The model reads your description. That's all it has.

The most counter-intuitive thing about tool use is this: the model doesn't understand your tool by its name or its code. It understands it through the description you wrote. And only through that.

If the description says "Looks up a crewmate", the model has to guess:

• Does it look up by alias? By name? By role?
• Does it return one or many?
• Is it read-only, or does it also modify?
• What happens if it doesn't find anything?

The model gets one of these guesses wrong and the tool ends up misused or ignored. The metric that matters is precision + recall:

• Recall. does the model invoke the tool when it should? (If not, it's hallucinating answers that should come from the tool.)
• Precision. does the model avoid invoking the tool when it shouldn't? (If not, it's calling functions unnecessarily, burning money and latency.)

Both go up when the description is specific and bounded: it says when YES, when NO, and what it guarantees.

When a description says something the model couldn't infer from the name alone, it's adding value. When it repeats what the name already says, it's wasting space.

Pick the correct option on the right.