Most prompt guides for physicians are nearly useless.
They recommend commands such as “summarize this article,” “write a patient explanation,” or “create a differential diagnosis.” Those are tasks, not competent prompts. They provide no clinical context, evidentiary standard, method for managing uncertainty, or protection against plausible-sounding fabrication.
Many physicians now use AI. Few have learned how to instruct it properly. Typing a question into a chatbot is not clinical AI competence.
A useful clinical prompt should define the context, exact task, evidence standard, limits of inference, desired output, and what the AI should do when essential information is missing.
Here are five prompts that are genuinely useful for obstetricians, gynecologists, and other clinicians.
1. Analyze a Clinical Case Without Premature Closure
“Act as a clinical reasoning assistant, not as the treating physician. Analyze the following deidentified case. First summarize the clinical problem in one sentence. Then identify the most likely diagnosis and the dangerous diagnoses that must not be missed. For each possibility, list the findings that support it and those that weaken it. Separate documented facts from assumptions. Identify missing information that could materially change management. Do not invent history, examination findings, laboratory results, or imaging. End with the three most time-sensitive clinical actions and explain why each matters. Case: [INSERT CASE].”
This prompt does more than request a differential diagnosis. It forces the AI to search for dangerous alternatives, identify missing data, and expose assumptions that might otherwise remain hidden.
2. Verify a Clinical Claim or Recommendation
“Evaluate the following claim using current authoritative guidelines and high-quality peer-reviewed evidence: ‘[INSERT CLAIM].’ Classify it as supported, partly supported, unsupported, or contradicted. Distinguish randomized evidence, observational evidence, expert consensus, and biological plausibility. Provide absolute risks when available, not only relative effects. Identify the relevant population, intervention, comparator, outcomes, and major limitations. Do not quote a guideline unless you have verified its exact wording. Provide traceable references with DOI or PMID. Explicitly state when evidence or a citation cannot be verified.”
This prompt is especially useful before repeating a claim in a lecture, manuscript, patient handout, policy statement, or social-media post. It instructs the AI to assess evidence rather than merely find language that supports the proposed conclusion.
3. Prepare an Informed-Consent Discussion
“Create an informed-consent framework for [INSERT CLINICAL DECISION]. Compare all clinically reasonable options, including expectant management or no intervention when applicable. For each option, describe its purpose, probable benefits, material risks, uncertainties, alternatives, and consequences of delay or refusal. Use absolute frequencies with a clearly stated time frame whenever reliable data exist. Separate common minor harms from rare serious harms. Clearly distinguish the clinician’s evidence-based recommendation from preference-sensitive choices. End with five teach-back questions that test understanding without asking, ‘Do you understand?’”
This produces a structure for counseling, not a script to be recited without judgment. Informed consent requires more than listing complications. Patients need to understand what the choices mean for them, including the consequences of postponing or declining recommended care.
4. Critically Appraise a Paper Before Changing Practice
“Critically appraise the attached paper for an ObGyn clinician. Identify the research question, design, population, intervention or exposure, comparator, primary outcome, effect estimates, and funding or conflicts of interest. Assess selection bias, measurement error, confounding, missing data, multiplicity, model specification, causal overstatement, and external validity. Distinguish what the data demonstrate from what the authors claim. Report absolute effects and confidence intervals. Do not rely on the abstract alone. Conclude whether the findings should change practice, justify further research, or currently have no meaningful clinical implication.”
A generic request to “summarize this paper” often reproduces the authors’ framing. Critical appraisal requires the AI to examine whether the methods and results justify that framing.
5. Audit a Manuscript, Guideline, Policy, or Patient Handout
“Audit the following text sentence by sentence. Flag claims that are unsupported, overstated, causally imprecise, ambiguously defined, inconsistent with the cited evidence, or potentially misleading to patients or clinicians. Pay particular attention to words such as ‘safe,’ ‘effective,’ ‘prevents,’ ‘causes,’ ‘recommended,’ and ‘standard of care.’ For every problem, provide: (1) the original claim, (2) the specific defect, (3) the evidence needed to support it, and (4) a more defensible revision. Verify every citation that materially supports a clinical claim. Do not soften valid criticism merely to make the language more agreeable. Text: [INSERT TEXT].”
This is far more useful than asking AI to “improve” a document. Improvement can mean making unsupported claims sound more polished and persuasive. An audit asks whether the claims are actually defensible.
Better Prompts Produce More Auditable Answers
These prompts require the AI to distinguish evidence from inference, disclose missing information, examine competing explanations, and admit when evidence cannot be verified. That makes its output easier for a clinician to inspect and challenge.
The objective is not to make AI sound more authoritative. It is to make its work more auditable.
Even an excellent prompt cannot guarantee a reliable answer. AI may fabricate references, misquote guidelines, overlook contraindications, or misunderstand patient-specific details. Protected health information should never be entered into a system that has not been approved for that purpose. Every clinically material output still requires professional verification.
But “AI can make mistakes” is not an excuse for physicians to remain incompetent in its use. Physicians also make mistakes, particularly when working from memory, under time pressure, or outside their narrow expertise. The responsible response is disciplined use: precise instructions, explicit evidence standards, independent verification, and professional accountability.
Prompting is not a clerical trick. In medicine, it is the translation of a clinical problem into instructions precise enough to produce an answer that can be examined, challenged, and verified.
That is now part of clinical AI competence.