Last year, Boston Children's Hospital hired me as their first-ever prompt engineer — a job that involves designing prompts for generative AI tools. Put simply, prompts are the requests or commands you type into ChatGPT or other large language models.
Children’s hospitals can use chatbots to enhance their operations, reduce administrative burdens, and augment staff’s work, but it can be difficult to know how to make the most of them. They are relatively new tools built on constantly evolving technology. My experience as an emergency room physician and a developer of successful AI applications has helped me learn how to maximize AI’s potential in health care settings. Here are five tips everyone can use.
1. Be specific
Provide parameters to narrow the model's potential output.
Try this: Instead of saying "Tell me about antibiotic resistance," say "Provide a detailed explanation of the mechanisms of carbapenem resistance in Klebsiella pneumoniae, including the role of carbapenemases, porin mutations, and efflux pumps. Discuss how these mechanisms impact treatment options and infection control measures in hospital settings."
Why it works: This technique provides focused and relevant responses by leveraging AI's natural language processing capabilities to parse detailed instructions. Reducing ambiguity and minimizing the model's reliance on broad generalizations results in more clinically precise information.
2. Ask for an explanation
Instruct the model to articulate the thought process behind its output.
Try this: "Walk me through the process of diagnosing and initially managing a patient presenting with symptoms suggestive of acute coronary syndrome. For each step, explain your reasoning, including what clinical findings, risk factors, and diagnostic tests you would consider, and how each piece of information influences your decision-making."
Why it works: This technique leverages the model's internal reasoning capabilities and guides it through a more structured decision tree. This results in a more transparent and potentially more accurate diagnostic or management approach by:
- Activating relevant knowledge across its neural networks.
- Encouraging a form of “self-attention" where the model reviews its own outputs.
- Reducing the likelihood of inaccuracies by grounding each step in previous reasoning.
3. Provide examples
Provide sample inputs and outputs to fine-tune the model for a specific task.
Try this: Provide an example of a clinical case presentation, then ask for a similar presentation for a different condition.
"Generate a concise clinical case presentation for morning rounds. Format it like this example: Patient: 67-year-old female; Chief Complaint: Shortness of breath; HPI: Gradual onset over three days, worse with exertion; PMH: Hypertension, Type 2 Diabetes; Vitals: BP 150/90, HR 98, RR 24, O2 Sat 92% on RA; Key Exam Findings: Bilateral lower extremity edema, crackles in lung bases; Key Lab/Imaging Results: BNP elevated, CXR shows pulmonary edema; Assessment: Acute decompensated heart failure; Plan: 1) IV lasix, 2) serial troponins, 3) cardiology consult.
Now, create a similar case presentation for a patient with suspected community-acquired pneumonia."
Why it works: The technique leverages the model's ability to recognize patterns and adapt its output style based on examples. This results in more consistent and clinically appropriate responses by:
- Calibrating the model's output format and style.
- Implicitly defining the level of detail or complexity required.
- Demonstrating domain-specific conventions or medical jargon usage.
4. Assign a specific role
Tell the model to provide an output based on a particular perspective.
Try this: "Assume the role of an experienced infectious disease specialist. A 45-year-old patient has returned from a trip to Southeast Asia with fever, muscle aches, and a rash. Walk me through your diagnostic approach, potential differential diagnoses, and initial management plan."
Why it works: This technique leverages the model's ability to adapt its output based on a specified perspective or expertise. Using role assignment prompts allows you to potentially access more specialized knowledge and decision-making frameworks, simulating expert consultation in various medical specialties. This can be especially useful for education, case discussions, or when considering different specialist perspectives on a complex case. This technique works by:
- Encouraging the model to draw upon specialized knowledge relevant to the given role.
- Framing the response in a way that's consistent with the expertise and decision-making process of the specified role.
- Providing insights that mimic the thought process of a specialist.
5. Customize communications
When generating a communication, explain your preferred tone and style.
Try this: When you need the AI to generate text in a specific tone or style, explicitly state it in your prompt. For example:
"Write a professional email to a patient explaining their upcoming cardiac stress test procedure. The tone should be informative yet reassuring."
"Compose an enthusiastic email to the hospital staff announcing a successful Joint Commission accreditation. The tone should be celebratory while maintaining professionalism."
Why this works: This technique takes advantage of the AI's ability to modulate its language based on specified parameters. Whether it's a formal report, patient education material, or an internal memo, this approach ensures that the AI-generated content aligns with the norms and expectations of the communication by:
- Allowing you to tailor the AI's output to suit different audiences and contexts within health care settings.
- Maintaining appropriate communication standards in various professional scenarios.
- Generating more effective and situation-specific content.
