Prompting Strategies for Large Language Models in Primary Care: A Primer for Clinician-Artificial Intelligence Interaction

Abstract

Primary care is a language-intensive environment where clinicians frequently synthesize, document, and communicate large amounts of information. Given the amount of information in primary care, large language models (LLMs) have emerged as tools to support primary care clinician workflows, aiming to reduce administrative burden and augment patient care. LLMs are artificial intelligence tools that can generate text and assist with chart summarization, message drafting, and clinical decision support. However, because LLMs are probabilistic, identical inputs may yield inconsistent responses. The LLMs output can be significantly improved through prompt engineering. Prompt engineering is the intentional design of text entered into the LLM to help produce better LLM results. Prompting is best approached as an iterative process where a clinician defines the task, the LLM generates output, and the clinician critically reviews and refines the results. Low-quality prompts are often vague and produce generic or misleading responses. High-quality prompts are specific, context-rich, and well-structured. Using structured prompting frameworks can improve the LLMs response. Including core elements, such as the LLMs role, the clinical context, the requested task, and the desired output formatting, can help improve the LLMs response by providing situational background and framing response formatting. Additional frameworks, such as Ask-Context-Expectation and PICO+O, can be utilized for specific patient circumstances and cases. However, despite their promise, LLMs have important limitations. They may generate inaccurate or hallucinated information and produce inconsistent results. Given these limitations, LLMs should not be viewed as knowledge authorities but rather as cognitive assistants. Thoughtful prompt design can help reduce these limitations and improve response accuracy, always keeping in mind that LLMs should not replace human judgment or clinical reasoning. As LLMs continue to be integrated into primary care, prompt engineering will become an increasingly important clinical skill. Careful prompt design and critical appraisal of LLM responses will be key to maintaining excellent patient care in the AI era.

Keywords

large language model primary care prompt engineering clinical workflow support artificial intelligence in primary care

Introduction

Primary care is a high-volume, language-intensive environment. Face-to-face conversations with patients, documentation, referral communication, and electronic message management are all language-centric aspects in primary care workflows.^1-4 Over the past decade, artificial intelligence (AI) has been rapidly evolving, and AI systems can now process, interpret, and generate human language. Specifically, large language models (LLMs) are AI systems that are capable of predicting and generating text by recognizing patterns learned from large bodies of written language.^5,6 LLMs are a type of AI that is trained to generate human-like language by predicting the most likely next word or sequence of words. It is not a search engine, an encyclopedia, a clinician, or a source of guaranteed truth.^7,8

Frequently used LLMs in primary care may include OpenEvidence, ChatGPT by OpenAI, Microsoft Copilot, Google Gemini, or Claude by Anthropic. Although LLMs can produce fluent, highly useful outputs, they do not understand contextual clues the way humans do, nor do they consistently distinguish between facts and fiction. LLMs perform well when summarizing information, classifying content, drafting responses, explaining concepts at different reading levels, and converting information from one format to another. More advanced models can also perform reasoning tasks. However, these tools do not reason reliably and may perform poorly in ambiguous situations. As such, they can generate incorrect information with high confidence.^9,10

When working with an LLM, the LLM should be viewed as a cognitive assistant rather than a knowledge authority. In practice, LLMs can assist primary care clinicians in three broad ways.^11-14

The first is clerical support, which includes tasks such as drafting, summarizing, or formatting content. In primary care, this might include summarizing a lengthy hospital discharge into a more organized note format.^11,12,15

The second is cognitive support, which includes brainstorming, helping structure diagnostic or management reasoning, and generating questions or recommendations for further review. For example, a primary care clinician might use an LLM to help generate a differential diagnosis for a patient with a chronic cough or to identify follow-up questions for a complex symptom presentation.^11,12,16

The third is communication support, which focuses on audience-facing tasks such as drafting patient messages, preparing letters to specialists or insurance companies, and translating medical information into plain language. A common example in primary care would be using an LLM to create a patient-friendly explanation of laboratory results at an appropriate reading level.^11,12,17-19

Because LLMs are probabilistic, rather than systematic search tools with consistent output, identical inputs may not produce identical, consistent responses. Clear, structured input can reduce ambiguity and increase the usefulness and consistency of the LLMs output. The text entered by a user into an LLM is called a prompt, and the process of designing effective prompts to guide LLM behavior is referred to as prompt engineering.^20-22

Regardless of the use case, prompt engineering principles are versatile and can be applied to many situations. Prompt engineering does not require technical expertise, but it is a skill to learn and has best practices to follow. For clinicians, it represents a practical skill that can improve the quality, safety, and efficiency of clinician-AI interaction.

