Simon Willison
@simonw
Since Anthropic publish their system prompts we can generate a diff between Claude Opus 4.6 and 4.7 - here are my notes on what's changed https://t.co/IQHuvLGmwO
100retweets925likes
View on XSimon Willison Surfaces Claude Opus 4.7 System Prompt Changes Toward Autonomy
· Updated
Simon Willison, creator of Datasette, generated a git-style diff of Anthropic's published system prompts to identify behavioral changes in Claude Opus 4.7. He surfaced a new slides agent called Claude in PowerPoint and a mandatory
tool_search call that prevents the model from prematurely claiming it lacks a capability.- New agent
- Claude in PowerPoint
- Tool search
- Mandatory before capability refusal
- Behavioral shift
- Act before clarifying
- Knowledge cutoff
- January 2026
- New safety tags
- Child safety and disordered eating
These prompt changes signal a shift from reactive conversational assistance toward proactive autonomous agency. By instructing the model to make a "reasonable attempt" rather than "interviewing" the user, Anthropic is following a broader industry trend. This builds on the recent Opus 4.7 launch emphasizing self-verification for autonomous agents.
Users should expect more proactive tool use and fewer clarifying questions. Stricter safety guardrails for child safety and disordered eating are now active, alongside a new January 2026 knowledge cutoff. These updates coincide with increased rate limits designed to handle the higher token consumption of autonomous reasoning.
Still wondering? A few quick answers below.
Simon Willison's analysis revealed that Anthropic updated the Claude Opus 4.7 system prompt with a new PowerPoint agent and a mandatory tool search function. The diff also surfaced a new January 2026 knowledge cutoff and expanded safety instructions for child safety and disordered eating, signaling a move toward more autonomous behavior.
The shift toward autonomous action means Claude is now instructed to make a reasonable attempt at a task before asking for clarification. Instead of interviewing the user for minor details, the model will proactively use tools like search or calendar access to resolve ambiguities, reducing conversational friction and speeding up task execution.
Users should expect Claude to be more proactive and less verbose, as the model now prioritizes acting with tools over asking questions. While the model will be more autonomous, it also features stricter safety guardrails for specific topics and a more current knowledge base that reflects events through early 2026.
