Anthropic renamed its “developer platform” to the “Claude Platform” and expanded the list of built-in tools. The 4.7 prompt now explicitly calls out “Claude in Chrome” for autonomous browsing, “Claude in Excel” for spreadsheets and “Claude in Powerpoint” for slides—tools that Claude Cowork can invoke. It also introduces a new tool_search API step: before saying it can’t perform an action, Claude first checks available tools (location access, calendar lookup, data fetches) and only admits defeat if none match. Safety and clarity sections got a major overhaul. Child-safety instructions are wrapped in a <critical_child_safety_instructions> tag and stress that once Claude refuses a request for safety reasons, it must stay cautious in that conversation. If you ask to end the chat, Claude will respect it rather than push you to stay. A fresh <acting_vs_clarifying> block tells Claude to proceed with reasonable defaults when details are missing, to call tools before asking questions and to finish tasks completely instead of stopping mid-response. Anthropic also trimmed sections aimed at curbing comic asterisks and words like “honestly,” since 4.7 behaves better without those guardrails. New guidelines warn Claude against giving precise diet or exercise numbers if disordered eating is suspected. And to counter screenshot-style attacks, an <evenhandedness> rule now lets Claude decline one-word yes/no prompts on complex issues and offer nuanced explanations instead. Finally, the outdated Trump-president clarification was removed, reflecting the updated knowledge cutoff of January 2026.