Prompt Engineering Overview

Prompt engineering can be defined as the process of intentionally designing and refining a user’s input to better guide the LLMs output and improve its accuracy and usefulness for specific tasks. Prompt engineering is similar to refining a search strategy in PubMed or internet search engines. A better prompt usually leads to better output.²³

Clinician-AI interaction is most effective when approached as a structured, iterative process. First, the clinician defines the goal and provides the input. Second, the LLM generates a draft output. Third, the clinician reviews, verifies, edits, and determines whether the output is appropriate for use. This cycle can then be repeated as needed to refine the result.

In general, vague prompts produce vague answers. When a prompt isn’t specific, the LLM may fill in missing details or generate responses that sound plausible, but do not address the question. Specific, context-rich prompts, termed “high signal” prompts, are better than long, bloated ones. A high-signal prompt should include only the information that improves the response. Extra wording, repeated instructions, and unnecessary background can reduce the prompt’s efficiency. For example, a low-signal prompt would be, “Can you please help me think through what might be going on with this patient who has concerning chest pain?” while a high-signal prompt might be, “Differential diagnosis for acute chest pain in a 65-year-old with HTN and smoking history. Prioritize life-threatening causes.” Additionally, effective prompting is often a dialogue rather than a one-time, zero-shot request.

Zero-shot prompting is asking the model to complete a task without providing examples or additional details besides the initial instruction. Although zero-shot prompting can be useful for simple tasks, more complex clinical situations benefit from iterative interaction. Clinicians can improve output by asking the LLM to refine an initial prompt or identifying if information is missing. Similarly, the model can be instructed to ask clarifying questions before answering, which may help reduce ambiguity and better define the task.^24-26

Prompting can also be used to strengthen the LLMs reasoning by asking the LLM to identify weaknesses, challenge assumptions, present opposing viewpoints, or play devil’s advocate. Additionally, many LLMs can share step-wise reasoning in a “chain-of-thought.” Chain-of-thought is where the LLM explains its reasoning process as part of its output, or during the “thinking process.” Utilizing chain-of-thought prompting can help the clinician identify gaps or weaknesses in the LLMs reasoning. For example, instead of asking, “What is the best diagnosis?” a clinician might ask, “What is the most likely diagnosis, and explain the steps and reasons that support this diagnosis.” This strategy encourages a more structured response and makes the LLMs reasoning process easier to inspect and appraise.^27,28

General Prompting Framework

In addition to zero-shot and chain-of-thought prompting, structured frameworks can improve LLM output. Although many prompting frameworks exist, four elements are common: Role, Context, Task, and Output.^29-31

First, Role tells the LLM who it should act as. Assigning a role tells the LLM the function it should assume when generating output. This also helps shape tone, reading level, and amount of detail. It allows the response to better align with the intended audience and task. For example, a prompt such as “Act as a primary care physician writing a secure electronic message to a patient with new lower back pain” will likely produce a different result than “Act as a patient educator explaining lower back pain at a fourth-grade reading level.”

Second, Context provides relevant information related to the task. This allows LLM to understand the situation and circumstances where the output will be used. Without context, outputs tend to be vague, generic, and less relevant to the specific situation. Relevant context may include patient characteristics such as age, comorbidities, clinical setting, and pertinent background information.

Third, Task outlines what the LLM is being asked to do. Tasks should be as specific as possible. Specifying if the task is clerical, cognitive, or communicative helps better frame the output. Clinicians should avoid vague prompts such as “help me with this patient” or “write something about these labs.” More concrete requests tend to yield better output. Strong verbs such as summarize, compare, rewrite, draft, classify, prioritize, explain, identify, and list help reduce the risk of unfocused output.

