Anthropic expands child safety rules and tool use in Claude Opus 4.7 system prompt

Anthropic published updated system prompt instructions for Claude Opus 4.7 on April 16, 2026, roughly two months after the previous version released in February. The company maintains a public archive of system prompts dating back to Claude 3 in July 2024, making it the only major AI lab to regularly publish such documentation. Researcher Simon Willison analyzed the git diff between versions and highlighted the substantive changes.

The most structurally significant update wraps child safety guidance in a new XML tag labeled critical_child_safety_instructions, signaling elevated priority. The rules now explicitly state that once Claude refuses any child safety request, all subsequent conversation turns must be approached with extreme caution—a ratcheting mechanism that prevents workarounds through topic switching or reframing.

Claude's tool interaction logic underwent substantial revision. A new section titled acting_vs_clarifying instructs the model to attempt reasonable completion of underspecified requests rather than asking clarifying questions upfront. When ambiguity exists and tools could resolve it—searches, location lookups, calendar checks—Claude is now directed to call the tool before requesting user input. This reflects a broader shift toward autonomous tool use.

A complementary addition documents the tool_search mechanism, which instructs Claude to query available tools before declaring missing capabilities. The prompt specifies that claims like 'I don't have access to X' are only valid after tool_search confirms no matching tool exists. This addresses a known failure mode where models incorrectly claim capability gaps.

Safety guardrails expanded in other areas. A new section on disordered eating instructs Claude to avoid providing specific nutrition targets, diet numbers, or exercise plans if users display eating disorder indicators, even if framed as health guidance. The rationale explicitly acknowledges that detailed prescriptions could trigger harmful behavior. A revised evenhandedness section guards against 'screenshot attacks' where users demand single-word answers to contested questions; Claude can now decline and offer nuanced reasoning instead.

Less prominent but indicative of model behavior improvements: instructions encouraging conciseness were added, legacy rules about avoiding asterisk-based actions were removed (suggesting the newer model doesn't default to that style), and a specific clarification about Donald Trump's presidency status was dropped—reflecting an updated knowledge cutoff to January 2026 that makes such corrections unnecessary.

Anthropic also documented available tools through a direct query to Claude. Twenty-four named tools are now exposed, including conversation_search, tool_search, web_search, bash_tool, and new visualization and connector recommendations features. The tool list appears unchanged from Opus 4.6, indicating that capability expansion in this version focused on instruction refinement rather than new tool deployment.