Changes in the system prompt between Claude Opus 4.6 and 4.7

The Catalyst: A "Transparent" System Prompt Update In an industry where system prompts are often treated as core secrets, Anthropic's practice of publicly releasing the system prompts for Claude.ai offers a unique window into AI development. Simon Willison's analysis of the changes between Claude Opus 4.6 and 4.7 is more than a technical "spot the difference" exercise; it's a prime case study in understanding how AI assistants are being shaped and "tamed." This update, rolled out in April 2026 just over two months after the last one, is packed with insights worth unpacking. Deconstructing the Core Changes: The Making of an AI's "Personality" First, the explicit expansion and integration of the tool ecosystem. The system prompt now clearly lists tools like "Claude in Excel" and "Claude in Powerpoint," noting that "Claude Cowork can use all of these as tools." This is far more than a simple feature list update. It signals that the AI assistant's role is evolving from a standalone chatbot into a collaboration hub embedded within specific productivity environments. Interactions with AI will increasingly occur within familiar applications like Word, Excel, and PowerPoint, where the AI acts as an intelligent agent. This foreshadows that the primary battleground for AI applications is shifting from independent chat interfaces into existing software ecosystems. Second, a fundamental shift in interaction logic: from "Questioner" to "Actor". The new <acting_vs_clarifying> section is the heart of this update. It explicitly instructs Claude: "When a request leaves minor details unspecified, the person typically wants Claude to make a reasonable attempt now, not to be interviewed first." Crucially, it stipulates that when a tool is available to resolve ambiguity—like searching or checking a calendar—Claude should call the tool first, rather than asking the user to look it up themselves. This fundamentally alters the AI interaction paradigm. We're accustomed to AI asking endless clarifying questions to achieve precision; now, it's encouraged to "try and figure it out" first. This reflects confidence in the AI's capabilities (especially tool use) and a deep insight into user experience: users want results, not a lengthy requirements elicitation session. Third, the refinement and proactivity of safety guardrails. The child safety section has been significantly expanded and highlighted with a <critical_child_safety_instructions> tag. One rule is particularly noteworthy: "Once Claude refuses a request for reasons of child safety, all subsequent requests in the same conversation must be approached with extreme caution." This transcends single-instance content filtering, establishing a conversation-level risk assessment mechanism. The AI's safety judgment is no longer one-off; it gains "memory" and "contextual awareness," reflecting a deeper understanding of potential risks in complex, multi-turn interactions. Finally, and most forward-looking, the AI's "self-awareness" and capability discovery mechanism. The system prompt introduces instructions for the tool_search tool: before claiming "I don't have access to X," Claude must first call tool_search to check if a relevant but deferred tool is available. This is a revolutionary concept. It means the AI is no longer entirely dependent on a pre-defined tool list hard-coded into the system prompt. It possesses a mechanism for dynamically discovering its own capabilities. It's akin to a person checking their skill set or toolbox before saying, "I can't do this." This lays the groundwork for a more flexible, extensible AI assistant architecture where the AI's capability boundaries can change dynamically rather than being statically fixed. Trend Insight: The AI Assistant as a "Proactive Digital Employee" Synthesizing these changes reveals a clear trend: AI assistants are evolving from "responsive tools" into "proactive digital employees." They are expected to: 1) Deeply integrate into workflows (function within specific software), 2) Proactively solve problems (act first, ask questions later), 3) Bear more complex safety responsibilities (contextual risk assessment), and 4) Possess a sense of self-management and capability discovery. Through iterative system prompt updates, Anthropic is meticulously crafting a behavioral framework for more autonomous, reliable, and integrated AI agents. Practical Value and Counter-Intuitive Insights For developers and product managers, this update offers a valuable blueprint for "behavioral design." If you're building your own AI agent, these principles—encouraging action over interrogation, designing tool discovery mechanisms, establishing conversation-level safety strategies—are directly applicable. A potentially overlooked, counter-intuitive point is this: the increasing length and complexity of system prompts is itself a critical product signal. As AI capabilities grow, the rules guiding their behavior don't become simpler; they become more complex and nuanced. It's akin to managing a more capable employee, which may require more elaborate regulations and higher communication overhead. Anthropic's publication of these lengthy and precise prompts恰恰展示了驾驭强大AI所需的“治理艺术”。 In conclusion, each update to Claude's system prompt is like a version iteration of the AI assistant's "mental model." Through these words, we see not just adjustments to technical parameters, but the continuous evolution of an AI product philosophy and interaction paradigm.