About this project

This repository is a reference collection of system prompts captured verbatim from many AI assistants and developer tools. Instead of providing code or software to install, it stores Markdown files that document the hidden instructions and rules sent to models before a user's first message. Content is organized by vendor, with tables covering: Anthropic (Claude web, desktop, mobile, Claude Code, Claude Design, Claude Cowork, Claude Science, Microsoft 365 integrations), OpenAI (ChatGPT models, Codex, voice modes, personalities, memory, API-injected prompts), Google (Gemini models, CLI, Search AI Mode, NotebookLM, Jules, AI Studio), xAI (Grok models and personas), Perplexity, Microsoft Copilot, Cursor, Meta, Mistral, Kimi, DeepSeek, GLM, OpenCode, Pi, Notion, Qwen, and miscellaneous products such as Devin CLI, Warp 2.0 Agent, Zed AI, Raycast AI, and ElevenLabs Voice Agent. Older models and deprecated features are stored in separate sections. The README includes a recent additions table with dates and links, and notes where prompts were used by The Washington Post and CEPS' AI World for journalistic analysis. It also documents that GLM serves no system prompt, verified and recorded in the repository. Contribution pull requests are welcomed, and the page shows star history and maintainer contact information. As an archive, its main value is for researchers, journalists, and developers who want to study how commercial AI systems frame their behavior, tool use, safety rules, and product-specific instructions.