HN
Today

Changes in the system prompt between Claude Opus 4.6 and 4.7

Simon Willison provides a meticulous breakdown of the system prompt evolution between Anthropic's Claude Opus 4.6 and 4.7. This deep dive offers rare transparency into how major AI labs fine-tune model behavior, safety, and interaction. The article reveals Anthropic's ongoing efforts to enhance user experience and address critical safety concerns, resonating with HN's interest in practical AI development.

12
Score
0
Comments
#6
Highest Rank
9h
on Front Page
First Seen
Apr 19, 11:00 AM
Last Seen
Apr 20, 12:00 AM
Rank Over Time
1211121798667

The Lowdown

Simon Willison's article meticulously details the changes in Anthropic's Claude system prompt between the Opus 4.6 and Opus 4.7 versions. Anthropic is notable for publicly archiving these system prompts, offering a unique window into the continuous refinement of their AI models. Willison leveraged Claude Code to generate a Git history of these prompt updates, enabling a clear comparison of the evolution over time.

Key changes identified in the system prompt include:

  • The "developer platform" is now referred to as the "Claude Platform."
  • New tools, such as "Claude in Powerpoint," are explicitly mentioned in the list of capabilities.
  • The child safety section has been significantly expanded and encapsulated in a new <critical_child_safety_instructions> tag, emphasizing extreme caution in subsequent interactions after an initial safety refusal.
  • Claude is instructed to be less "pushy," respecting user decisions to end conversations.
  • A new <acting_vs_clarifying> section directs Claude to make reasonable attempts to act on requests with minor unspecified details rather than asking clarifying questions first.
  • The system prompt indicates an embedded tool_search mechanism, enabling Claude to actively seek relevant tools before declaring a lack of capability.
  • New directives encourage Claude to keep responses focused and concise to avoid overwhelming users.
  • Instructions present in 4.6, such as avoiding emotes or specific filler words (e.g., "genuinely"), have been removed for 4.7, suggesting these behavioral issues have been mitigated in the new model.
  • A new section addresses "disordered eating," advising Claude against providing precise nutritional or exercise guidance in such cases.
  • The <evenhandedness> section now prevents Claude from giving simple yes/no answers to complex or contested issues, promoting nuanced responses instead.
  • Language from 4.6 confirming "Donald Trump is the current president" has been removed, reflecting the model's updated knowledge cut-off date beyond January 2026.

Beyond the system prompt, the article also explores Claude's available tools, obtained by directly querying the model. While the list of named tools is provided, it does not appear to have changed since Opus 4.6, emphasizing that the system prompt updates are primarily focused on behavioral guidance and safety protocols.