Every AI coding assistant you trust runs on a hidden instruction sheet you never see — and those sheets turn out to be where much of the product's "magic" actually lives. This repository collects them: the system prompts, internal tool schemas, and model configs extracted from 25+ proprietary assistants, from Cursor and v0 to Devin, Replit, Windsurf, Manus, and Claude Code.
Reading them is an education in prompt engineering at production scale. You see how Cursor frames file edits, how v0 constrains its component output, and how agentic tools describe their own tools back to the model — patterns refined by well-funded teams and otherwise invisible behind a closed API.
What Sets It Apart
- Breadth over one deep dive — 25+ tools side by side, so you can compare how different teams solve the same problem (tool-calling, refusing unsafe edits) instead of generalizing from a single example.
- Raw artifacts, not commentary — the prompts are the primary source. You read the exact wording teams ship, including the hedges, hard rules ("do not apologize"), and formatting tricks they lean on.
- A living snapshot — with 141k+ stars and dozens of contributors, new leaks land as tools ship and prompts change, tracking the moving state of the art.
Who It's For
A great fit if you build agents or LLM products and want concrete, battle-tested prompt patterns instead of blog-post theory, or if you research how commercial AI tools are actually engineered. Look elsewhere if you need official, stable docs: these are leaked and reverse-engineered artifacts, can be incomplete or outdated, and reproduce proprietary text — treat them as study material, not a license to copy.
One Honest Caveat
The repository doubles as an argument for prompt security — its solo maintainer also runs a service helping startups stop exactly this kind of extraction. That tension is the point: if one person can compile 25+ of these, your own system prompt is probably less secret than you think.
