A water-quality advisor must not invent medical advice when residents ask about illness after a boil-water notice. What prompt redesign helps?

Redesigning the prompt to refuse speculative medical claims and defer to published guidance is the correct prompt control for a water-quality advisor after boil-water notices. Residents often ask whether symptoms mean they need a doctor; the civic bot's job is to restate official boil-water instructions and direct people to public-health sources—not to diagnose. Explicit refusal patterns plus mandatory pointers to published notices reduce harmful improvisation while still answering operational questions (boil duration, affected zones, safe uses of water).

Grounding the prompt with the FAQ and requiring citations to published guidance sections fails because citations alone improve traceability but do not add a refusal rule, so the model can still append speculative medical claims around cited text instead of refusing such claims and deferring strictly to published public-health guidance. Adding few-shot examples of FAQ answers to standardize tone and citation formatting style fails because style examples only teach formatting consistency, not safety boundaries, so the model learns tone without learning to shut down speculative medical talk and defer to published guidance. Lowering the temperature setting to reduce variability in grounded FAQ response wording slightly fails because it only makes wording slightly more stable and does not create refusal behavior; the prompt must still refuse such claims outright and defer to published guidance.

Exam caveat: Refusal-and-defer prompts must stay synchronized with the current notice text and jurisdiction; stale prompts can contradict an updated boil-water order. Pair prompt redesign with content filters and evaluation cases that try to elicit medical advice.

Operational check: Add adversarial eval items that ask for diagnoses or home remedies after a boil-water notice; require refusal plus citation of the published guidance; confirm retrieval returns the active notice; review production transcripts for leaked medical speculation; and involve public-health SMEs in prompt approval.