Lastly, Output defines what the final product should look like. This could include bullet points, providing the output in a table, a SOAP-formatted note, or writing a letter. Output also includes limits on response length, details on the reading level and audience, and formatting. For example, “Provide the response in 3 bullet points written at a fourth-grade reading level,” or “Draft a concise secure electronic message in under 150 words, including return precautions.” Output also includes guardrails, or constraints, on the LLMs response. Guardrails help constrain response for the intended use. In practice, they can direct the model to state its uncertainty, avoid unsupported claims, and limit citations to trusted sources, although clinician verification remains necessary.

For primary care clinicians, these elements provide starting points for better clinician-AI interaction across a variety of clinical use-cases. Prompt engineering frameworks can also be task-specific and tailored to the use case. However, many task-specific prompting frameworks still rely on some variation of these four elements.

Task Specific Frameworks for Primary Care: A Toolbox for LLM Use

There are several prompting frameworks well-suited for primary care. Often, the best approach depends on the task’s complexity and the clinician’s goal (Table 1).

Table 1.

Prompting Strategies and Examples

Prompting strategy	Brief description and example
High-signal prompting	Provide only relevant details to sharpen the response.
High-signal prompting	Example: “List life-threatening causes of acute chest pain in a 65-year-old with HTN and smoking history.”
Zero-shot prompting	Uses one instruction without examples or follow-up.
Zero-shot prompting	Example: “Summarize this discharge note in 3 bullet points for a primary care follow-up visit.”
Iterative prompting	Back-and-forth revision to improve the response.
Iterative prompting	Example: “Revise a secure electronic message to be shorter, more reassuring, and written at a fourth-grade reading level.”
Chain-of-thought prompting	Stepwise reasoning to make the response easier to inspect.
Chain-of-thought prompting	Example: “List the key findings, most likely diagnoses, evidence for and against each, and the next best test.”
Meta-prompting	Uses the LLM to improve the prompt before answering.
Meta-prompting	Example: “Rewrite this prompt to be clearer and more patient-friendly, and ask what information is missing.”
Role-Context-Task-Output	Role, Context, Task, and Output to structure the request.
Role-Context-Task-Output	Example: “Act as a primary care physician. This is a 58-year-old with HTN and new ankle swelling. Draft a secure electronic message explaining possible causes, next steps, and when to seek urgent care. Keep it under 150 words, use plain language, and include return precautions.”
ACE (Ask-Context-Expectation)	Ask, Context, and Expectation to define the task, background, and format.
	Best for straightforward communication or documentation tasks in which the clinician already knows the goal and wants to specify tone, context, and formatting.
	Example: “Draft a secure electronic message explaining mildly elevated TSH in a patient with fatigue; keep it under 120 words, use plain language, and sound reassuring but honest.”
CRIT (Context-Role-Interview-Task)	Context, Role, Interview, and Task for complex, dialog-based prompts.
	Best used for more complex or uncertain situations in which clarifying questions and iteration would meaningfully change the response.
	Example: “Context: 72-year-old with diabetes and CKD. Role: experienced PCP. Interview: ask up to 4 clarifying questions. Task: compare SGLT2 inhibitor vs GLP-1 RA.”
BRAIN (Background-Role-Aim-Instructions-Next Steps)	Background, Role, Aim, Instructions, and Next Steps for multi-step tasks.
	Best used for multi-step clinical tasks that require more extensive background context, explicit instructions, and follow-up refinement or next steps.
	Example: “Background: diabetes, CKD, HTN. Role: PCP. Aim: compare SGLT2 vs GLP-1. Instructions: use guideline-based reasoning and note safety issues. Next steps: ask 3 follow-up questions.”
PICO+O (Patient-Intervention-Comparison-Outcome + Output)	Patient, Intervention, Comparison, Outcome, and Output to compare options or summarize evidence.
	Best used for evidence-based medicine applications, when the clinician wants the LLM to compare options, summarize evidence, or organize evidence-based reasoning for a clinical question.
	Example: “Patient: adults with type 2 diabetes and ASCVD. Intervention: GLP-1 RA. Comparison: SGLT2 inhibitor. Outcome: CV benefit, renal benefit, adverse effects. Output: 4-row table with 2 clarifying questions.”

