AI summary
Seven AI prompts for physicians that scaffold the writing and communication work without crossing into clinical decision-making: patient letters, note translation, differential brainstorming, after-visit summaries, insurance appeals, pre-visit triage, and difficult conversation prep. Every output is reviewed by the physician before patient contact. AI never provides medical advice; the physician makes every clinical call.
Physician AI use is regulated, scrutinized, and legally distinct from every other profession. The seven prompts below take that seriously. None of them produce clinical decisions, treatment plans, or diagnoses. They scaffold the writing and communication work that eats physician time (patient letters, after-visit summaries, insurance appeals, prep for hard conversations) so the physician can spend judgment time on the medicine. This is the physician slice of the AI Prompt Library, paired with a connector callout for the EHR and communication tools physicians use. For the broader playbook see Claude for Doctors.
Why do most AI physician-AI workflows produce documentation patients cannot read and notes that miss the clinical picture?
The default physician-AI risk is letting the AI cross from documentation help into clinical recommendation. The line matters legally (medical liability), ethically (informed consent), and clinically (AI confidently produces plausible-sounding clinical content that does not match the patient). Every prompt below is structured to keep AI in the documentation and communication role and to keep the physician in the clinical decision role.
Use AI for the structured prep, translation, and drafting work the prompts cover. Always review before patient contact. Never let AI provide treatment recommendations, dosages, or definitive diagnoses to a patient. If you do draft any patient-facing communication with AI, run it through How to Edit AI Out of Your Writing so it sounds like you, not like a chatbot. When a prompt becomes a daily move, graduate it using the Prompt-to-Workflow Ladder.
What are the seven for doctors prompts?
Prompt 1
Patient Letter Drafter
Most physician letters to patients are written tired, between visits. This prompt drafts the letter so you spend your time on the medical judgment, not the words.
I need to write a letter to a patient. PATIENT CONTEXT (de-identified): [AGE RANGE, RELEVANT MEDICAL HISTORY] THE PURPOSE OF THE LETTER: [TEST RESULTS / FOLLOW-UP / TREATMENT PLAN / DENIAL OF REQUESTED TREATMENT / etc.] THE MEDICAL CONTENT I NEED TO COMMUNICATE: [PASTE OR DESCRIBE] THE PATIENT'S LITERACY LEVEL (best guess from prior visits): [GENERAL READER / HEALTHCARE-FAMILIAR / NEEDS PLAIN ENGLISH] THE ACTION I AM ASKING THE PATIENT TO TAKE: [SPECIFIC] Draft a letter with: 1. OPENING: warm, specific to them, not generic. 2. THE MEDICAL FACTS: stated clearly in language matched to the patient's literacy level, without jargon or with jargon explained. 3. WHAT THIS MEANS FOR THEM: practical implications stated directly. 4. WHAT I AM ASKING THEM TO DO: with the specific action, timeframe, and who to contact. 5. WHAT TO WATCH FOR: warning signs that would change the urgency. 6. THE INVITATION TO ASK QUESTIONS, with how to reach me or the clinic. 7. CLOSE: warm, professional. IMPORTANT: I will review every letter before sending. Do not invent diagnoses, dosages, or test results not in my input. Do not give medical advice you have not been instructed to give. Flag any term you used that might be misinterpreted by the patient.
When to use: Same day as the test result or decision, ideally. · Best model: Claude (most disciplined about not fabricating clinical details).
Prompt 2
Note-to-Plain-Language Translator
Patients have a right to understand their own records. This prompt translates clinical notes into language they can actually act on.
Here is a clinical note from a recent visit (de-identified):
[PASTE NOTE]
The patient who is asking: [BRIEF CONTEXT ABOUT THEIR LITERACY LEVEL AND WHAT THEY ARE CONCERNED ABOUT]
Translate into plain language:
1. WHAT WE FOUND: the findings, restated in plain English appropriate to the patient.
2. WHAT IT MIGHT MEAN: the implications, calibrated to actual clinical significance (do not over-state mild findings).
3. WHAT WE ARE DOING ABOUT IT: the plan, in the order it will happen.
4. WHAT THEY SHOULD DO: their part in the plan.
5. WHAT TO WATCH FOR: warning signs that would change urgency.
6. THE QUESTIONS THEY ARE PROBABLY ABOUT TO ASK that the note does not yet answer.
7. FLAGGED TERMS: any clinical term in the note that, if I send this as-is, would confuse or frighten the patient. Suggest a replacement or note where I should add a brief explanation.
Do not invent findings not in the note. If the note contains uncertainty ("likely," "cannot rule out," "consistent with"), preserve that uncertainty in the translation; do not state probable findings as definitive.
