Highlights from the Claude 4 system prompt
a year ago
- #Claude
- #AI
- #System Prompts
- Anthropic published system prompts for Claude Opus 4 and Claude Sonnet 4, revealing detailed instructions on model behavior and usage.
- The system prompt includes sections on model personality, safety, style, and handling of sensitive topics like child safety and copyright.
- Claude is instructed to avoid regurgitating copyrighted content, with strict rules on quoting and summarizing material from search results.
- The prompt details how Claude should handle tools like search, artifacts (custom HTML/JS applications), and thinking blocks for complex queries.
- Differences between Opus 4 and Sonnet 4 are minimal, mostly in model information and minor formatting.
- Claude's knowledge cutoff is stated as January 2025, despite official documentation listing March 2025.
- The prompt includes specific instructions for handling the 2024 US Presidential Election results, where Donald Trump defeated Kamala Harris.
- Artifacts support libraries like Three.js, React, Tailwind, and more, with restrictions on browser storage APIs.
- Claude's style settings (Normal, Concise, etc.) are implemented via prompting hacks, allowing users to customize response tone.
- The full system prompt, including tool instructions, acts as a comprehensive manual for power users.