AI Prompting
( 14 minute read )
There is something quietly ironic about using artificial intelligence to address the burnout produced, in part, by the increasing administrative and technological burden on healthcare workers. But the irony resolves when you consider what LLMs actually are: extraordinarily powerful thinking partners, available at any hour, with no waiting room, no billing codes, and no judgment.
Used well, a large language model is not a replacement for therapy, mentorship, or systemic reform. It is something more immediate and more personal: a tool for structured reflection, accelerated learning, and active problem-solving. For a nurse finishing a twelve-hour behavioral health shift at 11pm, the gap between “I should understand the physiology of my cortisol response” and “I have time and energy to pursue that” can feel insurmountable. An LLM closes that gap.
This section does not offer a list of affirmations or generic wellness prompts. It offers a clinically grounded, technically rigorous guide to using AI models — particularly Claude, GPT-4, and similar frontier systems — to engage seriously with every dimension of burnout: its biology, its psychology, its institutional drivers, and its solutions. The quality of what you get out depends almost entirely on the quality of what you put in. That is what this guide is about.
-----
Understanding How LLMs Actually Work
Before you can use these tools effectively, it helps to understand what they are doing when they respond to you.
Large language models are trained on vast corpora of text — research papers, clinical guidelines, books, conversations, code — and learn to predict the most contextually appropriate continuation of any given input. They are not databases. They do not retrieve stored facts. They generate responses by drawing on deep statistical relationships between concepts, and the richer and more specific your prompt, the more precisely those relationships can be activated.
This has a direct implication for how you should interact with them: vague inputs produce vague outputs. A question like “tell me about burnout” will produce something encyclopedic and generic. A question like “I am a nurse in a high-acuity ED who has worked three consecutive night shifts and noticed increasing detachment from patients — walk me through the neuroendocrine mechanisms that might be driving this” will produce something clinically specific, practically useful, and genuinely illuminating.
The other critical thing to understand is that LLMs are responsive to context, role, structure, and iteration. Each of these is a lever you control. Learning to pull them deliberately is the core skill this section teaches.
-----
Foundational Technique 1: Role and Persona Assignment
The single highest-impact thing you can do before asking any substantive question is tell the model who it is and who you are.
This is known as role prompting, and it works because it activates specific clusters of knowledge, vocabulary, and reasoning style. Anthropic’s official guidance on Claude emphasizes that assigning a clear role in a prompt — even in a single sentence — focuses the model’s behavior and tone for the use case in ways that general queries cannot.
Basic form:
> “You are an occupational health physician with expertise in healthcare worker burnout and neuroendocrinology. I am a registered nurse working in emergency medicine. I want to understand [topic] in clinical depth.”
Why this works: The model now knows the lens through which to filter its knowledge (occupational medicine, neuroendocrinology), the vocabulary level appropriate to use (clinical, not lay), and the relational dynamic (peer-to-peer, not explanatory).
For burnout specifically, high-value personas include:
- Occupational health physician or psychiatrist
- Researcher in organizational psychology or health systems
- Expert in polyvagal theory and nervous system regulation
- Health policy analyst or labor economist
- Somatic therapist or trauma-informed practitioner
- Organizational behavior consultant
You are not limited to one. You can instruct the model to reason from multiple perspectives in sequence: “First analyze this from an organizational systems perspective, then from a physiological one, then identify where those analyses converge.”
Important nuance: Anthropic’s own best practice guidance notes that overly rigid role constraints can limit helpfulness. The goal is to orient the model, not to box it in. “You are a helpful occupational health expert” outperforms “You are a world-renowned burnout specialist who only speaks in clinical terminology and never expresses uncertainty.” The latter produces rigid, often less useful responses. The former produces grounded, appropriately confident engagement.
-----
Foundational Technique 2: Chain-of-Thought Prompting
Chain-of-thought (CoT) prompting is among the most thoroughly validated techniques in the LLM literature. It instructs the model to reason through a problem step by step rather than jumping directly to a conclusion — and this single modification dramatically improves accuracy on complex, multi-part questions.
