Anthropic publishes the system prompts for their Claude models. That’s unusual in this industry — and it gives someone like Simon Willison the perfect opportunity to dig deep. On Friday, the AI blogger and developer did exactly that: he converted all published system prompts into a git repository, using fake commit dates for each version. The result? You can run git diff between Claude Opus 4.6 (February 2026) and Opus 4.7 (April 2026) and see exactly what changed.
What Changed Between 4.6 and 4.7
The changes are revealing — not because they’re technically revolutionary, but because they show where Anthropic is focusing its attention.
Child safety significantly tightened: Instructions for protecting minors were moved into a dedicated <critical_child_safety_instructions> block and substantially expanded. Once Claude refuses a request for child safety reasons, all subsequent requests in the same conversation must be approached “with extreme caution.”
Less pushy: Claude was instructed to respect when users want to end conversations instead of pushing for more interaction. The model should also “make a reasonable attempt now, not be interviewed first” — in other words, fewer clarifying questions, more action.
Tool search instead of ‘I can’t’: There’s now a tool_search mechanism. Before Claude claims it doesn’t have access to a tool, it must first check whether a matching tool exists but simply hasn’t been loaded yet.
Shorter and more concise: New instructions emphasize focused, brief responses. Caveats should be mentioned briefly, not elaborated in full paragraphs.
Platform rename: The “Developer Platform” is now called “Claude Platform.” New additions to the system prompt include “Claude in PowerPoint” and other product integrations.
Why This Matters
System prompts are the operating system of an AI model — they define how it behaves, what it can do, and what it can’t. The fact that Anthropic publishes these is a transparency signal that OpenAI and Google haven’t matched. And Willison’s idea of turning them into a git timeline is brilliant: instead of slogging through markdown documents, any developer can trace changes using familiar tools.
The shifts in the prompt also reveal what users have been running into in practice: too many clarifying questions, overly long responses, missing tool recognition. Anthropic is responding to real feedback — and the system prompt is the proof.
Sources: