Anthropic updates Claude Opus 4.7 system prompt and tool handling

Anthropic has published the system prompt changes for Claude Opus 4.7, giving a rare look at how a major chat model is steered behind the scenes. The company has been unusually open about releasing prompt archives for its user-facing assistants, and this latest update continues that pattern by showing what changed between Opus 4.6 and 4.7.

A system prompt is the hidden set of instructions that shapes how an AI assistant responds. It sits above the user’s message and can affect tone, safety behavior, tool use, refusal patterns, and how the model handles uncertainty. For people building AI automation systems, those instructions are often as important as the underlying model weights, because they determine how the assistant behaves in production.

The most visible change in the 4.7 prompt is stronger child-safety language. Anthropic has been tightening this area across its models, and the updated instructions make clear that the assistant is expected to handle requests involving minors with more caution and stricter boundaries than before. That matters because general-purpose chat systems are increasingly used in settings where age-sensitive content, moderation, and escalation rules need to be explicit rather than implied.

The prompt also adds new guidance around tool use. In current AI assistants, tools can mean anything from web search and file access to internal functions, databases, or external APIs. When a model is allowed to call tools, the quality of the system prompt often decides whether it asks for clarification, searches first, or takes a shortcut that could produce a wrong answer.

Another part of the update focuses on ambiguity handling. That usually means telling the model what to do when user intent is unclear, when the question could have multiple interpretations, or when the assistant does not have enough context to answer cleanly. In practice, this can change whether the model makes an assumption, asks a follow-up question, or states uncertainty more directly.

Simon Willison, who highlighted the changes in his write-up, has long tracked prompt disclosures and model behavior across major AI vendors. Anthropic’s prompt archives have become a useful reference point for comparing how different versions of Claude are instructed to behave, especially as vendors add more tool routing and more guardrails around sensitive topics.

The updated knowledge cutoff language is also notable. The prompt now refers to January 2026, which signals what the model is expected to know by default and what lies beyond its built-in training horizon. Knowledge cutoffs matter because users often assume a chat model has current awareness when, in reality, its answers are bounded by the point at which its knowledge was frozen unless live tools are used.

Prompt updates like this do not usually make headlines outside the AI circle, but they can reveal meaningful shifts in product behavior. A change in system instructions can alter how often the assistant refuses, whether it consults tools before answering, and how it frames uncertainty when the information is incomplete.

Anthropic has been publishing these prompt diffs as part of its public-facing documentation practice, and the 4.7 version is another example of the company treating the system prompt as a visible part of the product rather than a fully hidden implementation detail. The published changes show new safety language, new tool instructions, new ambiguity handling, and a knowledge cutoff set to January 2026.