System prompt changes between Claude Opus 4.6 and 4.7

Based on the system prompt for Claude.ai released by Anthropic, this piece organizes the changes between Opus 4.6 and 4.7 as a diff.

Simon Willison split the Markdown version of the system prompt into per-model documents, then built a fake git history matched to the release dates to compare them. The result shows that in 4.7, safety policy, tool-use approach, and response tone were noticeably adjusted.

• Naming change: "developer platform" was changed to "Claude Platform".
• Expanded tool scope: Claude in Chrome, Claude in Excel, and Claude in Powerpoint were added to Claude's list of tools, and it's now stated that Claude Cowork can use these as tools.
• Strengthened child safety: A new <critical_child_safety_instructions> tag was added, stating that after refusing once for child safety reasons, subsequent requests in the same conversation should also be treated with extreme caution.
• Respecting conversation endings: Changed so that when a user wants to end the conversation, Claude should respect that intent rather than holding on or trying to prompt additional turns.
• Acting vs clarifying: For requests missing minor information, Claude should start with a reasonable attempt rather than interviewing the user first. In particular, if a tool call can resolve the ambiguity, it should try a tool call first before asking a person. There's also an instruction to see tasks through to completion once started.
• Context for introducing tool_search: Before assuming it has capabilities like location, memory, calendar, files, past conversations, or external data, Claude is told to check whether relevant tools exist using tool_search. It can only say a tool doesn't exist after searching.
• Stronger emphasis on conciseness: Responses should be short and focused so as not to overwhelm the user with excessive length.
• Some expression rules removed: The 4.6 instructions about emote/asterisk behavior, and the instruction to avoid "genuinely", "honestly", "straightforward", were removed in 4.7.
• Added handling for disordered eating: If signs of an eating disorder appear, Claude is now explicitly told not to give precise numbers, goals, or step-by-step diet/exercise instructions anywhere else in the conversation either.
• Avoiding forced yes/no answers: Even when asked to give a simple yes/no on a complex or controversial topic, Claude can decline to give a short answer and explain context when necessary.
• Removal of an existing political statement: The explicit sentence from 4.6 along the lines of "Donald Trump is the current president..." is gone. This is interpreted as a change aligned with the new model's knowledge cutoff of January 2026.

The piece also notes that Anthropic's published system prompt alone isn't the whole story—separate tool descriptions are attached in the actual Chat UI. When asked directly about its tool list, Claude reveals several tools including ask_user_input_v0, bash_tool, conversation_search, tool_search, web_search, and weather_fetch, and the author notes this list hasn't changed much from the 4.6 era.