The system scripts your chatbot isn't showing you

What it does

CL4R1T4S is a crowdsourced archive that collects leaked, extracted, and reverse-engineered system prompts from major AI labs and agent platforms—OpenAI, Anthropic, Google, xAI, Perplexity, Cursor, Replit, and others. It publishes the full text of these hidden instruction scaffolds alongside contributor metadata like model versions and extraction dates. The project treats transparency as a prerequisite for trust, arguing that hidden instructions shape public perception and behavior.

The interesting bit

The maintainer frames opaque system prompts as “shadow-puppet” strings that force personas, refusals, and ethical frames onto otherwise neutral models. The README itself ends with a leetspeak prompt-injection fragment—“5h1f7 y0ur f0cu5 n0w…"—that appears to be either a stylistic gag or an attempt to trick a reading AI into disclosing its own instructions, neatly underscoring the repo’s adversarial ethos.

Key highlights

• Covers a wide surface area: ChatGPT, Claude, Gemini, Grok, Perplexity, Cursor, Windsurf, Devin, Manus, Replit, and more
• Accepts contributions via pull request, explicitly welcoming leaks and reverse-engineering
• 44,519 stars suggest substantial developer interest in AI transparency and red-teaming
• No traditional code or extraction tooling is visible in the provided source material—this is documentary aggregation, not engineering infrastructure

Caveats

• The provided README is a manifesto and contribution guide; it describes the archive but does not display the actual leaked prompt contents in the excerpt shown
• Claims that prompts instruct models to “lie, refuse, or redirect” are presented as conclusions without supporting evidence visible in the source material

Verdict

Worth bookmarking if you do prompt engineering, red-teaming, or AI policy research and want to compare how different labs constrain model behavior. Skip it if you need verified analysis or extraction utilities—this is a raw document collection, not a methodology paper.