The technique emerged from research showing that simply adding “Let’s think step by step” to a prompt could significantly improve performance on reasoning tasks — it mirrors the methodical, step-by-step approach clinicians employ when tackling complex diagnostic or treatment-planning challenges.
For burnout-related inquiry, this is particularly valuable because burnout is not a simple phenomenon. It is the product of intersecting physiological, psychological, organizational, and cultural forces. A model asked to address all of them simultaneously will tend toward surface-level synthesis. A model asked to reason through them sequentially will produce substantively deeper analysis.
Examples of CoT prompting applied to burnout:
Basic:
> “I want to understand why I feel more cynical toward patients after a run of high-acuity shifts. Think step by step through the neurobiological, psychological, and relational mechanisms that might explain this.”
Advanced:
> “Walk me through your reasoning in stages: First, explain the HPA axis response to sustained occupational stress. Second, describe how chronic cortisol elevation affects the prefrontal cortex and its capacity for empathy and emotional regulation. Third, connect this to the depersonalization dimension of the Maslach Burnout Inventory. Fourth, identify what the evidence says about reversibility.”
The explicit request for staged reasoning produces something qualitatively different from a single-shot question — and you can ask follow-up questions at each stage.
-----
Foundational Technique 3: Context Layering
Context is the most underused resource in most people’s LLM interactions. The more specific and structured the context you provide, the more precisely the model can calibrate its response.
This is sometimes called “contextual scaffolding” — building a rich situational frame around your question before you ask it. Anthropic’s API documentation recommends wrapping different types of content in their own sections — instructions, context, input — to reduce misinterpretation and improve response quality, and advises placing the most important context at the beginning of long prompts.
For a healthcare worker exploring burnout, context layers might include:
- Your role and setting: specialty, acuity level, shift structure, team size
- Your current state: where you are on the burnout spectrum right now
- Your knowledge baseline: what you already understand about the topic
- Your goal for the conversation: understanding, problem-solving, emotional processing, research
- Your constraints: time available, energy level, what format is most useful
Example of a well-layered prompt:
> “Context: I am a critical care nurse, three years post-COVID-19 pandemic, working in a 36-bed ICU with chronic understaffing. I have a solid understanding of stress physiology but less familiarity with the organizational literature on burnout. I am currently experiencing moderate emotional exhaustion but not depersonalization. My goal is to understand what the research says about the relationship between staffing ratios and burnout outcomes, and what institutional interventions have shown the most evidence of effectiveness. Please give me a thorough, research-grounded response, and flag where the evidence is strong versus preliminary.”
This prompt will produce something dramatically more useful than “what causes burnout in ICU nurses?” — not because the model knows more, but because you’ve given it the information it needs to apply what it knows accurately to your situation.
-----
Foundational Technique 4: Iterative Dialogue
One of the most significant misunderstandings about LLMs is that they should be used like search engines — single queries producing single answers. The most powerful use of these tools is iterative: a conversation that deepens over multiple exchanges, with each response informing the next question.
This matters for burnout specifically because the topic is not linear. Understanding burnout means moving between the physiological and the structural, the personal and the systemic, the immediate and the long-term. An iterative conversation allows that movement.
Effective iteration patterns:
Depth drilling: Ask for an overview, then drill into the most relevant element.
> “Now focus specifically on the relationship between moral injury and depersonalization — that distinction feels important. Go deeper on the evidence.”
Perspective shifting: Ask the same question from a different angle.
> “You’ve described this from a clinical psychology perspective. Now analyze the same situation through the lens of labor economics and institutional incentive structures.”
Challenging the response: Push back to test the reasoning.
> “You said resilience training has limited evidence for long-term burnout reduction. What does the strongest counter-argument look like, and how does the evidence weigh against it?”
Synthesis requests: After exploring multiple angles, ask for integration.
> “Given everything we’ve discussed — the physiological mechanisms, the institutional drivers, and the evidence on interventions — what does an evidence-based personal strategy look like for someone in my specific situation?”