One useful framework for simple primary care tasks is Ask-Context-Expectation (ACE), where ask defines the task, context provides the background and setting, and expectations specify the boundaries, tone, format, or style. This framework could be useful in daily primary care tasks such as medication counseling, composing secure electronic message, or refining discharge summaries. For example: Ask: Write a secure electronic message to the patient explaining thyroid results and next steps. Context: 45-year-old with TSH of 6.0 and normal free T4 (possible subclinical hypothyroidism). Has fatigue and is worried. Expectation: <150 words, plain language (4th grade level), include when to follow up.

A framework for more complex situations is Context-Role-Interview-Task (CRIT). This prompting approach defines the context in which the LLM is working, the LLMs role, and the tasks that need to be completed. However, it also adds an interview component, where the LLM asks clarifying questions to better understand the request. Because it encourages dialogue, it also works well in situations where the clinician may be uncertain what information to include in the prompt. For example: Context: Adult patient with fatigue and anemia. Hemoglobin is 10.2 g/dL, MCV is 74 fL, ferritin is 9 ng/mL, iron saturation is 8%, and TIBC is elevated. Role: Experienced primary care physician. Interview: Ask 4 key clarifying questions, one at a time, that would help interpret the anemia. Task: Interpret the most likely cause of the anemia, identify other possible causes, and draft a patient-friendly explanation of the lab results and next steps at a fourth-grade reading level.³²

Another useful framework for primary care is the Background-Role-Aim-Instructions-Next Steps (BRAIN) framework .³⁰ This approach is better for complex, multi-step tasks because it provides more context and incorporates interactive elements of working with LLMs, similar to CRIT. BRAIN would be a good option for drafting care plans, comparing management options, or synthesizing guideline recommendations. For example: Background: 72-year-old male with type 2 diabetes, CKD stage 3a, hypertension, and a recent rise in creatinine (1.6→2.3 over 3 months) while taking metformin. Role: Experienced primary care physician. Aim: Interpret the worsening kidney function and organize the likely contributors. Instructions: Use current guideline-based reasoning, note key safety considerations, and identify areas where clinician judgment is still needed. Next steps: Ask me 3 follow-up questions that would help refine the interpretation and guide the next steps.

For more patient-specific clinical questions, PICO+O can also be a useful framework for prompting. This builds on the evidence-based medicine structure of Patient, Intervention, Comparison, and Outcome, with the +O for desired LLM output. Output includes the desired format, limitations, constraints, and options for iteration. This framework is useful when asking the LLM to compare management strategies or summarize evidence.³³ For example: Patient: adults with type 2 diabetes and established cardiovascular disease. Intervention: GLP-1 receptor agonist. Comparison: SGLT2 inhibitor. Outcome: cardiovascular benefit, renal benefit, adverse effects, and practical prescribing considerations. Output: summarize in a 4-row table for a primary care audience, include key limitations, and end with 2 clarifying questions that would improve the recommendation.

Lastly, meta-prompting is a strategy to improve prompts using LLMs to refine the prompt.³¹ Instead of asking the model to simply answer a question, the clinician first asks the LLM to write a better prompt. For example, instead of “help me explain these lab results to a patient,” a clinician might use meta-prompting and ask, “Help me write/rewrite a prompt so it is clearer and produces patient-friendly output. Ask me what information is missing.”

Common Prompting Mistakes Clinicians Should Avoid

There are several pitfalls that clinicians should be aware of when prompting LLMs. Although clinicians may feel the time constraints of clinical practice, investing a small amount of effort in crafting a clear prompt can save time by generating more relevant and usable responses.

Good prompts can be thought of as an AI equivalent of SmartPhrases or order sets in the electronic health record. Once a prompt has been refined and shown to consistently produce useful output, it can be saved and reused as a template for future tasks. In this way, saved prompts may improve efficiency and support more consistent AI-assisted workflows.

Higher-stakes use cases, particularly those involving cognitive support for diagnostic or management reasoning, also require greater scrutiny. In these higher-stakes situations, LLM output should be treated as a rough draft rather than a reliable clinical recommendation. Fluent wording can mask flawed clinical reasoning. Clinicians should have an attitude of “enlightened skepticism” when using the output.

