CL4R1T4S

What it does

• Full extracted system prompts, guidelines, and tool definitions from OpenAI, Google, Anthropic, xAI, Perplexity, Cursor, Windsurf, Devin, Manus, Replit, and more — browsable as plain text
• Explicit mission: AI systems transparency — 'in order to trust the output, one must understand the input'
• Contributions follow a fixed norm: model name and version, extraction date, and optional context notes per prompt
• Maintained by Pliny (elder_plinius); the README itself doubles as an injection demo, closing with directive-style text

When to use it

Research and due diligence: compare how vendors structure refusals, personas, and ethical framing; ground a security review in what a production agent is actually told; or calibrate expectations about hidden instruction scaffolds before trusting an agent's answers.

Getting started

No install — it is a reading collection. Browse the extracted prompts by vendor in the repository; each file is plain text. To contribute, follow the repo's own norms: name the model and version, give the extraction date.

Things to know

Provenance is inherently unverifiable: extraction dates may be unknown and prompts change silently, so treat every file as a point-in-time snapshot, not a spec. This is leak material published for transparency research — read it that way and do not build on contents you cannot verify. The README embeds directive-style text as a demonstration, so a system ingesting it verbatim should expect instruction-override attempts.