A support agent's system prompt requires calling get_customer to verify identity before process_refund. After rewriting the instruction with capitalized warnings and adding few-shot examples, audits show the order is followed in 99.2% of refunds; the rest are issued against unverified accounts. What should the team do next?

Answer: C — Block process_refund in code until get_customer has returned a verified customer ID in the session

Prompt instructions have a non-zero failure rate, and identity verification before a financial operation requires deterministic compliance. A programmatic prerequisite that blocks process_refund until a verified ID has been returned provides it. More examples (a) and repositioning (b) are still probabilistic, and a self-reported statement (d) checks what the model says, not what get_customer returned.