The Art of the Ask

What Does "Prompting" Actually Mean?

If you've spent any time reading about AI, you've encountered the term "prompting"—probably dozens of times, used in ways that seem contradictory or confusing. Before we go further, let's untangle this, because the word carries at least three distinct meanings in AI discussions, and conflating them creates unnecessary confusion.

For this module, we're primarily concerned with the first two meanings—what you type and how to type it better. But understanding that third layer helps explain why the same question can produce different results in different AI tools, even when they use the same underlying technology.

A better way to think about it: prompting is the skill of clear communication applied to a new medium. You're not learning arcane techniques—you're learning to communicate effectively with a particular kind of interlocutor that processes language differently than humans do. The same principles that make you a clear communicator with colleagues, patients, and trainees apply here, just with some adjustments for the AI's particular characteristics.

The Clinical Parallel: You Already Know This

Here's the thing: if you've practiced medicine for any length of time, you already understand the core principles of effective prompting. You just call it something else.

Consider the patient history. Every clinician knows that a well-taken history does most of the diagnostic work. "I don't feel good" gives you almost nothing to work with. "I've had progressive shortness of breath over two weeks, worse with exertion, no chest pain, but I noticed my ankles are swelling and I've gained about five pounds even though I'm not eating more"—that gives you a direction.

The same principle applies to AI. A vague prompt produces vague output. A specific, well-structured prompt produces focused, useful output. The clinical intuition you've developed about what information matters for a given question transfers directly.

Consider how you'd present a case to a specialist:

Notice what you've done: you've established the patient, the timeline, the key symptoms, the relevant background, what you've already ruled out, what you're considering, and what you actually want from the consultant. That's expert-level prompting of a human intelligence. The same structure works for artificial intelligence.

Why Good Input Matters

Large language models are pattern-matching engines trained on vast amounts of text. They predict what text should come next based on statistical patterns learned during training. When you give an LLM a vague prompt, you're essentially asking it to guess what you mean. When you give it a specific, well-structured prompt, you're constraining the space of reasonable responses, making it much more likely to produce something useful.

Compare these two prompts about diabetes management:

Vague:

Specific:

The first prompt will generate something encyclopedic and unhelpful—a survey of diabetes treatment written for no particular audience with no particular purpose. The second will generate a clinically relevant comparison tailored to a specific patient scenario.

The Anatomy of an Effective Prompt

Think of a good prompt as having the same structure as a good case presentation. Not every prompt needs every element, but knowing the components helps you include what matters.

Let's see how these components work together:

This prompt gives the AI everything it needs to produce something genuinely useful on the first try.

Specificity: The Most Valuable Habit

If you take one thing from this module, let it be this: specificity costs nothing but transforms results.

Every time you're about to submit a prompt, pause and ask: what am I leaving implicit that I could make explicit? What assumptions am I hoping the AI will correctly guess?

Notice that the specific versions take longer to write. This is actually the point. The few seconds you spend thinking about what you really need get repaid in responses that actually help, rather than responses you have to fix, regenerate, or supplement.

Role Assignment: When It Helps (and When It Doesn't)

Early AI prompting guides emphasized "persona" heavily—telling the AI to "act as" a specialist. With modern models (GPT-4o, Claude 3.5+, Gemini 1.5), this technique has become less essential. Today's models are much better at inferring appropriate framing from the task itself. If you ask a detailed clinical question, the model responds clinically without being told to "act as a physician."

That said, role assignment still helps in specific situations:

• Adjusting communication style: "Explain this to a patient with an 8th-grade reading level" or "Write this for a medical student" shapes tone and depth.
• Simulation and practice: "Act as a patient with poorly controlled diabetes who is resistant to starting insulin" creates useful role-play scenarios for training.
• Perspective-taking: "What would a clinical pharmacist flag about this regimen?" can surface considerations you might miss.

The key insight: specificity about your task and context matters more than persona. A detailed, well-structured prompt without any role assignment will usually outperform a vague prompt that starts with "Act as an expert." Focus your effort on context, task, and format first.

Structured Prompting: Breaking Complex Tasks into Steps

When you're asking an AI to do something complex, asking for everything at once often produces worse results than breaking the task into logical steps. This is sometimes called "chain of thought" prompting, and it mirrors how clinicians actually reason through complex cases.

This step-by-step approach produces more thorough reasoning than simply asking "What's wrong with this patient?" It also makes the AI's thinking visible, which helps you evaluate whether its reasoning is sound.

The Power of Examples: Showing Rather Than Telling

If you want a specific format, style, or approach, showing an example is often more effective than describing what you want. This is sometimes called "few-shot prompting."

