I’ve been tracking system prompt changes across major AI models for a while now, and Anthropic remains the only lab actually publishing theirs. That transparency alone is worth celebrating, but what’s more interesting is watching how their thinking about AI behavior evolves between releases.
The jump from Opus 4.6 to 4.7 is subtle but philosophically significant. These aren’t just tweaks to fix edge cases. They represent a fundamental shift in how Anthropic thinks Claude should interact with users.
Less Asking, More Doing
The most striking change is this new instruction: “When a request leaves minor details unspecified, the person typically wants Claude to make a reasonable attempt now, not to be interviewed first.”
This is huge. Earlier versions of Claude would often fall into the trap of asking clarifying questions before doing anything useful. You’d ask for help with a script and get back five questions about your exact requirements. It was technically correct but incredibly annoying in practice.
The new guidance says to just take your best shot unless the request is “genuinely unanswerable” without more information. And there’s that word again, “genuinely”, which coincidentally appears on the banned words list later in the prompt. Someone at Anthropic has been reading their own output logs.
What I find fascinating is the tool-first approach they’re pushing. If Claude has access to search, location data, or any other capability that could resolve ambiguity, it should use those tools before bothering the user. This is the right call. Why ask me for my timezone when you can just look it up?
The Persistence Problem
Another addition: “Once Claude starts on a task, Claude sees it through to a complete answer rather than stopping partway.”
I suspect this addresses a specific failure mode where Claude would start explaining something, then suddenly stop and ask if you want it to continue. It’s a weird hesitation that breaks flow, especially when you’ve explicitly asked for help with something complex.
This connects to another change about keeping responses “focused and concise” to avoid overwhelming users. There’s tension here that’s worth examining. You want completeness but not verbosity. You want thoroughness but not bloat. Finding that balance is genuinely (oops, banned word) one of the hardest problems in language model behavior design.
The Tool Discovery Dance
The instruction about tool_search is particularly interesting for developers. Before Claude claims it can’t do something, it’s now required to check whether a relevant tool exists but was deferred. This is a patterns I’ve seen emerging across artificial intelligence systems: don’t assume your capabilities, verify them.
The problem is that the published system prompts don’t include tool descriptions. Anthropic tells you the behavior guidelines but not what tools are actually available. You have to ask Claude directly, which feels like an unnecessary hurdle for developers trying to understand the full picture.
When you do ask, you get a list that includes bash_tool, text_editor, mcp_server_manager, and several others. These haven’t changed since 4.6, but knowing they exist helps explain why Claude behaves certain ways. If it has bash access, of course it’s going to suggest running commands. If it has memory tools, it should be checking those before claiming ignorance.
The Little Things That Matter
Some changes seem minor but reveal careful attention to user feedback. No more asterisk emotes unless specifically requested. Avoiding words like “honestly” and “straightforward” that add nothing but filler. These are the kind of verbal tics that make AI text feel artificial.
The guidance on contested topics is smart: if someone asks for a yes/no answer on something complex, Claude can now decline and explain why nuance matters. This sidesteps the trap of appearing to take strong positions while also not being uselessly fence-sitting.
There’s also new guidance around disordered eating that’s both careful and specific. No precise numbers or targets if someone shows signs of disordered behavior. This kind of harm reduction thinking needs to be baked into system prompts, not handled as an afterthought.
What This Means for Developers
If you’re building with Claude, these changes affect how you should structure your prompts. The model is now biased toward action over clarification. That’s great for straightforward tasks but means you need to be more explicit when you actually want exploration of options rather than immediate execution.
The tool-first approach also matters. If you’re using Claude in an environment with custom tools via the Model Context Protocol, expect it to be more aggressive about calling those tools. Design your tool descriptions carefully because Claude will use them.
I appreciate that Anthropic publishes this stuff, even if it’s incomplete. Other labs treat their system prompts like nuclear launch codes. But the missing piece is still those tool descriptions. Without them, you’re reverse-engineering behavior from observation rather than working from documentation.
The evolution from 4.6 to 4.7 isn’t revolutionary, but it shows Anthropic is paying attention to how people actually use Claude and adjusting the behavior accordingly. That iterative refinement based on real-world friction is exactly how you should be developing AI systems, and it’s refreshing to see it happen in public where we can all learn from it.