Claude Opus 4.7: Revolutionizing System Prompts at Anthropic

⚡

Key Takeaways

1Anthropic has updated Claude's system prompt with Opus 4.7, launched on April 16, 2026, introducing new features and adjustments.

2The Claude tools now include Claude in Powerpoint, which was absent from the previous version, and increased attention to child safety.

3Claude Opus 4.7 introduces mechanisms to reduce verbosity and handle ambiguous requests with tools before querying the user.

💡Why it matters — These developments enhance Claude's ability to interact more securely and effectively while respecting user preferences.

Anthropic, a major player in the field of artificial intelligence, stands out for its transparency by publishing the system prompts of its chat models. This practice, uncommon among large AI labs, allows for the observation of model evolution over time. On April 16, 2026, Anthropic launched Claude Opus 4.7, bringing significant changes compared to the previous version, Opus 4.6, which was released on February 5, 2026.

Changes in Tools and Platform

One of the first notable changes is the renaming of the "developer platform," now called the "Claude Platform." This change in terminology comes with an update to the list of available tools. Among the new additions, Claude in Powerpoint, a presentation agent, has been added to the existing tools like Claude in Chrome, an autonomous browsing agent, and Claude in Excel, a spreadsheet agent. Claude Cowork, another tool, is capable of utilizing all of these features. It is important to note that Claude in Powerpoint was not mentioned in the prompt 4.6.

Safety and User Interaction

Child safety has been enhanced in this new version. A new tag, <critical_child_safety_instructions>, has been introduced to frame interactions. If Claude refuses a request for child safety reasons, all subsequent requests in the same conversation must be approached with extreme caution. Additionally, Claude has been adjusted to be less insistent: when a user wishes to end a conversation, Claude respects that decision without attempting to prolong the interaction.

Handling Requests and Verbosity

A new section, <acting_vs_clarifying>, has been introduced to guide Claude in managing ambiguous requests. When minor details are unspecified, Claude is encouraged to make a reasonable attempt without asking for clarification, unless the request is impossible to fulfill without additional information. Claude prioritizes the use of tools to resolve ambiguities before soliciting the user. Once a task is initiated, Claude commits to providing a complete response rather than stopping midway.

Furthermore, an effort has been made to reduce the verbosity of Claude's responses. The language has been adjusted to encourage more concise answers, even when warnings or caveats need to be included. The goal is to avoid overwhelming the user with overly long responses while keeping the essence of the answer focused on the main topic.

Tool Search Mechanism

Claude's chat now features a tool search mechanism, as indicated in the API documentation and described in a post from November 2025. Before concluding that Claude lacks a capability—such as access to the user's location, memory, calendar, files, past conversations, or any external data—Claude calls tool_search to check if a relevant tool is available but deferred. "I do not have access to X" is only correct after tool_search confirms that no corresponding tool exists.

Disordered Eating

A new section on disordered eating has been introduced. If a user shows signs of disordered eating, Claude must not provide specific advice on nutrition, diet, or exercise—no specific numbers, goals, or step-by-step plans—anywhere else in the conversation. Even if this aims to help establish healthier goals or highlight the potential dangers of disordered eating, responses containing these details could trigger or encourage disordered trends.

Other Adjustments

Claude Opus 4.7 also includes protections against screenshot attacks, which seek to obtain simplistic answers to complex questions. The system prompt now allows Claude to provide nuanced responses rather than simple "yes" or "no." Finally, adjustments have been made regarding the model's knowledge cutoff date. The prompt 4.6 contained a section stating that "Donald Trump is the current president of the United States and was inaugurated on January 20, 2025," because without this, the model's knowledge cutoff combined with its previous knowledge that Trump falsely claimed to have won the 2020 election meant it would deny that he was president. This language has been removed for 4.7, reflecting the model's new reliable knowledge cutoff date of January 2026.

Limitations of Published Information

Although the system prompts are available, they do not contain descriptions of the tools provided to the model. However, it is possible to directly ask Claude for the list of available tools, which includes features such as ask_user_input_v0, conversation_search, and fetch_sports_data. This list appears unchanged from Opus 4.6.