Skip to content

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):

VendorAI ProductsNotes
AnthropicClaude Fable 5, Opus 4.8, Claude Code, Claude DesignMost detailed prompts, includes tool definitions
OpenAIGPT-5.5 Thinking, GPT-5.5 Instant, GPT-5.5 CodexIncludes API and web versions
GoogleGemini 3.5 Flash, Gemini 3.1 Pro, Antigravity CLIIncludes tool JSON definitions
xAIGrok ExpertUnique style
MicrosoftGitHub Copilot, VS Code Copilot Agent, Copilot macOS AppDeveloper tool prompts
PerplexityPerplexity ComputerSearch-enhanced AI
OthersCursor, Zed AI, Docker Gordon AIDeveloper tools

Must-Read for Beginners

  1. 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.

  2. Claude Fable 5 System Prompt — The diff vs Opus 4.8 is fascinating, showing Anthropic's strategy adjustments across models.

  3. GPT-5.5 Thinking System Prompt — OpenAI's thinking model prompt, with a distinctly different style from Claude.

Must-Read for Developers

  1. Claude Code System Prompt — Claude Code's complete prompt with all tool definitions. Key to understanding AI coding assistant design.

  2. VS Code Copilot Agent System Prompt — Microsoft's AI coding assistant prompt. Most insightful when compared with Claude Code.

  3. 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.


See AI's hidden cards, understand it, master it. 🦖

MIT Licensed