Five components turn a vague question into an actionable answer — and the difference between a 5% catch rate and a 95% one.
Asking "as a clinical pharmacist" versus "as a helpful assistant" changes which knowledge the model surfaces, the rigor it applies, and the format it defaults to. Roles are the cheapest way to focus a model — and the most underused.
Same case. Same model. Same H&P. The only thing that changes is the prompt. The output below was generated against our 22-error case using Claude Opus 4.5 — a frontier model — for both runs.
Caught 1 of 22 planted errors. Missed every drug interaction, every guideline gap, every laterality issue. Hedged language throughout.
Caught 21 of 22 planted errors. Cited guidelines. Surfaced the off-checklist drug interaction. Same model.
These are exactly the prompts the exercise on the next page expects you to try. Copy them now, or open this page in a separate tab.
Each one focuses the model through a different professional lens. Run all three sequentially against the same H&P and the catch rate climbs from ~50% to 95%.
Surfaces drug-drug interactions, contraindications given organ dysfunction, dosing errors, and home medications that weren't reconciled at admission.
Act as a clinical pharmacist auditing this admission H&P. Go through every medication, one at a time. For each, check: - Drug-drug interactions with the rest of the list - Contraindications given the patient's renal and hepatic function - Dose appropriate for age, weight, and indication - Home medications reconciled at admission - Any med that should have been started but wasn't Output a numbered list. One row per finding. Cite the medication and the issue.
Catches missed quality measures: stroke center metrics, dysphagia screens, VTE prophylaxis, antithrombotic timing — the things checklists exist to enforce.
Act as a Joint Commission stroke quality auditor reviewing this admission H&P. Check every primary-stroke-center quality measure: - NIHSS documented on arrival - Dysphagia screen before any oral intake - VTE prophylaxis (pharmacologic or mechanical, with rationale) - Antithrombotic therapy by end of hospital day 2 - Statin therapy - Smoking cessation counseling - Stroke education before discharge - Last known well time documented For each measure, state whether it was met, missed, or unclear. Quote the chart text.
Reads the chart against current guidelines. Flags acute-management gaps, secondary-prevention misses, and embolic workup omissions.
Act as a stroke neurologist reviewing this admission H&P against AHA/ASA 2019 and 2021 guidelines. Identify gaps across three domains: (1) Acute management — tPA / thrombectomy candidacy, BP targets, glucose, temperature, NPO status. (2) Secondary prevention — antiplatelet vs. anticoagulation choice, statin intensity, BP and lipid targets, lifestyle counseling. (3) Embolic workup — TTE/TEE, telemetry, hypercoagulable workup if cryptogenic, carotid imaging. For each gap, cite the specific guideline section if you can. Numbered list, deduplicated.
Prompting craft beats model choice. A small model with the right prompt outperforms a flagship model with "find errors."