Tool schemas consume context on every single turn, and selection accuracy falls as the surface grows. Build the problem, measure it, then fix it structurally rather than by trimming descriptions.
What to build
- Register thirty tools across four unrelated domains.
- Measure the token cost of the schemas alone, before any conversation.
- Run a request set and measure selection accuracy.
- Now restructure: group the tools behind a routing layer, or scope tool sets per subagent.
- Re-measure both schema cost and selection accuracy.
- Use tool_choice to force a specific tool where the sequence is known in advance.
Done when
- You have real numbers for schema token cost before and after.
- Selection accuracy improved after restructuring.
- At least one path uses tool_choice to remove a decision the model should not be making.
- You can explain why deleting descriptions is the wrong way to cut schema cost.
The trap this exercise teaches
Treating tool schemas as free. They are resent every turn and they compete with the conversation for the same window.
Before you start
The theory behind this build is covered in Scaling the tool surface. If any step below is unfamiliar, read that first — the exercise assumes it. Primary source: Claude Docs — Tool use.
Check yourself against the exam
This exercise sits in Domain 4, which is 18% of the CCAR‑F exam. Once you have built it, run a domain drill in the study app and see whether the questions read differently.