The AI will match the reading level, structure, and tone of your example, producing consistent materials without you having to describe every stylistic element.

Iterative Refinement: The Conversation Continues

One of the most underused features of AI assistants is their ability to have conversations. You don't have to get your prompt perfect on the first try. You can refine.

• If the AI produces something too technical: "That's too complex. Simplify it for a patient with an 8th-grade reading level."
• If it's too long: "Good content, but condense to half the length."
• If it missed your point: "I was asking specifically about dosing considerations for patients with renal impairment."
• If you want it to reconsider: "What would be the argument against your recommendation?"

Here's a practical example of iteration in action:

The key insight is that prompting isn't a single transaction. It's a dialogue. Treat it like a conversation with a knowledgeable but not omniscient colleague.

Prompt Patterns for Clinical Work

Certain prompt structures work well for common clinical tasks. Here are patterns you can adapt:

These templates aren't magic formulas. They're starting points that remind you to include relevant context, specify your task, and constrain the output appropriately.

Common Pitfalls and How to Avoid Them

Working with Uncertainty: Prompting for Nuance

One hallmark of clinical expertise is comfort with uncertainty. AI tools, by contrast, often default to confident-sounding responses even when uncertainty would be more appropriate. You can prompt for more nuanced responses:

You can also explicitly request acknowledgment of limitations:

• "After answering, note any important caveats or situations where this guidance might not apply."
• "If the evidence is limited or conflicting, say so explicitly rather than presenting a definitive answer."
• "What would you want to know that you don't have information about in order to answer this better?"

When Prompting Reaches Its Limits

Effective prompting can dramatically improve AI outputs, but it can't fix fundamental limitations. It's worth knowing when you're bumping against walls that better prompting won't move.

• Knowledge cutoffs: Most AI models were trained on data up to a certain date and don't know about subsequent events, guidelines, or research.
• Hallucination: AI models can generate plausible-sounding but false information. No prompt can guarantee accuracy. Verification remains essential.
• Proprietary information: The AI doesn't have access to your hospital's protocols, your patient's chart, or your organization's policies unless you provide it.
• Tasks requiring real-world interaction: The AI can help you plan a conversation, but it can't have it. It can suggest what to look for on exam, but it can't perform it.
• Logical reasoning edge cases: AI models can still fail at edge cases, multi-step logic, or situations requiring common sense.

Personalizing Your AI: Custom Instructions

Every time you start a new conversation with an AI, it knows nothing about you. You're starting from scratch—explaining your role, your preferences, your context. But all three major AI platforms let you set custom instructions that apply to every conversation automatically. This is like giving a new colleague an orientation on their first day, so you don't have to re-explain the basics every time you work together.

Custom instructions are particularly valuable for clinicians because your context is consistent: your specialty, your patient population, how you prefer information formatted, what you typically need AI help with. Setting this up once saves repetition and produces more relevant outputs from the start.

What to Include in Custom Instructions

Think about what you'd tell a knowledgeable assistant on their first day:

Setting Up Custom Instructions by Platform

Example Custom Instructions for Clinicians

Here's a template you can adapt:

Testing and Refining Your Setup

After setting custom instructions:

1. Start a new conversation and ask a typical question without adding context. See if the response reflects your preferences.
2. Check the tone and format. Is it as concise as you wanted? Using the right structure?
3. Try different task types—patient education, clinical questions, administrative help. Make sure your instructions work across your common use cases.
4. Iterate. If something isn't working, go back and refine. "Be concise" might need to become "Limit responses to 3-4 paragraphs maximum unless I ask for more detail."

Well-configured custom instructions won't make every prompt perfect, but they'll give you a meaningful head start. Instead of explaining your context every time, you can jump straight to the specific question—and get responses that are already calibrated to your needs.

Practical Exercises

To build prompting skills, try these exercises:

1. The Specificity Upgrade: Take a question you'd naturally ask an AI and make it more specific. Then more specific again. See how outputs differ.
2. The Format Requirement: Ask for information in three formats: paragraph, bullet points, and table. Notice which is most useful when.
3. The Step-by-Step Approach: Take a complex clinical question and break it into sequential steps. Compare to asking all at once.
4. The Example Provision: When asking for content in a particular style, provide an example. Compare to asking without one.
5. The Iteration Practice: Start with a basic prompt, then refine through 3-4 rounds of follow-up. Notice how output improves.
6. The Cross-Check: Ask a clinical question, then follow up with "What's the strongest argument against what you just said?"
7. The Audience Shift: Ask the AI to explain a concept to a patient, then to a medical student, then to a specialist. Notice how context shapes output better than "act as" instructions.

Readings

Podcasts & Video

Books