A team adds a tool with a vague description and an unrestricted free-text argument. The model frequently chooses it incorrectly. Which improvement is most useful?

1. Accept all generated arguments without checking
2. Define a narrow purpose, typed arguments, and clear use conditions
3. Hide successful tool responses from the model
4. Rename every tool to the same short name