Prompt style effects on large language model differential diagnosis output in adolescent catatonia and psychosis

Prompt engineering shapes large language model (LLM) output, but practicing clinicians have little guidance on which prompting style suits a given clinical task. A single publicly available, de-identified pediatric psychiatric case report was submitted to a base, unauthenticated ChatGPT-5.5 session under four prompt conditions: zero-shot, chain-of-thought, role/persona, and structured/constrained output. Each output was scored from 0 (feature absent) to 2 (complete) on four criteria: diagnostic accuracy, causative agent recognition, safety completeness, and calibration. Output was compared against the published clinical course. Total scores were zero-shot 7/8, chain-of-thought 6/8, role-persona 7/8, and structured 7/8. All four conditions ranked bipolar I disorder above substance-induced psychotic disorder, whereas the treating team's documented differential favored unspecified and substance-induced categories without a confirmed bipolar diagnosis. All four conditions recommended neuroimaging, electroencephalography, and lumbar puncture, exceeding the workup the treating team actually performed. Across four prompt styles, core diagnostic content was materially the same; the styles differed in the precision of probability quantification and in reasoning transparency, not in the underlying differential generated.