System Prompts Leaks

Best for: prompt engineers and researchers who want to read how frontier labs actually instruct their own models.

This repository collects leaked system prompts captured verbatim — the hidden instructions given to ChatGPT, Claude, Gemini, Grok and other assistants before a user's first message. Maintained by Asgeir Thorsen and updated regularly, it covers Anthropic, OpenAI, Google and xAI models along with Cursor, Copilot, VS Code and Perplexity. Released under CC0, effectively public domain.

What makes it valuable: it is primary source material. Reading how well-resourced labs handle refusals, tool use, formatting and tone teaches more about practical prompt design than most written guidance, because these prompts are shaped by enormous amounts of real-world feedback. Its credibility is reflected in citations by The Washington Post and CEPS, both of which built published work from the repository's files.

Where it falls short: these prompts are extracted rather than officially published, so accuracy and completeness cannot be guaranteed, and they go stale quickly as models are updated. Treat entries as a snapshot rather than current documentation.

Details: JavaScript tooling, CC0 1.0.