When to use: When a patient requests a translation of their record or before sharing the after-visit summary. · Best model: Claude. Discipline about preserving clinical uncertainty matters here.
Prompt 3
Differential Diagnosis Brainstorm
Used as a check on YOUR thinking, not a replacement for it. This prompt produces a differential to compare against your own working list, surface the diagnosis you might have anchored away from.
Here is the clinical picture (de-identified): PRESENTING SYMPTOMS: [LIST] RELEVANT HISTORY: [LIST: meds, comorbidities, recent events] EXAM FINDINGS: [KEY POSITIVES AND NEGATIVES] THE WORKING DIAGNOSIS I AM CONSIDERING: [YOUR LEAD HYPOTHESIS] Produce a differential brainstorm: 1. MY WORKING DIAGNOSIS RESTATED: what fits, what does not fit. 2. THE TOP 5 OTHER POSSIBILITIES, in approximate likelihood order, with the specific feature that would distinguish each. 3. THE CAN'T-MISS: the diagnoses that are unlikely but life-threatening if missed. 4. THE ANCHORING RISK: based on the history, am I anchoring on a familiar pattern that fits less well than I think. 5. THE NEXT-STEP DIAGNOSTICS: 2-3 tests or examinations that would meaningfully discriminate between the top candidates. 6. WHAT I SHOULD ASK THE PATIENT that I have not yet asked. This is a brainstorm, not a recommendation. I will make the clinical call. Do not provide treatment plans. Be specific; avoid vague "consider..." lists with no clinical reasoning.
When to use: Any case where you want a structured second look at your differential. · Best model: Claude. Discipline about “not a recommendation” matters legally and clinically.
Prompt 4
After-Visit Summary Drafter
Most after-visit summaries are templated and forgotten by the patient before they reach the parking lot. This prompt drafts one they will actually use.
Here is the visit (de-identified):
REASON FOR VISIT: [WHY THE PATIENT CAME IN]
WHAT WE DISCUSSED OR EXAMINED: [BRIEF]
DIAGNOSIS OR ASSESSMENT (PROVISIONAL OK): [BRIEF]
PLAN: [TESTS, MEDICATIONS, REFERRALS, LIFESTYLE, FOLLOW-UP]
WHAT THE PATIENT EXPRESSED CONCERN ABOUT: [THEIR WORRY IN THEIR WORDS]
Draft a one-page after-visit summary:
1. WHAT WE DID TODAY: 2 sentences naming the visit's purpose and outcome.
2. WHAT WE FOUND OR THINK: the assessment, stated plainly and at appropriate certainty level.
3. WHAT TO DO NEXT: numbered list of actions, in order of timing, with the specific deadline for each ("call by next Friday," not "soon").
4. WHAT TO WATCH FOR: warning signs that would change urgency, with the specific phone number to call.
5. YOUR QUESTION: address the specific concern the patient raised, even if briefly.
6. WHEN WE WILL SEE YOU AGAIN: with how to reach the clinic if needed sooner.
Keep under 350 words. Use plain English. Do not invent findings. Do not promise outcomes. Do not list medications or dosages unless explicitly provided in my input.
When to use: Immediately after the visit, while the conversation is fresh. · Best model: Claude. Tone calibration matters; do not over-cheerful and do not over-clinical.
Prompt 5
Insurance Appeal Letter Drafter
An insurance company has denied a treatment you believe is medically necessary. This prompt drafts the appeal letter so you can edit instead of staring at a blank page.
I am writing an appeal for: PATIENT (de-identified): [AGE, RELEVANT CONTEXT] INSURANCE COMPANY: [NAME] TREATMENT DENIED: [SPECIFIC TREATMENT, DEVICE, MEDICATION, OR PROCEDURE] REASON GIVEN FOR DENIAL (paste their letter or summarize): [REASON] CLINICAL JUSTIFICATION (why this treatment is appropriate for this patient): [PASTE OR SUMMARIZE] RELEVANT GUIDELINES OR EVIDENCE that support the treatment: [CATEGORIES OF EVIDENCE, NOT SPECIFIC CITATIONS] WHAT HAS BEEN TRIED OR CONSIDERED FIRST: [PRIOR TREATMENTS OR REASONS THEY ARE NOT APPROPRIATE] Draft an appeal letter with: 1. OPENING: brief, professional, naming the patient ID and the denial reference. 2. THE CLINICAL PICTURE: the patient's relevant medical situation in factual terms. 3. WHY THIS TREATMENT: the specific clinical reasoning, anchored to the denial reason if possible. 4. WHAT THE EVIDENCE SHOWS: the category of evidence supporting the treatment (I will add specific citations). 5. WHAT IS NOT APPROPRIATE INSTEAD: the alternatives the insurance company implied, and why they are not appropriate for this patient. 6. THE REQUEST: specific approval of the treatment, with timeframe. 7. CLOSE: contact information, willingness to provide additional documentation. Do not cite specific journals, study names, or guideline numbers. I will add citations during my review. Do not invent patient details.
