Peeking at the cheat sheets inside ChatGPT and Claude

What it does This repository collects the hidden system prompts extracted from major commercial AI assistants—Anthropic’s Claude, OpenAI’s GPT and Codex, Google’s Gemini, xAI’s Grok, and tools like Cursor, Copilot, and Perplexity. It organizes these instruction strings by vendor, model version, and variant, including API flavors, personality modes, and tool-specific configurations. The collection is maintained with recent updates through May 2026 and has been cited by The Washington Post.

The interesting bit System prompts are the model’s unseen employee handbook: they dictate tone, enforce refusals, inject tool schemas, and slip in hidden context like dates or environment details. Seeing them raw and side-by-side turns abstract AI behavior into concrete, comparable policy documents, offering a rare look at the scaffolding behind the conversational facade.

Key highlights

• Covers Anthropic, OpenAI, Google, xAI, Perplexity, and miscellaneous tools (Cursor, VS Code Copilot, Docker Gordon, Zed, etc.)
• Tracks versioned releases, including API variants, personality modes (Friendly, Pragmatic, Cynical), and tool-specific prompts
• Includes both current and archived prompts, showing how vendor instructions evolve across model generations
• Cited in The Washington Post as a resource for understanding hidden AI rules
• Maintained with recent updates through May 2026

Caveats

• The README does not explain how prompts are extracted, so provenance and completeness for each leak are unclear
• Some official prompts are explicitly frozen at their release date and not updated, while others are refreshed regularly
• No analysis or diff tooling is provided; it is purely a raw prompt collection

Verdict Worth bookmarking if you build prompts, study model behavior, or simply want to know why ChatGPT sounds so chipper. Skip it if you are looking for executable code or a framework for prompt injection defense.