A GitHub repository has publicly mirrored what it identifies as the system prompt for Claude Fable 5, giving outside observers a rare look at the instructions and product framing behind Anthropic's latest broadly available model tier.
The file, posted in the public repository CL4R1T4S, is labeled as a Claude Fable 5 system prompt and spans more than 1,500 lines. The document is hosted as a third-party artifact rather than an official Anthropic publication, but it appears to reproduce detailed behavioral guidance, product descriptions and safety rules associated with the model.
According to the repository listing, the file was updated on June 15, 2026. GitHub also shows the document in a public project with substantial attention, including tens of thousands of stars and thousands of forks, underscoring the level of interest in model internals and prompt engineering materials.
The mirrored document describes Claude Fable 5 as the first model in Anthropic's Claude 5 family and says it sits within a new Mythos-class tier positioned above Claude Opus in capability. It also distinguishes between Claude Fable 5 and Claude Mythos 5, saying the two share the same underlying model but differ in access and safety controls.
The prompt text includes instructions on how the assistant should respond to users, including product information, safety boundaries, and advice on prompting. It also refers to multiple access channels such as web, mobile, desktop, API, Claude Code, and other tools. The document further says Claude should avoid certain content, including harmful instructions, malicious code, and material involving minors.
One notable instruction in the file says Claude should never use a specific type of voice note block, even if such blocks appear in conversation history. The document also lays out policies on when the model should search official documentation, how it should describe product capabilities, and how it should handle refusals.
System prompts are often closely held because they can reveal how a model is instructed to behave, what safety policies it follows, and how a company positions a product. Publicly accessible copies can become reference points for researchers, developers and hobbyists studying model behavior, though they may not reflect official, current documentation in full.
In this case, the mirrored file appears to be a third-party version of internal or semi-internal guidance rather than an authenticated release from Anthropic. That means readers should be cautious about treating it as authoritative without corroboration from the company or its documentation.
Still, the posting offers a snapshot of how one of Anthropic's latest models is described in the wild, including the way it is framed against other Claude offerings and how its safety restrictions are articulated.
The file's appearance also highlights an ongoing pattern in the AI sector: prompts, policies and configuration files sometimes leak, are republished, or circulate through public repositories, creating an informal record of model governance that can be difficult for companies to control once it spreads online.