Group of 2 Under the hood
Claude now has explicit guidance on how to safely learn a connector tool's argument and result shapes before publishing a page that uses it
What
A new block of guidance text instructs Claude on how to handle connector (MCP) tools when publishing something like a Claude Doc or Artifact page:
- Take argument names for a loaded connector tool from the tool's own input schema, not from guesswork.
- Only learn a tool's result shape by making a safe, real call to it, never by inferring it from a write call.
- Use
describeTool(server, tool)when the published page needs to read a tool's schema at view time, since the authoring session cannot see that information before publishing. - Disclose to the user, rather than embedding in the page itself, whenever a tool's shape had to be guessed.
Why
This reduces the chance of a published page calling a connector tool with wrong argument names or misinterpreting its results, since the model is directed to verify tool shapes safely instead of guessing.