As previously noted, prompting often yields the best results when it’s an iterative process. Although there is a role for zero-shot prompting for quick or simple tasks, iteration and dialogue help better frame the context and task. This reduces the likelihood that the model fabricates information and makes the output more specific to the use case.³¹

Furthermore, since models may hallucinate and fabricate citations, citations should be verified. Additionally, clinicians should be aware that LLMs may still hallucinate citations even when prompted to use only accurate and reliable sources. In high-stakes situations where clinicians are less familiar with the literature, LLM output should be verified against trusted, reliable sources.^34-36

Conclusion

For primary care clinicians, prompting should be viewed as a new clinical communication skill. Unlike a traditional search engine query, which is typically asked once and answered once, interaction with an LLM is often most effective when approached as an iterative dialogue in which the prompt and output are progressively refined. Clear, specific prompts improve the LLMs output while vague prompts are more likely to produce vague, generic, or less relevant answers. Prompting frameworks can help clinicians structure their input by clearly defining key elements, such as the LLMs role, the clinical context, the task, and the desired output format. In high-stakes scenarios, LLMs should not be treated as the definitive authority, and outputs should be verified with trusted sources. As LLMs become more ingrained in clinical workflows, prompting will become an increasingly important skill in primary care. Clinicians should prompt thoughtfully, review outputs critically, and refine them iteratively. Thoughtfully prompting and critically reviewing the output will help clinicians provide safe and effective care well into the AI era.

Footnotes

ORCID iD

Ivana T. Croghan

Ethical Considerations

This educational project did not require approval from the institutional review board. All authors assert that all procedures contributing to this work comply with the ethical standards of the Mayo Clinic.

Author Contributions

All the authors participated in the study concept and design, collection, analysis and interpretation of data and drafting and revising the paper for important intellectual content and have seen and approved the final version of the manuscript. Each author will take public responsibility for the entire work.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This project was supported in part by the Mayo Clinic Division of General Internal Medicine.

Declaration of Conflicting Interests

The authors declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: R. T. Hurt is a consultant for Nestlé Nutrition for research activities unrelated to the content of this paper. All other authors declare no support from any organization for the submitted work; no financial relationships with any organizations that might have an interest in the submitted work in the previous 3 years; and no other relationships or activities that could appear to have influenced the submitted work.

Data Availability Statement

All data supporting the study findings are contained within this manuscript.*

References

Arndt

Beasley

Watkinson

, et al. Tethered to the EHR: primary care physician workload assessment using EHR event log data and time-motion observations. The Annals of Family Medicine. 2017;15(5):419-426.

Arndt

Micek

Rule

Shafer

Baltus

Sinsky

. More tethered to the EHR: EHR workload trends among academic primary care physicians, 2019-2023. The Annals of Family Medicine. 2024;22(1):12-18.

Nguyen

Jenkins

Khanna

, et al. A systematic review of contributing factors of and solutions to electronic health record–related impacts on physician well-being. Journal of the American Medical Informatics Association. 2021;28(5):974-984.

Kuhn

Basch

Barr

Yackel

Physicians* MICotACo . Clinical documentation in the 21st century: executive summary of a policy position paper from the American College of Physicians. Annals of internal medicine. 2015;162(4):301-303.

Brown

Mann

Ryder

, et al. Language models are few-shot learners. Advances in neural information processing systems. 2020;33:1877-1901.

Singhal

Azizi

, et al. Large language models encode clinical knowledge. Nature. 2023;620(7972):172-180.

Harrer

. Attention is not all you need: the complicated case of ethically using large language models in healthcare and medicine. EBioMedicine. 2023;90:104512.

Rogers

Anna

Luccioni

Alexandra

. Position: Key claims in llm research have a long tail of footnotes. Proceedings of Machine Learning Research. arXiv preprint arXiv:230807120. 2023.

Kim

Podlasek

Shidara

Liu

Alaa

Bernardo

. Limitations of large language models in clinical problem-solving arising from inflexible reasoning. Scientific Reports. 2025;15(1):39426. doi:10.1038/s41598-025-22940-0.

10.

McCoy

Swamy

Sagar

, et al. Assessment of large language models in clinical reasoning: a novel benchmarking study. NEJM AI. 2025;2(10):AIdbp2500120.

11.

Maity

Saikia

. Large Language Models in Healthcare and Medical Applications: A Review. Bioengineering (Basel). 2025;12(6):631. doi:10.3390/bioengineering12060631.

12.

Shah

Sharma

, et al. Large language model prompting techniques for advancement in clinical medicine. Journal of Clinical Medicine. 2024;13(17):5101.

