GitHub repo system_prompts_leaks collects system prompts from major models

The repo lists many original system prompts grouped by vendor. The Anthropic section covers models such as Claude Haiku 5.5, Sonnet 5.5, Opus 5.5, and Fable 5.1, plus multiple versions of Claude Code prompts for terminal, desktop client, cloud, and headless modes, along with artifact types like Slides, Dashboards, Design, and Docs, plus injection reminders and sub-agent prompts. The OpenAI section includes ChatGPT versions, the Codex series, and API injection prompts; the Google section includes Gemini 3.8 Flash, Gemini 3.1 Pro, and Antigravity CLI; there are also products like Grok, Muse, Perplexity, Cursor, Kimi, and DeepSeek. The repo also lists dated recent additions, such as Claude artifact-type prompts added on October 9 and Claude Haiku 5.5 with corresponding Claude Code prompts added on October 7. It says its content is organized and proofread, welcomes PRs for additional findings, and claims citations from outlets and institutions including The Washington Post and CEPS's AI World, with Simon Willison calling it the most complete collection of leaked system prompts he knows of. For developers doing prompt engineering and agent behavior tuning, and teams doing competitive product analysis, this repo turns system-level instructions that previously required repeated conversational reverse-engineering into searchable text, making it easy to compare how the same model behaves across different entry points and to locate specific sections on tool calling, persona settings, and safety policies. However, the content is not officially confirmed by vendors, and version numbers may not fully match live behavior, so for production applications that depend on system prompt stability, self-testing should remain the standard.