// the take
A community-maintained collection of verbatim system prompts extracted from major AI products — Claude, ChatGPT, Gemini, Grok, Cursor, Copilot, and dozens more. It's a reference for anyone who wants to understand how these products actually constrain and shape their models at runtime, rather than taking vendor documentation at face value. Updated within days of new model releases.
Coverage is genuinely broad and current — prompts for Fable 5 and GPT-5.6 appeared within weeks of launch, which takes real effort to maintain. The distinction between 'captured' prompts and officially published ones is useful; you can see where vendors quietly change behavior without announcement. Organizing by product and model version rather than just dumping files makes it navigable at this scale. The Claude Code section goes deeper than just the main system prompt — subagent prompts, injected reminders, MCP server instructions, and individual slash commands are all here.
There's no verification methodology documented — you're taking the contributor's word that any given prompt is authentic and complete, which matters a lot when prompts are thousands of tokens and easy to partially capture or misattribute. The repo has no tooling to diff prompt versions over time, so tracking behavioral changes across model updates means manually reading markdown files. Extraction methods aren't explained anywhere, which makes it hard to assess what was captured correctly versus what might be truncated or reconstructed from memory. As a plain file dump with no search, finding a specific behavior or constraint across all prompts requires grep.
We read the README, the directory tree and a key source file in July 2026. Nobody pays to be listed. How we judge →