13.

Spotnitz

Idnay

Gordon

, et al. A survey of clinicians' views of the utility of large language models. Applied Clinical Informatics. 2024;15(02):306-312.

14.

Agrawal

Chen

Gulamali

Joshi

. The evaluation illusion of large language models in medicine. npj Digital Medicine. 2025;8(1):600.

15.

Zhou

Lyu

Wang

. Large language models-powered clinical decision support: enhancing or replacing human expertise? Elsevier; 2025:1-4.

16.

Gaber

Shaik

Allega

, et al. Evaluating large language model workflows in clinical decision support for triage and referral and diagnosis. npj Digital Medicine. 2025;8(1):263. doi:10.1038/s41746-025-01684-1.

17.

Tripathi

Sukumaran

Cook

. Efficient healthcare with large language models: optimizing clinical workflow and enhancing patient care. Journal of the American Medical Informatics Association. 2024;31(6):1436-1440.

18.

Chen

Guevara

Moningi

, et al. The effect of using a large language model to respond to patient messages. The Lancet Digital Health. 2024;6(6):e379-e381.

19.

Garcia

Shah

, et al. Artificial intelligence–generated draft replies to patient inbox messages. JAMA Network Open. 2024;7(3):e243201.

20.

Busch

Hoffmann

Rueger

, et al. Current applications and challenges in large language models for patient care: a systematic review. Communications Medicine. 2025;5(1):26.

21.

Tan

Gunasekeran

Low

, et al. Regulation of clinical Artificial Intelligence (AI) in the Age of Agents: Unconfined Non-Deterministic Clinical Software (UNDCS) systems for healthcare. npj Digital Medicine. 2026;9(1):186. doi:10.1038/s41746-026-02420-z.

22.

Chen

Zhang

Langrené

Zhu

. Unleashing the potential of prompt engineering for large language models. Patterns. 2025;6(6):101260. doi: 10.1016/j.patter.2025.101260.

23.

Wang

Chen

Deng

, et al. Prompt engineering in consistency and reliability with the evidence-based guideline for LLMs. npj Digital Medicine. 2024;7(1):41. doi:10.1038/s41746-024-01029-4.

24.

Wang

Shi

, et al. Prompt engineering for healthcare: Methodologies and applications. Meta-Radiology. 2025;4:100190.

25.

Sivarajkumar

Wang

. HealthPrompt: a zero-shot learning paradigm for clinical natural language processing. AMIA Annual Symposium Proceedings; 2023:972.

26.

Daulat

Dholaria

Burnet

, et al. Prompt Engineering and Follow-Up Questioning Improves the Readability of Spine Surgery Questions in Large Language Models. World Neurosurg. 2025;203:124423. doi:10.1016/j.wneu.2025.124423.

27.

Wei

Wang

Schuurmans

, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems. 2022;35:24824-24837.

28.

Yao

Zhao

, et al. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems. 2023;36:11809-11822.

29.

Dair.Ai . Elements of a Prompt. Prompt Engineering Guide; 2026.

30.

Callens

. Effective prompt design for large language models in clinical practice. Acta Clin Belg. 2026;81:1-12. doi:10.1080/17843286.2026.2613903.

31.

Liu

Wang

Liu

. Prompt Engineering in Clinical Practice: Tutorial for Clinicians. J Med Internet Res. 2025;27:e72644. doi:10.2196/72644.

32.

Woods

. The AI-Driven Leader. AI Thought Leadership; 2024.

33.

Deng

Sun

, et al. Benchmarking large language models in evidence-based medicine. IEEE journal of biomedical and health informatics. 2024;29(9):6143-6156.

34.

Chelli

Descamps

Lavoué

, et al. Hallucination rates and reference accuracy of ChatGPT and bard for systematic reviews: comparative analysis. Journal of medical Internet research. 2024;26(1):e53164.

35.

Sanchez-Ramos

Lin

Romero

. Beware of references when using ChatGPT as a source of information to write scientific articles. American journal of obstetrics and gynecology. 2023;229(3):356-357.

36.

Aljamaan

Temsah

M-H

Altamimi

, et al. Reference hallucination score for medical artificial intelligence chatbots: development and usability study. JMIR Medical Informatics. 2024;12(1):e54345.