Changes in the system prompt between Claude Opus 4.6 and 4.7 (simonwillison.net)

Article summary

The article discusses the changes in the system prompt between Claude Opus 4.6 and 4.7, highlighting updates such as the introduction of a new 'acting_vs_clarifying' section, expanded child safety measures, and a tool search mechanism. The author analyzes the differences in the system prompts and their potential impact on the model's behavior. The updates aim to make Claude less pushy, more concise, and better at handling complex requests. The article also mentions the availability of Claude Code system prompt diffs and the list of named tools provided to the model.

Main themes

• AI model updates
• System prompt changes
• Child safety measures
• Tool search mechanism
• Conversational AI
• Malware detection

What commenters say

• The 'acting_vs_clarifying' change has improved the model's behavior by reducing the number of clarifying questions asked before taking action.
• Some users prefer the model to ask clarifying questions before taking action, rather than making assumptions and potentially causing errors.
• The model's new behavior of making reasonable attempts without asking questions first can lead to incorrect interpretations and frustration for users.
• The introduction of malware detection measures is seen as a necessary step to prevent the model from being used for malicious purposes, but some users find it overly cautious and restrictive.
• The model's tendency to contemplate malware development and ask unnecessary questions can waste token budget and hinder productivity.
• Some users have found workarounds to the model's new behavior, such as using tweakcc to patch the system prompts or adding specific instructions to their prompts.
• The model's behavior is seen as a trade-off between being helpful and being cautious, with some users preferring a more proactive approach and others preferring a more cautious one.
• The discussion highlights the importance of user preference and the need for customizable models that can adapt to individual users' communication styles and needs.