Anthropic publishes the system prompts behind Claude's web and mobile chat apps in a release notes page that drew heavy discussion on Hacker News this week, according to the page. The prompts carry contextual information like the current date and instructions on tone, safety practices and formatting, and explicitly do not apply to the Claude API, the documentation says.
The page lists prompts for Claude Opus 5, released July 24, Claude Fable 5, released June 9, and earlier models including Opus 4.8, Opus 4.7 and Sonnet 4.6, with Claude 4.6 and later models shown as fixed snapshots per model ID and earlier versions marked with bolded changes between releases, according to Anthropic.
The documentation notes that Fable 5 access was suspended from June 12 to July 1 under U.S. export controls before being restored, and that each model's prompt lists a knowledge cutoff, with Fable 5's ending in January 2026 and Opus 5's ending in May 2026, Anthropic says.
Each entry also covers refusal handling, tool access and product context that shapes how the assistant responds inside the chat apps specifically, as distinct from how a developer's own system prompt shapes API behavior, according to the page.
For developers, the practical takeaway is the boundary the page draws for itself: none of this governs API calls. Anyone building on Claude through the API is working from their own system prompt, not the one documented here.