Inside the Claude Opus 4.7 System Prompt Update

Simon Willison's analysis of the Claude Opus 4.7 system prompt reveals a significant push toward autonomous tool use and expanded safety protocols. Key changes include the addition of a PowerPoint agent, stricter guidelines for child safety and disordered eating, and a more concise communication style. The update also reflects an advanced knowledge cutoff of January 2026, allowing the model to handle recent political facts without specific hardcoded prompts.

Key Points

Claude is now directed to prioritize taking action with tools over asking clarifying questions when details are minor or tools can provide the answer.
Safety instructions have been significantly expanded, featuring a new critical child safety tag and specific prohibitions on providing diet or exercise plans to users showing signs of disordered eating.
The system prompt introduces 'Claude in Powerpoint' as a new productivity agent and details a 'tool_search' mechanism to check for deferred capabilities.
Anthropic has refined Claude's conversational style to be more concise, less repetitive, and more resistant to 'yes/no' traps on controversial topics.
The removal of specific instructions about the 2025 U.S. inauguration indicates the model now has a reliable knowledge cutoff of January 2026.

Sentiment

The community is predominantly critical of the changes revealed in the system prompt update. While some users appreciate the transparency of Anthropic's published prompts and find individual changes reasonable, the overall tone is one of frustration — particularly around the acting-vs-clarifying shift, safety guardrail bloat, and malware paranoia. Power users feel Anthropic is making choices that serve the lowest common denominator at the expense of experienced users, and several are exploring alternatives or building agent-neutral codebases.

In Agreement

The acting-vs-clarifying change reduces friction for experienced users who find excessive clarification questions annoying
The eating disorder section is a reasonable, common-sense addition given Anthropic's liability exposure and the prevalence of eating disorders
System prompt safety measures represent standard product liability protection, similar to warning labels on consumer products
The system prompt serves as a useful hotfix mechanism that can be updated without retraining the model
The third-person 'Claude does X' style helps the model maintain a consistent identity and roleplay its character more effectively

Opposed

Claude should ask clarifying questions instead of guessing, as wrong assumptions waste far more tokens and time than upfront questions would
Adding niche safety guardrails like eating disorder sections bloats the system prompt and degrades quality for all users
The malware paranoia is severely impacting legitimate coding work, with Claude wasting tokens checking every file for malware and sometimes refusing to edit ordinary code
The growing system prompt represents concerning censorship and moral control by a single corporation over what users can learn and discuss
Anthropic may be optimizing for metrics that look superficially better when the model avoids asking questions, prioritizing engagement over user satisfaction
Claude 4.5 was significantly better than subsequent versions, and users are exploring alternatives or building agent-neutral codebases to reduce vendor lock-in