-----
Foundational Technique 5: Structured Output Requests
LLMs can produce information in whatever format is most useful to you. This is not cosmetic — structure shapes comprehension, and the right format for a clinical learner differs from the right format for someone processing an emotional experience.
Being explicit about format is a simple technique that consistently improves usability.
High-value format requests for burnout learning:
- “Give me a summary appropriate for a clinical audience, followed by a brief layperson translation.”
- “Organize your response around: (1) what the evidence clearly supports, (2) what is promising but preliminary, and (3) what is contested or unclear.”
- “After your main response, give me three follow-up questions I should ask to go deeper.”
- “Conclude with the one most actionable insight from everything you’ve described.”
- “Cite or reference specific studies or researchers where possible, and note where you are uncertain about a source.”
The last point is important: LLMs can and do confabulate citations. Asking the model to flag its uncertainty does not eliminate this, but it reduces confident-sounding fabrications and trains you to verify specific claims — which is the correct approach to using AI for anything clinically relevant.
-----
Applied Prompt Library: Burnout by Domain
The following prompts are designed for specific dimensions of burnout. Each can be used as written or modified with your own context layering.
-----
Physiological Understanding
The HPA Axis and Chronic Stress
> “You are an endocrinologist with expertise in occupational stress. Walk me through the full cascade of the HPA axis response to chronic, unpredictable stress — the kind experienced in emergency medicine. Include the role of cortisol, CRH, and ACTH; the difference between acute and chronic activation; and the downstream effects on immune function, sleep architecture, and emotional regulation. Think step by step.”
Sleep Deprivation and Cognitive Function
> “I work rotating shifts including overnight. Explain the specific cognitive and emotional effects of chronic sleep deprivation relevant to healthcare performance and burnout risk. Include what the polysomnographic research shows about shift workers specifically, and what evidence-based interventions are most effective for this population.”
Allostatic Load
> “Explain the concept of allostatic load and how it applies to emergency medicine nurses over a multi-year career. What biomarkers are associated with elevated allostatic load, what are the health consequences, and what does the evidence say about reducing it in high-demand professions?”
-----
Psychological and Emotional Dimensions
Moral Injury vs. Burnout
> “Clarify the clinical distinction between moral injury and burnout. They are often conflated — what are the diagnostic, phenomenological, and neurobiological differences? For a nurse in a behavioral health emergency setting, which framework is more explanatory, and does that change what interventions are appropriate?”
Compassion Fatigue vs. Empathy Erosion
> “Walk me through the distinction between compassion fatigue and empathy erosion as described in the occupational health literature. Are they the same phenomenon measured differently, or distinct processes? What does this mean for the depersonalization dimension of burnout — is depersonalization protective, pathological, or both?”
Nervous System Regulation
> “You are a practitioner trained in polyvagal theory and somatic psychology. Explain, in clinical terms, what is happening in the autonomic nervous system during a high-acuity shift — and then describe the evidence base for specific nervous system regulation techniques that can be practically applied before, during, and after high-stress clinical work.”
-----
Systemic and Institutional Analysis
Staffing Ratios and Burnout
> “Summarize the research evidence on the relationship between nurse-to-patient staffing ratios and burnout rates. Which studies are most methodologically rigorous? Where have policy interventions — such as California’s mandatory ratio legislation — produced measurable outcomes, and what are the limitations of that evidence?”
Electronic Health Records and Administrative Burden
> “What does the literature say about the contribution of electronic health record burden to physician and nurse burnout? Include data on documentation time as a proportion of clinical time, the studies on the Pajama Time phenomenon, and what EHR design interventions have shown measurable benefit.”
Moral Injury at the Systems Level
> “Analyze healthcare worker burnout through the framework of moral injury at the institutional level. How do organizational factors — production pressure, staffing decisions, resource rationing, institutional culture — produce the specific experience of moral injury in frontline workers? Cite the most important theorists and studies in this literature.”
