Claude Opus 4.7 system prompt changes detailed in diff analysis
A diff between the Claude.ai system prompts for Opus 4.6 and Opus 4.7 reveals notable changes including new tool integrations, expanded child safety rules, reduced verbosity guidance, and new protections against screenshot-based prompt attacks.
Score breakdown
Developers and practitioners building on Claude can use this diff to understand exactly how Anthropic is shaping model behavior — including new tool-discovery mechanics via `tool_search`, stricter safety escalation rules, and reduced verbosity defaults — which directly affects how Claude-powered agents will respond in production.
- 01Opus 4.7 shipped on April 16, 2026; the previous Opus 4.6 system prompt was dated February 5, 2026.
- 02Anthropic is described as the only major AI lab to publicly publish system prompts for its user-facing chat systems.
- 03New tools listed in the Opus 4.7 prompt include 'Claude in Chrome', 'Claude in Excel', and 'Claude in PowerPoint' — the last was absent from the 4.6 prompt.
The article examines the git diff between the Claude.ai system prompts for Opus 4.6 and Opus 4.7, using a Git history constructed with Claude Code to track prompt evolution over time. Anthropic's system prompt archive now stretches back to Claude 3 in July 2024, and the article highlights the most significant changes between the two recent versions. On the tooling side, the "developer platform" has been renamed to the "Claude Platform", and the prompt now explicitly lists "Claude in Chrome" (a browsing agent), "Claude in Excel" (a spreadsheet agent), and "Claude in PowerPoint" (a slides agent) — the last of which was absent from the 4.6 prompt. A new `tool_search` mechanism is also described: before Claude claims it lacks access to something like a user's location, memory, or calendar, it must call `tool_search` to confirm no matching tool exists.
Conversely, guidance was added to keep responses focused and concise.
Several behavioral guardrails were updated or added. The child safety section was greatly expanded and wrapped in a new tag, with a notable rule that once Claude refuses a request for child safety reasons, all subsequent requests in the same conversation must be approached with extreme caution. New language discourages Claude from being overly cautious about ambiguity — it should attempt tasks with reasonable assumptions rather than asking clarifying questions upfront, and should use available tools to resolve ambiguity rather than asking the user to do lookups themselves. Conversely, guidance was added to keep responses focused and concise. A section warning against the use of emotes or asterisk-style actions and discouraging words like "genuinely" and "honestly" was removed, presumably because Opus 4.7 no longer exhibits those behaviors. New sections address disordered eating (Claude should avoid specific numerical diet or exercise guidance if a user shows signs of the condition) and screenshot-based prompt attacks that try to force a simple yes/no answer on complex or contested topics.
Key facts
- 01Opus 4.7 shipped on April 16, 2026; the previous Opus 4.6 system prompt was dated February 5, 2026.
- 02Anthropic is described as the only major AI lab to publicly publish system prompts for its user-facing chat systems.
- 03New tools listed in the Opus 4.7 prompt include 'Claude in Chrome', 'Claude in Excel', and 'Claude in PowerPoint' — the last was absent from the 4.6 prompt.
- 04A new `tool_search` mechanism requires Claude to check for available tools before claiming it lacks access to a capability.
- 05