A Prompt Improved Twice

In practice

The first attempt is a single line: summarize these requests. The output is readable but unusable: it mixes urgent leaks with paint touch-ups, names no buildings, and changes shape every week. Before changing anything, the coordinator writes her criteria: every request listed with building and room, sorted by urgency, safety issues flagged, and nothing invented. She keeps three past weeks of requests as test cases, one with a request that gives no room number. Round one adds the elements. The instruction comes first and says to extract and sort; the context says the reader is the maintenance supervisor; the requests are pasted below a separator; and the output format is a table with building, room, issue, urgency and a safety flag. Two test weeks now pass. The third fails: for the request with no room number, the model fills in a plausible one. Round two adds the method's grounding: use only what the requests state, and where a detail is missing, write not stated and add it to a short list of questions for the requester. She changes only that, reruns all three weeks, and all three now pass. She saves the prompt in the team library with its test cases and a note on why the not-stated rule is there.