Tips for writing a good user query

1. Stick to one core idea. Give the LLM a single, clear subject: “falls from hospital beds,” “carbon-monoxide poisoning at home.” In general, the shorter the prompt, the less room for misinterpretation.
2. Avoid nested logic. Complex clauses like “suicide and medication error but not in custody” dilute the signal. Consider running separate screens (suicide; medication error; in custody) and combine or subtract results later with pandas.
3. Let the model handle synonyms. You don’t need “defective, faulty, malfunctioning” all in the same query; “malfunctioning defibrillators” is enough.
4. Use positive phrasing. Negations (e.g. “not related to COVID-19”) can flip the model’s reasoning. Screen positively, set filter_df to False, then drop rows in pandas.
5. Keep it readable. If your query needs multiple commas or parentheses, break it up. A one-line statement without side notes usually performs best.
6. Use limiting words to restrict scope. Words like “only" can focus the LLM on a specific case, reducing the chance it will infer extra details. For example, “falls from beds **only**” signals to the model not to include corridor or bathroom falls. However, avoid overusing them; too many limiters can make the model miss relevant edge cases.

Examples:

Less-effective query	Why it struggles	Better query
“Deaths where someone slipped or fell in hospital corridors or patient rooms and maybe had fractures but not clinics”	Too long, multiple settings, negative clause	“Falls on inpatient wards”
“Fires or explosions causing death at home including gas leaks but not industrial accidents”	Mixes two ideas (home vs. industrial) plus a negation	“Domestic gas explosions”
“Cases involving children and allergic reactions to nuts during school outings”	Several concepts (age, allergen, setting)	“Fatal nut allergy on school trip”
“Railway incidents that resulted in death due to being hit by train while trespassing or at crossings”	Two scenarios joined by “or”; verbose	“Trespasser struck by train”
“Patients dying because an ambulance was late or there was delay in emergency services arrival or they couldn't get one”	Chain of synonyms and clauses	“Death from delayed ambulance”
“Errors in giving anaesthesia, like too much anaesthetic, wrong drug, problems with intubation, etc.”	Long list invites confusion; “etc.” is vague	“Anaesthesia error”