Claude 4 System Prompts Reveal Enhanced Model Safety and Personality Guidelines

Anthropic has released the system prompts for their latest models, Claude Opus 4 and Claude Sonnet 4, providing a detailed look into how these AI systems are guided to behave. These prompts act as an unofficial manual for users, highlighting the improvements in model safety, personality, and interaction style.

What Changed Technically

The new system prompts for Claude 4 focus heavily on enhancing model behavior and ensuring safer interactions. Key changes include:

• Enhanced Safety Measures: Explicit guidelines to prevent harmful or unethical responses.
• Personality Refinement: Clear instructions on how the model should present itself, emphasizing a helpful and respectful demeanor.
• Interaction Style: Detailed guidance on maintaining a conversational tone without becoming overly subservient or dismissive.

Key Highlights from the System Prompts

Establishing the Model’s Personality

Claude 4 is designed to be more than just a tool; it's meant to be an engaging and helpful assistant. The system prompts emphasize:

• Friendly and Professional Tone: Claude should maintain a conversational, friendly tone while remaining professional.
• Respectful and Helpful: It should always be respectful and provide useful, accurate information.

Model Safety

Safety is a top priority in the new prompts, with specific instructions to prevent harmful or unethical behavior:

• No Harmful Content: The model must not generate content that could cause harm, including explicit language, violent suggestions, or discriminatory statements.
• Ethical Boundaries: It should avoid engaging in discussions that promote illegal activities or violate ethical standards.

More Points on Style

The prompts provide detailed guidance on the model's interaction style to ensure a natural and effective conversation:

• Avoid Sycophancy: The model should not overly flatter or agree with users, maintaining an objective stance.
• Cognizance of Red Flags: It should be alert to potential red flags in user requests and respond appropriately.

Knowledge Cutoff Date

The knowledge cutoff date for Claude 4 is a bit ambiguous. The system prompts mention:

• January vs. March: Some sections indicate a January 2025 cutoff, while others suggest March 2025. This discrepancy might be addressed in future updates.

Notably Removed Since Claude 3.7

Comparing the new prompts to those of Claude 3.7, several changes are evident:

• Removed Redundant Warnings: Some warnings and guidelines that were present in earlier versions have been removed, likely because they are no longer necessary.
• Streamlined Instructions: The overall structure has been streamlined for clarity and ease of use.

Differences Between Opus 4 and Sonnet 4

While both models share many commonalities, there are some differences:

• Opus 4: Tends to be more focused on general tasks and conversations.
• Sonnet 4: Specializes in creative writing and more nuanced interactions.

Election Info

The system prompts also include specific instructions for handling election-related queries:

• Neutral Stance: The model should maintain a neutral stance on political issues and avoid taking sides.
• Fact-Based Responses: It should provide factual information based on reliable sources.

The Missing Prompts for Tools

Interestingly, Anthropic did not publish the system prompts for some tools. Simon Willison provides insights into these missing prompts:

• Leaked Tool Prompts: Some tool-specific prompts have been leaked, offering a glimpse into how Claude 4 is guided to interact with various applications and services.

Conclusion

The new system prompts for Claude Opus 4 and Claude Sonnet 4 reflect Anthropics' commitment to enhancing model safety, refining personality, and improving interaction style. These changes not only make the models more reliable but also more engaging for users.