The 4-Part EHR Prompt Formula
Every strong AI prompt for clinical documentation has four components. Missing any one of them is usually why output requires extensive editing.
[Role + EHR] + [Task] + [Clinical Context] + [Format] = Ready-to-use note
Breaking it down:
[Role + EHR] Your specialty and EHR name. I am an internist using Epic EHR.
[Task] What you need: note type, document type, or code lookup. Write a progress note/Generate MDM documentation/Look up ICD-10 code
[Clinical Context] Age, sex, diagnosis, relevant details. No PHI. 45yo male, T2DM, A1c 10.2, refusing metformin.
[Format] How to structure the output for your EHR. Use SOAP format. Order A&P by acuity. Optimize for copy-paste into Epic.
What This Looks Like in Practice
Same patient, two prompts: an 18-year-old admitted with cough and dyspnea on exertion.
❌ Vague prompt: Write a SOAP note for a patient with cough and shortness of breath.
Result: generic placeholders, no EHR formatting, requires extensive editing.
✅ Specific prompt (See it in action using the formula): I am a hospitalist using Epic. Write an H&P for an 18yo male admitted for cough and dyspnea on exertion, likely CAP vs. atypical pneumonia. Include MDM with problem, data, and risk elements. Order A&P by acuity. Optimize for moderate-complexity E/M billing.
Result: MDM paragraph with all three elements, formatted for Epic, ready to copy. Try Doximity Ask → The upgrade is simple: add your specialty, your EHR name, and your coding level goal to any prompt you already use.
5 Mistakes to Avoid
1. Pasting AI output without reading it. AI can generate plausible-sounding but incorrect drug doses, lab values, or clinical facts. Every AI-generated note requires your review before signing.
2. Including PHI in a non-HIPAA-compliant tool. Sending identifiable patient data to a consumer AI tool (like the free version of ChatGPT) is a HIPAA breach. Use de-identified clinical details or a HIPAA-compliant tool like Doximity Ask. 3. Writing vague prompts for complex notes. The more complex the note, the more specific your prompt needs to be. A hospitalist note with five active problems needs context, structure, and format instructions.
4. Never verifying coding suggestions. AI can suggest ICD-10 and CPT codes confidently and incorrectly. Always verify against your documentation before submitting.
5. Signing notes that "look fine" without reading. Rushing through AI-generated content is a legal and clinical risk. Read every sentence before signing.
Your First Week: A 5-Day Plan
Start with one task, compare the output to your usual work, and expand from there.
- Monday: Draft one progress note from your bullet points. Compare to your usual note and see what needed editing.
- Tuesday: Ask AI for MDM documentation on your most complex patient. Check all three elements: problems, data, risk.
- Wednesday: Draft a discharge summary or consult note using the full 4-part formula.
- Thursday: Create one SmartPhrase (or equivalent) for your most common diagnosis and load it into your EHR.
- Friday: Look up the 5 ICD-10 codes you use most often. Save as a quick-reference note.
Remember: Your EHR remains the system of record. AI helps you move faster, it does not replace clinical judgment, legal responsibility, or your review of every note before signing.
