Designing tool schemas — CCA-F Exam Prep

×
L2.060
REAL STORYSplit screen. Left side: a clean, well-labeled tool schema with clear descriptions and typed parameters. Claude uses it perfectly, green checkmarks. Right side: a vague, poorly named schema with missing descriptions. Claude guesses wrong, red X marks everywhere. Developer pulling his hair out.

Same tool. Same model. One schema worked. The other failed 40% of the time.

A fintech team built two tools: get_balance and transfer_funds. Both worked perfectly in unit tests. But in production, Claude kept calling get_balance when users asked to send money. 40% error rate.

The fix took 15 minutes. They didn't change the code. They didn't retrain anything. They rewrote the tool descriptions.