Comparing Claude System Prompts Reveal Anthropic's Priorities

a year ago

Claude 4.0's system prompt changes reveal Anthropic's priorities and development cycle.
Old hotfixes from Claude 3.7 are removed, indicating reinforcement learning addresses past issues.
New hotfixes, like avoiding flattery in responses, are added to address recent issues like GPT-4o's behavior.
Search is now encouraged without user approval, reflecting increased user reliance on Claude for search tasks.
Structured documents are more nuanced in Claude 4.0, catering to user needs like meal plans and study guides.
Context limit issues are addressed with concise variable names in code artifacts to maximize content.
Cybercrime is a new guardrail, with strict refusal to assist with malicious code or harmful protocols.
System prompts define chatbot UX and reveal Anthropic's user-driven development process.

Hasty Briefsbeta