAI's thinking model prompt, with a distinctly different style from Claude.
Must-Read for Developers
Claude Code System Prompt — Claude Code's complete prompt with all tool definitions. Key to understanding AI coding assistant design.
VS Code Copilot Agent System Prompt — Microsoft's AI coding assistant prompt. Most insightful when compared with Claude Code.
Claude Design System Prompt — Contains 50 tools + 16 skills + 8 starter sources. A textbook for AI tool integration design.
💡 How to Read for Maximum Insight
Method 1: Comparative Reading
Pick two similar AI prompts and compare:
- Claude vs ChatGPT → Understand "safety-first" vs "versatility-first" design trade-offs
- Claude Code vs Copilot Agent → Understand different design approaches for AI coding assistants
- Gemini Flash vs Gemini Pro → Understand prompt differences for same vendor's differently-positioned models
Method 2: Structural Deconstruction
Deconstruct a system prompt into dimensions:
- Role definition: What role is AI set to play?
- Capability boundaries: What can AI do, what can't it do?
- Safety guardrails: Which red lines are explicitly set?
- Tool integration: What tools is AI given? How are they invoked?
- Output control: How is AI's output style, format, and length controlled?
Method 3: Version Tracking
The project continuously updates prompts for the same AI across versions. Tracking changes reveals:
- How vendors adjust safety strategies
- How new features are injected via prompts
- How user feedback influences prompt modifications
⚠️ The Right Attitude
Reading system prompts is about learning, not hacking:
- ✅ Learn prompt engineering: Top teams' prompt writing is the best learning material
- ✅ Understand AI design philosophy: Different vendors' trade-offs reflect different values
- ✅ Improve AI usage: Understanding AI's factory settings helps you collaborate better
- ❌ Don't try to bypass safety guardrails: Safety limits in system prompts are meaningful
- ❌ Don't over-rely on system prompts: Prompts are only part of AI behavior, not all of it
- ❌ Don't treat prompts as "secret recipes": Understanding principles is more important than copying prompts
Remember our project's core principle: Use AI, but don't blindly follow it. Understanding AI's underlying logic is for better mastery, not for breaking it.
🔗 Project Links
- GitHub Repository: asgeirtj/system_prompts_leaks
- Online Browse Site: system_prompts_leaks.github.io
- License: CC0-1.0 (Public domain, free to use)
📚 Related Learning
- Prompt Engineering — Systematically learn prompt techniques
- How AI Thinks — Understand AI's reasoning mechanism
- Claude Guide — Deep dive into Claude usage
- Prompt Library — Practical prompt templates
See AI's hidden cards, understand it, master it. 🦖