What's new in Claude Opus 4.7's system prompt

Anthropic publishes the system prompts used on their web/mobile/desktop apps (not the API) at platform.claude.com/docs/en/release-notes/system-prompts. A quick diff against Opus 4.6:

• New “4.7” model family. Opus 4.7 is the only member. Referenced model IDs update to claude-opus-4-7, claude-sonnet-4-6, and claude-haiku-4-5-20251001.

  “This iteration of Claude is Claude Opus 4.7 from the Claude 4.7 model family. The Claude 4.7 family currently consists of Claude Opus 4.7.”

  “The most recent Claude models are Claude Opus 4.7, Claude Sonnet 4.6, and Claude Haiku 4.5, the exact model strings for which are 'claude-opus-4-7', 'claude-sonnet-4-6', and 'claude-haiku-4-5-20251001' respectively.”

• Knowledge cutoff moves from end of May 2025 to end of January 2026.

  4.6: “Claude’s reliable knowledge cutoff date … is the end of May 2025.”

  4.7: “Claude’s reliable knowledge cutoff date … is the end of January 2026.”

• Claude Cowork gets expanded treatment, with explicit notes on the agent products it can drive as tools (Chrome, Excel, PowerPoint).

  4.6: “Cowork - a desktop tool for non-developers to automate file and task management.”

  4.7: “Claude can be used via Claude Cowork, an agentic knowledge work tool for non-developers that is available as a desktop app.”

  “Claude Cowork can use all of these as tools.”

• Dedicated <critical_child_safety_instructions> block. Previously general language, now a structured section with much more detail.

  4.7 adds a dedicated section: \<critical_child_safety_instructions\>

  “These child-safety requirements require special attention and care … Claude strictly follows these rules:”

• Refusal handling now includes an explicit nested child-safety subsection instead of keeping all of that guidance inline.

  4.6: “Claude cares deeply about child safety and is cautious about content involving minors …”

  4.7: \<critical_child_safety_instructions\>

• User wellbeing guidance grows, especially around crisis situations.

  “When discussing means restriction or safety planning with someone experiencing suicidal ideation or self-harm urges, Claude does not name, list, or describe specific methods …”

  “If Claude suspects the person may be experiencing a mental health crisis, Claude should avoid asking safety assessment questions.”

• Tool-use and capability guidance gets more specific, especially around acting vs clarifying and checking for available integrations before claiming Claude lacks a capability.

  “When a request leaves minor details unspecified, the person typically wants Claude to make a reasonable attempt now, not to be interviewed first.”

  “Before concluding Claude lacks a capability … Claude calls tool_search to check whether a relevant tool is available but deferred.”

The technical shift here is that Anthropic stopped treating tools as optional props. <capability_check> flips the default: Claude can’t say “I don’t have access to X” until tool_search says so, which is cool. “I can’t” is now a claim that has to be verified against the runtime, not a feeling the model is allowed to have. <acting_vs_clarifying> backs this up with three simple rules. One: if the request is vague, take a reasonable shot instead of asking follow-up questions. Two: if a tool can answer the question for you, use the tool instead of asking the user. Three: if the user asked for something to be done (send the message, add the task, post the update), actually do it — writing out a draft for them to copy-paste is not the same as doing it.

Another interesting thing is that the prompt now tells Claude some tools aren’t loaded yet and it has to go ask for them. Basically, you only bother saying that if there are too many tools to show the model all at once. So the setup seems to have changed.

Is Anthropic swapping the fixed toolbox for an on-demand tool library?