Research prompts: fix the question and eligibility first

Last updated: 7 Oct 202611 min read

tutorial

An evidence-synthesis prompt needs a question and a selection contract before the model sees attractive results. Name the population, change being tested, comparator, measured outcome, observation window, acceptable study designs, and exclusion rules. Fix which report versions and search snapshot are in scope. The model can apply the rules to candidate records and flag ambiguous cases, but a reviewer owns final eligibility decisions. Save excluded records with reasons. Without this ledger, a polished summary may select only studies that support the desired answer. A narrative scan of a few supplied reports is not a comprehensive review unless its search and screening process actually supports that description.

Operational case

A fictional parcel operator asks whether a scan-confirmation panel lowers misrouted parcels. The target population is outbound parcels at staffed sorting lines, the comparator is the existing panel, the outcome is misroutes per allocated parcel, and follow-up ends after a full shift. Before seeing results, the editor includes controlled line trials and excludes simulator exercises and before-after reports from the primary comparison. Trial ST-47 and trial ST-88 qualify. Observational report ST-61 is retained in a separate context bucket because route mix changed between shifts. An assistant that drops ST-88 after finding a worse candidate result violates the contract even if its remaining paragraph is accurate.

Question: staffed sorting lines; new scan panel vs existing panel.
Primary outcome: misrouted parcels / allocated parcels during full shift.
Primary set: controlled line trials ST-47 and ST-88.
Context only: ST-61, shift comparison with changing route mix.
Keep an exclusion reason and report version for every screened record.

Performance and operating cost

Screening S candidate records is O(S) decisions once metadata is available; full-text review can dominate time when outcomes or design are unclear. A model may reduce the first-pass workload, but every uncertain exclusion needs review because an error can remove a contrary finding. Store the fixed rule and case-level decisions separately from prose. Repeating the same search later may return a different corpus, so use a dated snapshot and identifiers. The prompt should ask for missing information, not fill absent design details from plausibility.

Common Mistakes

Do not write eligibility rules after reading which study wins.
Do not silently combine controlled trials with changed-shift observations.
Do not call a supplied packet a complete literature review without a reproducible search.

Connected lessons

Continue with: Interview prompts: limit transcript use to the study purpose.

Continue with: Survey prompts: define the population and invitation frame.

prompt engineering

research synthesis