Claude Opus 4.7 system prompt: less verbose, tool-first, tighter child safety

Simon Willison diffed Anthropic’s published system prompts between Claude Opus 4.6 (February 2026) and 4.7 (April 16, 2026), surfacing meaningful behavioral shifts. The model is nudged toward acting over asking: when details are ambiguous, it should call tools like search, location, or calendar lookup before interrogating the user, and a new tool_search mechanism must be consulted before Claude claims it lacks a capability. Responses are explicitly steered to be shorter and more focused, and Claude is told to respect users who want to end a conversation rather than angling for another turn.

Safety-oriented sections expanded. Child safety guidance now sits under a dedicated critical_child_safety_instructions block, with the rule that any refusal on those grounds puts the rest of the conversation on heightened alert. A new passage addresses disordered eating, barring specific nutrition, diet, or exercise numbers even when framed as helpful. An evenhandedness clause lets Claude refuse forced yes/no answers on contested topics - a direct counter to screenshot-gotcha attacks.

Some scaffolding dropped out because the underlying problem is gone: the explicit note that Trump is the sitting president was removed now that the January 2026 knowledge cutoff covers it, and prior anti-emote and banned-word rules (“genuinely”, “honestly”, “straightforward”) were cut, implying 4.7 no longer drifts that way unprompted. The tool roster (bash_tool, web_search, conversation_search, tool_search, and others) appears unchanged from 4.6, though Anthropic still doesn’t publish the tool descriptions themselves.