Pliny & CL4R1T4S: The GitHub Repo of Leaked AI System Prompts (51K Stars) + How to Find Free AI Models

Every time you open ChatGPT, Claude, or Gemini, there are hidden instructions already loaded before you type your first message. These are called system prompts - and most AI companies keep them private. A GitHub account called elder-plinius, operating under the online handle Pliny the Liberator, has been collecting and publishing them in a repo called CL4R1T4S.

The repo has over 51,300 stars as of October 2026. This guide explains what the repo actually contains, what it does not do, and how to find genuinely free AI models safely.

What Is CL4R1T4S?

CL4R1T4S is a GitHub repository maintained by elder-plinius (handle @elder_plinius on X). The repo was created on March 4, 2025, and its stated goal is "AI systems transparency." It collects extracted system prompts for a wide range of AI products including ChatGPT, Claude, Gemini, Grok, Perplexity, Cursor, Lovable, Replit, and others.

The license is GNU AGPL v3.0, which is a strong copyleft open-source license. As of October 20, 2026, the repo has approximately 51,300 stars and 10,500 forks. Contributions are accepted via pull request, with each submission expected to include the model name, version, and date of extraction. Elder-plinius is an anonymous online handle - nothing is known publicly about who the person is.

What Is a System Prompt?

When you send a message to an AI assistant, the model does not receive just your text. Before your words arrive, a block of instructions written by the company is already loaded. This block is the system prompt. It tells the model who it is, what it can discuss, what it must refuse, what tone to use, what tools it has access to, and more.

For example, a system prompt for a customer service AI might include instructions like: "You are a support agent for Company X. Only discuss topics related to our product. Do not discuss competitor products. Always be polite." A system prompt for a coding assistant might define what languages to prioritize, how to format code, and what safety policies to follow.

Companies treat these prompts as proprietary. They are not shown to users. CL4R1T4S collects them through extraction - using prompts and interactions that reveal what instructions the model is operating under - and publishes them publicly.

What CL4R1T4S Does NOT Do

This is the most important section to read clearly before you draw conclusions from the reel or from social media coverage.

• It does not give you free access to ChatGPT, Claude, or any other paid model. Reading a system prompt does not unlock the model behind it. The model itself runs on the company's servers and requires an API key or subscription.
• It does not contain a copy of any model. A system prompt is a text file - a few hundred to a few thousand words of instructions. It is not the model. The model is a neural network with billions of parameters, trained on massive compute. The two are completely separate things.
• It does not remove or bypass any AI's restrictions. Knowing what instructions an AI has been given is different from changing them. You cannot use this repo to make a paid AI behave differently than it is designed to.
• It is not about jailbreaking. The repo is framed around transparency - publishing what instructions exist - not around circumventing safety measures.

What reading system prompts does give you is understanding. You can see how companies configure their AI products - what they prioritize, what they restrict, and how they define the assistant's behavior. For people building with AI or studying how these systems work, that is genuinely useful information.

How to Find Free AI Models Safely

The reel covers three methods for finding free AI models. Here is what each one actually means in practice, with the caveats worth knowing.

Step 1 - Use the OpenRouter Free Filter

OpenRouter is a platform that routes requests to many different AI models through a single API. It has a free tier that includes models with $0 input and $0 output pricing. You can filter for these directly:

openrouter.ai/models?max_price=0

Important caveats: free models on OpenRouter have rate limits - you can only send a certain number of requests per minute or per day. The list of available free models changes as providers add or remove them. Do not rely on a specific number - check the filter yourself when you need it. Rate limit details are documented here: openrouter.ai/docs/api_reference/limits

Step 2 - Check X/Twitter After New Model Launches

When a major AI lab releases a new model, open-source developers often release their own fine-tunes, distillations, or alternative versions shortly after. These appear on X/Twitter, HuggingFace, and GitHub. Following AI researchers and developers on X is a practical way to hear about these quickly.

Safety check before you install anything from X:

• Check the repo's star count and how recent the activity is. A repo with no stars, no recent commits, and no issues is a red flag.
• Read the license. Make sure it permits the use you have in mind.
• Read install commands before running them. A curl | sh or pip install command runs code on your machine. Read it first.
• Check the issues tab. Real users leave real feedback. If something is broken or suspicious, someone has usually reported it.
• Do not paste API keys or secrets into unknown tools. A new tool asking for your OpenAI or Anthropic key should be treated carefully. Only use tools you trust or have audited.

Step 3 - Give Your AI Tool the Repo Link and Ask It to Install

If a repo has a clear README with installation steps, you can paste the GitHub URL into an AI coding assistant and ask it to read the README and follow the installation instructions. Tools like Claude Code, Codex CLI, or Cursor can do this effectively for well-documented repos.

The same safety rules apply here as in Step 2. An AI assistant will follow instructions in a README - including instructions to run code. Make sure you have read and understood what the install process does before you ask an AI to execute it.

What You Can Actually Learn from System Prompts

For anyone building AI-powered products or studying how these tools work, the CL4R1T4S repo is a practical resource. Here is what is worth paying attention to:

• Structure and length. How long are professional system prompts? How are they organized? The prompts for tools like Cursor and Lovable show how developer-focused AI products are configured for technical tasks.
• Tone and persona instructions. How do companies define their AI's personality, formality level, and response style? This matters when you are writing system prompts for your own applications.
• Safety and refusal patterns. What topics or request types does each company explicitly instruct their model to avoid? This varies significantly between providers and products.
• Tool and capability declarations. Some system prompts declare what tools the model has access to - web search, code execution, file reading. Seeing how these are described in the prompt helps when configuring similar setups.

The prompts in the repo are snapshots at specific points in time - each contribution includes the extraction date. Companies update their system prompts regularly, so older entries may not reflect the current behavior of a product. Use them as reference and learning material, not as definitive documentation.

Is the CL4R1T4S Repo Itself Trustworthy?

The repo contains text files - system prompts written in plain language. There is no executable code to run and no installation required to read the contents. The risk profile is therefore low compared to, say, a new AI tool asking for your API key.

The GNU AGPL v3.0 license is a recognized, standard open-source license. The repo has over 51,000 stars, an active fork count, and a long history of contributions going back to March 2025. These are reasonable signals of legitimacy for a text repository.

That said: individual prompt entries are contributed by the community. The repo does not independently verify that every submitted prompt is exactly what a company currently uses. Treat each entry as a community-gathered snapshot, not an official disclosure from the AI company in question.