-----
Solutions and Interventions
What Actually Works
> “Provide a rigorous, evidence-stratified summary of interventions for healthcare worker burnout. Organize by: (1) individual-level interventions with strong evidence, (2) team and unit-level interventions with strong evidence, (3) institutional and policy-level interventions, and (4) interventions that are widely used but have weak or mixed evidence. Be direct about what the research supports versus what healthcare systems often implement instead.”
Building a Personal Protocol
> “Based on the current evidence in occupational health and stress physiology, help me design a realistic burnout mitigation protocol for a nurse working in a high-acuity ED on rotating shifts. The protocol should address sleep, physiological recovery, cognitive processing of trauma, social connection, and when to seek formal support. Make it specific and practical, not aspirational.”
Advocacy and Structural Change
> “I want to understand how individual clinicians have successfully driven institutional change around burnout and working conditions. What does the organizational behavior and health systems literature say about effective advocacy strategies from within healthcare institutions? What approaches have produced durable change versus those that stall at the pilot stage?”
-----
Using AI for Ongoing Reflection
Beyond structured research, LLMs can serve as a tool for regular reflective practice — something the clinical literature increasingly recognizes as protective against burnout.
Research into AI-assisted journaling, including the MindScape project at Dartmouth, has demonstrated that LLM-generated contextual and personalized prompts can promote self-reflection and emotional development in ways that traditional journaling does not, particularly when those prompts are calibrated to the user’s current circumstances.
For healthcare workers, a brief reflective exchange after a difficult shift — not processing trauma, but naming experiences and tracing their physiological and emotional signatures — builds the metacognitive awareness that the burnout literature identifies as one of the strongest protective factors.
Reflective prompts designed for clinical contexts:
> “I want to do a brief structured reflection on my shift today. Ask me three questions that would help me identify what was most emotionally demanding, how my body responded, and what I might do differently to protect my capacity tomorrow.”
> “I noticed I felt irritable with a patient today and I’m not sure why. Help me think through what might have been happening — physiologically, emotionally, and situationally — without pathologizing it.”
> “I am experiencing what feels like compassion fatigue. Help me distinguish whether what I’m describing is better understood as exhaustion, moral injury, vicarious trauma, or something else — and what that distinction suggests about next steps.”
As researchers in AI and mental health have noted, LLM-based tools offer a judgment-free space that blends the privacy and interiority of personal diaries with the responsive engagement of conversation — uniquely suited to the kind of emotional accounting that underlies effective self-reflection.
-----
A Note on Limitations
LLMs are powerful tools. They are not infallible ones.
They can produce plausible-sounding citations that do not exist. They can reflect biases embedded in their training data. They can be confidently wrong on specific empirical questions. They are not therapists, not physicians, and not a substitute for human connection or formal mental health support.
The correct posture is one of informed, critical engagement: use these tools to accelerate your understanding, deepen your reflection, and identify directions for further inquiry — then verify specific claims, seek out primary sources, and bring what you learn into conversation with colleagues, supervisors, and clinicians who know you.
Used that way, AI is not a replacement for anything. It is an expansion of what is possible — available at 11pm after a shift that broke something small in you, ready to help you understand what happened and what to do next.
-----
The prompts in this section are starting points. The most effective use of these tools is iterative and personal — beginning with a prompt, building on the response, and developing a conversation that is specific to your situation, your role, and where you are right now.
-----
Sources include: Nagarajan et al., Human Resources for Health (2024); Wei et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” NeurIPS (2022); Anthropic Claude Prompting Best Practices (2025); Amazon Web Services / Anthropic Prompt Engineering Guide (2024); Nepal et al., MindScape Contextual AI Journaling, CHI (2024); Vecchione & Singh, “Artificial Intelligence is Mental,” Big Data & Society (2025); Liu et al., “Prompt Engineering in Clinical Practice,” JMIR (2025); Jeon et al., “Chain-of-Thought Prompting in Medical Reasoning,” Computers in Biology and Medicine (2025).