Tools

Research prompts

A prompt per tool that makes the agent verify its own configuration against the API.

A configuration snippet only tells you what to write. These prompts tell you whether it worked.

Paste one into the tool itself, in the repository where you have configured it. The agent will look for the credential, send a real request, try streaming, and report the exact error if anything fails instead of guessing. Both languages say the same thing.

The prompts are written to keep your key out of the transcript: the agent must not print it, and must not write it into a file that git tracks. If an agent ever echoes your key, revoke it and create a new one.

What each prompt checks

  1. The credential is present, and where it was found. The key is never printed.
  2. A non-streaming request works, with the exact command and the model's reply.
  3. Streaming works, meaning events arrive as they are generated rather than in one buffered body.
  4. Failures are reported precisely, with the status code and the request that produced it, and no more than two fix attempts before it stops and asks.

Tools that need a bridge, and tools with a request-shape or tool-calling trap, get one extra step that checks exactly that. Cursor gets a different prompt on purpose: it cannot use a third-party endpoint, so its prompt audits the repository instead, which is the useful thing an agent can do there.