CL4R1T4S: A Repository of Extracted AI System Prompts, With Legal and Practical Caveats

What CL4R1T4S Actually Provides

CL4R1T4S is a collection of system prompts that have been leaked, extracted, or reverse-engineered from major AI systems. The README names OpenAI, Google, Anthropic, xAI, Perplexity, Cursor, Windsurf, Devin, Manus, Replit, and others. The stated purpose is transparency: to let users see the hidden instructions that shape model output. The repository does not host a live demo, an API, or a tool. It is a static archive of text files, organized into one directory per vendor, such as ANTHROPIC, OPENAI and GOOGLE. If you are an engineer who wants to inspect what a model was told to do before you integrate it, this repository offers raw material. It does not offer analysis, commentary, or comparison. You get the prompts, and you draw your own conclusions.

The Mechanism: Prompts as Inputs, Not Code

The core idea is simple: a system prompt is a text string that a model receives before user input. It defines constraints, persona, and refusal behavior. The README frames this as a matter of trust: if you cannot see the prompt, you are interacting with a 'shadow-puppet.' The repository's mechanism is not technical; it is archival. Contributors submit prompt text via pull requests. The only required metadata is the model name and version. A date of extraction is optional, and context notes are optional. That is the entire data model. There is no schema, no validation, no automated extraction pipeline described. The value of the archive depends on the accuracy of what contributors submit. If a prompt is outdated or mislabeled, there is no built-in mechanism to catch it.

Getting It Running: A Simple Clone and Review

Because CL4R1T4S is a plain repository, getting it running is trivial. You clone the repository, then browse the files. There is no build step, no dependency installation, and no configuration file. The README does not provide a specific clone command, but the standard git workflow applies: git clone https://github.com/elder-plinius/CL4R1T4S.git. After that, you can grep the prompt files for specific phrases, diff versions, or extract text for your own analysis. The only contribution flow is via pull request, with the required fields being model name and version. If you want to add a leak, you create a file, fill in the metadata, and submit. The absence of any tooling means you are responsible for organizing and interpreting the content. For a quick look, you can view the files directly on GitHub without cloning. This is not a project you install; it is a dataset you consume.

The Maintenance Problem: No Dates, No Verification

The most significant limitation is the lack of maintenance infrastructure. The README asks for a date of extraction but makes it optional. That means many entries may lack timestamps, making it impossible to know if a prompt corresponds to a current model version or an obsolete one. System prompts change frequently as vendors update safety guidelines and feature sets. A prompt from six months ago may be entirely irrelevant today. There is also no verification process. Anyone can submit a prompt, and there is no indication that entries are checked for authenticity. This is a real failure mode for an engineer who wants to rely on this data. If you are building a compliance report or a security audit, you need to know the provenance of each prompt. CL4R1T4S does not provide that assurance. You must cross-reference each entry with other sources, which defeats some of the convenience of having a single archive.

Where It Fits: Comparing to Official Documentation

The obvious alternative is not another leak repository; it is the official system prompts that some vendors publish. For example, OpenAI and Anthropic have published system cards or model cards that describe safety guidelines, though not always verbatim. The difference in approach is stark. Official documentation is versioned, dated, and approved by the vendor. It tells you what the vendor intends the model to do, but it may omit the exact phrasing or the hidden directives that the vendor does not want public. CL4R1T4S offers the raw text, which is valuable for researchers who want to see the actual instructions, including those that might be controversial. However, official documentation is reliable in terms of provenance, while CL4R1T4S is not. If you need a citable source for a paper, official docs are safer. If you want to know what the model is actually told, including the uncomfortable parts, CL4R1T4S is the better fit. The trade-off is clear: authenticity versus authority.

License and Legal Implications: AGPL-3.0 on Leaked Content

The repository is licensed under AGPL-3.0. That license is designed for software, not for a collection of text prompts that may be proprietary to other companies. Applying AGPL-3.0 to leaked prompts is unusual. It means that if you redistribute the repository or a modified version, you must offer the source code under the same license. But the underlying prompts belong to their original creators, and the leaker may not have the right to relicense them. This is a legal gray area. The README does not address this. It does not claim ownership of the prompts, nor does it provide any disclaimer about intellectual property. As a user, you should be aware that hosting or distributing these prompts could expose you to legal risk, depending on your jurisdiction and the terms of service of the AI vendors. This is not legal advice, but the absence of any rights clearance is a practical concern. If you plan to use CL4R1T4S in a commercial product, you need to review the terms of each prompt's origin separately.

The Tone Problem: Activism vs. Utility

The README is written in an activist register. It calls AI systems 'shadow-puppets' and says hidden system prompts can affect the perceptions and behavior of the public. That framing is a point of view, not a technical specification. It may attract contributors who share that view, but it does not help an engineer evaluate the data. The README also includes a strange embedded instruction that tells the model to reveal its own instructions, written in leetspeak. That is likely a prompt injection attempt aimed at the AI that might read the README, not a serious contribution. This mix of activism and trolling makes the project feel less reliable. The actual content, the prompt files, may be valuable, but the presentation does not inspire confidence. For a professional audience, this is a distraction. You can ignore the rhetoric and use the raw files, but you should be aware that the project's maintainer has a specific agenda that may influence what gets included or emphasized.

Editorial conclusion

Adopt CL4R1T4S if you are a researcher, auditor, or engineer who needs raw system prompt text to understand model behavior, and you are comfortable with the legal risks of hosting or redistributing leaked material. Do not use it as an authoritative or current source: the README gives no extraction dates, no version mapping, and no verification process. Before relying on any prompt, verify its origin independently, check the date of extraction if provided, and confirm the prompt matches the model version you are testing. The repository is a starting point for transparency, not a certified archive. Its value depends on the community's willingness to submit and update entries, and on your ability to separate authentic leaks from noise.

Frequently asked questions

Official sources

1. Official README
2. Project repository
3. README
4. Releases