Getting good output from an AI isn’t magic — it’s a skill. And like most clinical skills, it improves with deliberate practice. This guide covers the four core principles that separate a useful AI interaction from a frustrating one.

1. Set the Role

The single most effective thing you can do is tell the AI who it is and what it’s doing. LLMs are trained on enormous amounts of text and can adopt many personas — give it the right one.

Instead of: “Summarize this note.”

Try: “You are an experienced hospitalist physician. Summarize the following progress note in 3 bullet points suitable for a rapid team handoff.”

The role shapes vocabulary, level of detail, and clinical judgment in the output. A “second-year medical student” will explain more; an “attending cardiologist” will be more terse and technical.

2. Provide Context

AI has no memory of your patient, your institution, or your clinical setting. Every interaction starts fresh. Give it what it needs.

Useful context to provide:

The clinical question or goal
Relevant history (age, major diagnoses, current medications if relevant)
What output format you need
Who the audience is (patient, specialist, med student, etc.)

Example:

“The patient is a 72-year-old man with HFrEF (EF 30%), CKD stage 3, and type 2 diabetes. He was admitted for acute decompensated heart failure. Draft a discharge instruction sheet in plain language (8th grade reading level) covering fluid restriction, daily weights, and when to call the doctor.”

The more specific your context, the less editing you’ll need to do.

3. Be Specific About Format

Tell the AI exactly how you want the output structured. Common formats that work well in clinical contexts:

Bullet points — great for rapid review, handoffs, clinical reasoning lists
Numbered steps — good for procedures, protocols, and patient instructions
Table — useful for medication comparisons, differential diagnoses
Paragraph narrative — best for letters, referrals, summaries for patients or families
SOAP format — familiar structure for progress notes

Example format instruction:

“Provide your response as a table with three columns: Diagnosis, Supporting Evidence, Against. Limit to 5 rows.”

If you don’t specify format, you’ll get whatever format the AI defaults to — which may or may not be what you need.

4. Iterate

Your first prompt is rarely your best prompt. Treat AI interactions like a conversation — follow up, redirect, and refine.

Useful follow-up patterns:

“Make this more concise” — if the output is too long
“Explain your reasoning” — if you want to audit the AI’s clinical logic
“Reformat as bullet points” — if the structure isn’t what you need
“Add a section on [X]” — if something’s missing
“The patient also has [X]. How does that change your recommendations?” — to add context mid-conversation

Putting It Together

Here’s a before/after example:

Weak prompt: “Write a note.”

Strong prompt:

“You are an experienced hospitalist. Write a brief hospital course section for a discharge summary. The patient is a 65-year-old woman admitted for community-acquired pneumonia treated with ceftriaxone and azithromycin for 5 days with good clinical response. She was afebrile for 48 hours prior to discharge. She has a follow-up chest X-ray ordered in 6 weeks. Write in concise, professional prose suitable for a primary care physician.”

The strong prompt takes 30 extra seconds to write and will save you 5 minutes of editing.