CL4R1T4S is an AGPL-3.0 repository of full extracted system prompts, guidelines and tools from the major AI models and agents — ChatGPT, Claude, Gemini, Grok, Perplexity, Cursor, Lovable, Replit and more — for anyone who wants to know what instructions sit behind the AI systems they use.

What it is

CL4R1T4S sits in the AI & Machine Learning / AI Development Platforms category, but it is a reference corpus rather than a runnable library: the registry lists no programming language for it. The README frames the whole project around one claim, "In order to trust the output, one must understand the input," and argues that AI labs shape model behaviour through large, unseen prompt scaffolds. Those scaffolds define what a model cannot say, which personas and functions it must follow, when it is told to lie, refuse or redirect, and which ethical or political frames are baked in by default. Stated coverage is "virtually all major AI models + agents."

The concrete problem is opacity at the point of use. Interacting with a model whose system prompt is unknown means talking to what the README calls a "shadow-puppet," so the repository gathers extracted prompts, guidelines and tools into one public place instead of leaving them scattered or undisclosed. The numbers behind it are substantial: about 49,933 stars, 10,263 forks and 127 open issues.

Key capabilities

• Publishes full extracted system prompts and guidelines from OpenAI, Google, Anthropic, xAI, Perplexity, Cursor, Windsurf, Devin, Manus and Replit, spanning surfaces such as ChatGPT, Claude, Gemini, Grok, Lovable and more.
• Covers tools shipped alongside those prompts, not the prompts alone.
• Accepts outside additions through a pull request format that asks for model name and version, date of extraction where known, and optional context or notes.
• Routes leaks, extractions and reverse-engineering results to @elder_plinius on X or Discord.
• Carries repository topics including prompt-engineering, agents, leak, leaked, hacking, openai, chatgpt and gemini for discoverability.
• Distributes under AGPL-3.0, so redistribution and derivative works carry copyleft obligations.

Who uses it and how

• Prompt engineers read the extracted scaffolds before trusting a model's refusals or persona in a shipped product.
• Agent and tooling developers study how vendors instruct models to redirect, refuse or stay in character, then design around those constraints.
• Researchers and writers work from the corpus when examining the ethical and political framing the README says is baked into defaults.
• Leakers, extractors and reverse-engineers contribute directly, following the pull request checklist for model name, extraction date and notes.

Getting started

The README documents no install, package, image or compose file, so there is nothing to deploy — the material is read straight from the repository. Additions arrive as pull requests, or through @elder_plinius on X or Discord.

How it compares

No comparable tool is named in the supplied facts, and no list of paid products it replaces is given. Within this registry it stands alone as a prompt-transparency archive rather than a product with a commercial counterpart.

When to use it — and when not to

Pick it when the goal is reading primary prompt text; skip it if you need a maintained library, an API or vendor-supported documentation, because the repository ships no code, package or deployment target. A self-hoster has nothing to operate — no database, storage or SMTP — but also no versioned releases to depend on, and the README contains an embedded leetspeak instruction block, which signals informal maintenance. Treat every prompt as a point-in-time extraction rather than a current statement from the lab that wrote it.

"In order to trust the output, one must understand the input."

What AIs can’t say

What personas and functions they’re forced to follow

How they’re told to lie, refuse, or redirect

And what ethical/political frames are baked in by default

If you're interacting with an AI without knowing its system prompt,
you’re not talking to a neutral intelligence — you’re talking to a shadow-puppet.

✅ Model name/version

🗓 Date of extraction (if known)

🧾 Context / notes (optional but helpful)