AI prompts for doctors
Last updated
The safe and genuinely useful lane for AI in medicine is language, not judgment: translating findings into patient-friendly words, drafting discharge instructions at the right reading level, digesting literature you provide, and preparing difficult conversations. Nothing below asks AI for a diagnosis or treatment decision — that line is structural, not stylistic.
Privacy is non-negotiable: no patient-identifying data in consumer tools, ever. Every template works on de-identified, generalized inputs.
How to prompt as a physician
- De-identify by default: age band, relevant history in general terms — never names, dates of birth, rare identifying combinations.
- Set the reading level explicitly for patient materials ("8th-grade level", "B1 German") — it's the single biggest clarity lever.
- Provide the sources: for literature tasks, paste the abstracts/papers. General models fabricate citations and dosages with equal confidence.
- Keep clinical judgment out of scope: AI drafts language around YOUR conclusions; it never supplies the conclusion.
The five templates
The MRI report is clear to you. The patient heard "lesion" and stopped hearing anything else.
Rewrite the following explanation for a patient. Audience: [ADULT PATIENT, ~8TH-GRADE READING LEVEL / LANGUAGE: X]. Emotional context: [E.G. ANXIOUS, FIRST DIAGNOSIS].
Rules: everyday words with the medical term once in brackets; what it means for daily life before what it is technically; numbers as natural frequencies ("2 of 100 people") not percentages; no minimizing ("just", "only") and no catastrophizing; end with the 3 questions the patient probably wants to ask, answered in one sentence each.
What I want to convey (my clinical content, already decided):
[PASTE YOUR EXPLANATION IN CLINICAL TERMS — DE-IDENTIFIED]
Discharge in 20 minutes; the instructions need to survive the kitchen table, not just the chart.
Draft discharge instructions from my clinical notes. Audience: [PATIENT PROFILE, READING LEVEL / LANGUAGE — DE-IDENTIFIED].
Format, strictly: 1) "What happened" — 2 sentences; 2) "Your medications" — table: name, what it's for, when, with food?, until when; 3) "Do / Don't for the next [N] days" — max 5 each, concrete ("lift nothing heavier than a full kettle"); 4) "Come back immediately if…" — red flags as observable signs, not clinical terms; 5) follow-up appointments with what happens at each.
Max one page. Every item traceable to my notes — nothing added.
My notes (de-identified):
[PASTE]
Six new papers on a management question, one commute to absorb them.
Digest the following [N] abstracts/papers on [CLINICAL QUESTION]. Deliver: 1) comparison table — study, design, N, population, intervention, key outcome with effect size and CI, major limitation; 2) three-paragraph synthesis: where evidence agrees, where it conflicts (name studies), what remains open; 3) "Practice-relevant deltas" — only findings that would change something vs. [CURRENT GUIDELINE/PRACTICE YOU NAME], each marked [STRONG/WEAK SIGNAL]; 4) anything you could not verify from the provided text, listed separately. Work only from the material. No external claims, no invented references. Material: [PASTE ABSTRACTS, NUMBERED]
Bad news, treatment refusal, family conflict — the conversations that deserve rehearsal and never get it.
Help me prepare a difficult conversation. Situation (de-identified): [E.G. PROGRESSION UNDER THERAPY, FAMILY PUSHING FOR AGGRESSIVE TREATMENT PATIENT DOESN'T WANT]. My goals, ranked: [E.G. PATIENT UNDERSTANDS OPTIONS; AUTONOMY PROTECTED; FAMILY STAYS ENGAGED]. Deliver: 1) a 4-step conversation arc (opening words verbatim, key message in patient language, space for reaction, concrete next step); 2) the 5 hardest things they might say, each with a response that validates before it informs; 3) phrases to avoid and why; 4) then play the [PATIENT/RELATIVE]: respond as they might, and coach my replies after each exchange. Stay with communication — clinical decisions are made and not up for revision here.
The encounter is done, the note isn't. Your fragments in, a structured draft out — for your review.
Turn my fragments into a structured [NOTE TYPE, E.G. SOAP NOTE / REFERRAL LETTER] draft. Institution style: [PASTE A REDACTED EXAMPLE IF AVAILABLE]. Rules: use ONLY my fragments — no clinical content added, no differential invented, no normal-findings boilerplate I didn't state; where a section has no input, write [NOT DOCUMENTED] instead of filling it; keep my exact terms for findings; list any internal contradictions you notice at the end as questions. Fragments (de-identified): [PASTE SHORTHAND / DICTATION]
Frequently asked questions
Can I use ChatGPT or Claude with patient data?
Not with identifiable patient data in consumer versions — that violates privacy law in most jurisdictions (HIPAA, GDPR and equivalents). Options: rigorous de-identification (the templates assume it), or institution-approved deployments with signed agreements (BAA/AVV) and no-training guarantees. Your institution's policy is the authority.
Why no diagnosis or treatment prompts?
Because general-purpose models are not medical devices, hallucinate plausibly (including dosages and citations), and the accountability stays with you regardless. The reliable value today is language work around your judgment — which happens to be where clinician hours actually leak.
Is AI-drafted documentation legally safe?
As safe as your review makes it. The templates constrain drafts to your stated content ([NOT DOCUMENTED] markers, no additions), which keeps verification fast. Signed documentation is yours — same as with human scribes; several health systems now run exactly this pattern with ambient/scribe tools.
How is medical literature search different from the digest template?
Discovery needs grounded tools (PubMed, UpToDate-class resources, specialized evidence engines); general models fabricate references. The digest template deliberately starts AFTER discovery: you provide the papers, it structures and compares them — with a could-not-verify list as the honesty check.