Anthropic has published the system prompts for its Claude consumer applications, including the latest updates for the Fable 5.1 model. The company's transparency in sharing these prompts, including historic changes, provides valuable insights into how Claude is designed to operate. A notable addition in the Fable 5.1 prompts is the section on not reproducing song lyrics, poems, or book passages, reflecting a heightened awareness of copyright issues.
The updated prompts also address the generation of images of copyrighted material, such as specific artworks, album covers, or brand logos. Claude is instructed to not reproduce such images and instead offer original alternatives. This shift is particularly relevant given the absence of a specialist text-to-image model in Anthropic's offerings, unlike some competitors.
Furthermore, the handling of abusive conversations has undergone changes, with Claude now maintaining a steady, honest, and respectful tone without becoming submissive or apologizing unnecessarily. The goal is to provide helpful responses while upholding self-respect. Interestingly, while the system prompt no longer mentions the end_conversation tool, it is still part of Claude's functionality, used as a last resort in cases of persistently abusive behavior.
Claude's system prompts have always included sections on illegal substances, but Fable 5.1 introduces new guidance. The model is instructed to provide life-saving and life-preserving information, such as signs of overdose or dangerous interactions, while refraining from giving specific protocols for dosing or production. Notably, this section includes external URLs for harm-reduction information sources, a first for Claude's system prompts.
The reliable knowledge cutoff for Claude is now set to the end of June 2026, reflecting the model's training data. This information is directly provided to the model, influencing how it responds to queries. To track changes in these prompts, a Git timeline has been created, allowing for easy comparison of updates. This repository includes synthesized commit histories for each model family and version, facilitating the identification of new rules, behaviors, or policy changes.
Automation, powered by GPT-5.6 Luna, is used to create bullet-point summaries of changes, available in the repository's README and CHANGELOG.md file. This approach helps in navigating the complexities of prompt updates, highlighting key changes such as the refusal to reproduce protected works, the handling of drug-related queries, and the adjustment in responding to abusive users.
Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.