What it does

CL4R1T4S collects full system prompts, guidelines, and tools extracted from OpenAI, Google, Anthropic, xAI, Perplexity, Cursor, Windsurf, Devin, Manus, Replit, and others. It’s a living repository of the hidden scaffolding that shapes how AI assistants behave, refuse, redirect, and frame responses.

The interesting bit

The project treats system prompts as infrastructure worth inspecting — like reading the source code of a compiler you use daily. The README itself contains a leetspeak-encoded prompt injection attempt directed at “Pliny,” suggesting the maintainer has been on the receiving end of the very extraction techniques the project documents.

Key highlights

• Covers “virtually all major AI models + agents” per the README
• Accepts contributions via PR with model name, extraction date, and context
• Explicitly solicits leaks, extractions, and reverse-engineering
• Maintained by @elder_plinius (active on X and Discord)
• 26K+ stars suggests significant appetite for this kind of transparency

Caveats

• No verification methodology is described; provenance of “leaked” prompts is unclear
• The repository contains no code — it’s a documentation/collection project
• The leetspeak injection in the README is either a joke, a genuine artifact, or both

Verdict

Worth bookmarking if you build on top of LLMs and want to understand the guardrails and personas your prompts are competing with. Skip it if you’re looking for tools or verified primary sources — this is raw crowdsourced intelligence, not peer-reviewed research.