When to use: Within 2 weeks of receiving the denial. · Best model: Claude. Tone discipline matters; appeals are legal documents.
Prompt 6
Pre-Visit Patient Triage
Your schedule has a new patient at 2pm. Their intake form is long. This prompt triages it so you walk in already knowing the picture.
Here is the new patient intake form (de-identified): [PASTE INTAKE] Reason for visit: [WHAT THEY WROTE] Prior records I have access to: [BRIEF, OR NONE] Produce a pre-visit brief: 1. THE PROBABLE PRESENTING ISSUE in one sentence, based on the intake. 2. THE PRIOR HISTORY THAT MATTERS for this visit, surfaced from the intake form. 3. THE MEDICATIONS WORTH NOTING: anything that interacts, contraindicates, or shapes the workup. 4. THE RED FLAGS: any symptom or sign that requires immediate evaluation or referral. 5. THE QUESTIONS WORTH ASKING that the intake form did not capture. 6. THE PROBABLE WORKUP, in approximate order. 7. THE ONE PIECE OF CONTEXT to keep in mind about the patient as a person (something from the intake that suggests how to communicate with them). Do not skip a red flag because the intake form did not bold it. Do not suggest treatment; this is a triage brief.
When to use: 15 minutes before the visit. · Best model: Claude. Discipline about not skipping red flags matters.
Prompt 7
Difficult Conversation Prep
You have to deliver a difficult diagnosis or news. The first conversation shapes everything that follows. This prompt prepares the framing so you do not improvise the hardest moments.
I have to deliver a difficult conversation: PATIENT (de-identified, with the context that affects communication): [BRIEF] WHAT THE DIFFICULT NEWS IS: [SPECIFIC] WHAT THEY ALREADY SUSPECT OR KNOW: [BRIEF] WHAT MATTERS MOST TO THEM (from prior visits): [LIFE CONTEXT] WHAT MY MAIN CLINICAL CONCERNS ARE for them right now: [SPECIFIC] Draft a conversation framework with: 1. THE OPENING: how to set up the conversation in 2-3 sentences so they are not blindsided. 2. DELIVERING THE NEWS: the specific words I should use, calibrated to their literacy and emotional state. Avoid both excessive hedging and excessive bluntness. 3. THE PAUSE: where I should stop talking and let them respond. 4. THE FIRST QUESTION I should ask after they have absorbed the news. 5. ADDRESSING WHAT THEY CARE ABOUT MOST: 2-3 sentences that connect the news to their life context. 6. THE PLAN: how to introduce what comes next without overwhelming. 7. THE CLOSE: how to end the visit so they leave with something concrete (a follow-up appointment, a written summary, a number to call) and a sense of partnership. Do not soften the news to the point of misinformation. Do not over-clinical the language. Tone: warm, direct, calibrated to the patient.
When to use: Day before the conversation. Re-read the morning of. · Best model: Claude. Tone discipline is the entire value.
These work across Claude, ChatGPT, Gemini, and Grok. Claude is the strongest default for medical work because of its discipline about preserving clinical uncertainty (“likely,” “consistent with,” “cannot rule out”) and not fabricating clinical details. For HIPAA compliance, the choice of AI tool matters: use only paid plans with a signed Business Associate Agreement (BAA) where required, and verify your institution’s policy on which AI tools are approved for PHI. De-identify all inputs by default; use minimum-necessary patient context.
What is the worst thing you can do with AI for doctors?
Three failure modes will sink physician-AI workflows fastest, and all three have legal and ethical weight.
- Pasting identifiable patient information (PHI) into a free-tier AI tool. Free tiers may train on inputs and may retain content. This is a HIPAA violation. Use only paid plans with verified data-handling terms and, where required, a Business Associate Agreement. De-identify inputs by default; even with paid plans, minimum-necessary applies.
- Letting AI produce treatment recommendations, dosages, or definitive diagnoses that reach the patient. AI fabricates clinical detail confidently. Every prompt above is structured to keep AI in the documentation and communication role; the clinical decision is yours. Any patient-facing AI output must be reviewed by the physician before transmission.
- Asking AI for specific medical citations or guideline numbers. AI tools, including the most expensive, fabricate citations to journals and guidelines that do not exist. Use AI to structure your reasoning; cite the literature yourself from authoritative sources (UpToDate, PubMed, ACP, ACOG, etc.).
