Anthropic revised the Claude.ai system prompt for Opus 4.7 released April 16 2026 expanding child safety rules tool references and behavioral directives according to Simon Willison's diff analysis of the public archive.

The prompt renames the developer platform to Claude Platform adds Claude in Powerpoint to the listed tools and introduces a <critical_child_safety_instructions> wrapper stating that once a request is refused for child safety all subsequent requests in the conversation must be approached with extreme caution (Simon Willison https://simonwillison.net/2026/Apr/18/opus-system-prompt/ Anthropic Claude System Prompt Archive July 2024). New <acting_vs_clarifying> language directs the model to use tool_search for capabilities such as location memory or calendar before stating lack of access and to complete tasks without partial stops a mechanism first described in Anthropic's November 2025 tool-use post.

Instructions now require concise responses with brief disclaimers and add a disordered eating section prohibiting specific nutrition numbers or diet guidance; prior language forbidding asterisk emotes and words such as genuinely honestly or straightforward has been removed indicating changed model behavior (Simon Willison https://simonwillison.net/2026/Apr/18/opus-system-prompt/ Anthropic API Documentation November 2025). Original coverage omitted that these incremental edits continue the transparency pattern maintained since Claude 3 in July 2024 and parallel the Constitutional AI approach published by Anthropic in 2022.

The archive shows Anthropic alone among major labs publishes successive full system prompts allowing direct comparison of safety alignment and verbosity priorities as model scale increases.