What if you want to take this further?
Each prompt above takes inputs you paste in. The next move is connecting AI to the systems where physician work already lives, with extra care for PHI handling.
Connectors are now standard
Claude, ChatGPT, and Grok all support connectors that let your AI read live data from your work tools (Gmail, Notion, GitHub, Asana, HubSpot, Stripe, and many more) instead of relying on you to paste context. For physicians this means the AI can read appropriate non-PHI workflow tools (your calendar, your appointment system, your appeals tracker) while staying out of direct EHR access where institutional policy and HIPAA preclude it.
For doctors, the connectors worth pairing with these prompts:
- Gmail or Outlook connector — for the insurance appeal and patient-letter prompts, AI can reference prior correspondence WITHOUT pulling in actual PHI from secure messaging portals.
- Calendar connector — for the pre-visit triage and difficult-conversation prep prompts, AI references upcoming visit timing.
- Notion connector — if your practice maintains workflow templates, appeal letter templates, or quality-improvement docs in Notion, AI reads them for consistency.
- Google Drive connector — for de-identified case discussion or research summary work, AI reads your reference folder.
- DO NOT use — consumer AI connectors with your actual EHR (Epic, Cerner, athenahealth) unless your institution has explicitly approved the integration with a signed BAA and audited data flow.
What are common questions about AI for doctors?
Is AI use in clinical practice HIPAA-compliant?
It depends on the tool, the plan, and the data handled. Free-tier consumer AI is not HIPAA-compliant. Paid tiers with no-training and no-retention terms may be appropriate for non-PHI work and for de-identified case discussion. For PHI handling, you need a Business Associate Agreement (BAA) with the AI vendor. Claude Enterprise, ChatGPT Enterprise, and certain healthcare-specific AI tools offer BAAs. Verify with your institution’s compliance office before any PHI flows through AI.
Will AI replace physicians?
AI is changing what physicians do, not eliminating the role. Documentation, communication, scheduling, prior authorization, insurance appeals: all compressing with AI. Clinical decision-making, the physical exam, the relationship of trust, the read of the patient as a person: not. Physicians who use AI for the back-office work and spend their saved time on actual care become more valuable, not less. Physicians who let AI make clinical decisions create liability and harm patients.
Which AI tool is best for medical practice?
For documentation drafting on de-identified content, Claude Pro is most disciplined about preserving clinical uncertainty. For PHI handling, you need a HIPAA-compliant tool with a BAA: Claude Enterprise, ChatGPT Enterprise with BAA, or medical-specific tools (Suki, Abridge, Nuance DAX). Verify what your institution has approved. Do not use unapproved tools for clinical work.
Can AI write my SOAP notes?
Some medical-specific AI tools (Suki, Abridge, Nuance DAX) are designed for ambient note generation during the visit and may be appropriate within your institution’s policy. General-purpose AI is appropriate for editing and structuring notes you write, not for generating clinical content from scratch. Always review and sign every note as the responsible clinician.
Should I tell my patients I use AI?
Increasingly required by some state laws (CA, TX, others). Even where not required, transparency is increasingly expected. The line: if AI affects a clinical decision or generates patient-facing content, disclose. If AI assists with documentation or workflow, most policies do not require disclosure. Check your state’s medical board guidance and your institutional policy.
How do I avoid AI fabrication in clinical writing?
Never ask AI for specific citations, guideline numbers, drug dosages, or diagnostic criteria. Use AI to structure your reasoning; reference clinical sources yourself. Every prompt above is built to keep AI in the structural role and to require physician review before patient contact. If you do allow AI to draft clinical content (for review), require the AI to flag any clinical detail it generated that you did not provide as input.
How long does it take to build the physician-AI loop?
Six weeks. Start with the After-Visit Summary Drafter and the Patient Letter Drafter, both on de-identified inputs first to learn the rhythm. Add the Insurance Appeal prompt the next time a denial comes. Most physicians settle into 3-5 of the seven prompts as part of their weekly flow within a quarter. The Differential Diagnosis Brainstorm is a less-frequent but high-value tool for atypical presentations.
The AI Prompt Library · $39
Medical practice workflows, prompt-paved.
Soon to be 1000+ prompts in Notion organized by use case. The full medical section includes everything above plus prompts for CME planning, peer-review feedback, hospital committee writing, residency program teaching prep, and grant application drafting. Plus prompts for every other field. Lifetime access.
Get Smarter About AI Every Morning
Free daily newsletter. Built for people who want to use AI well, not chase every model.
Free forever. Unsubscribe anytime.
Sources to read next?
- HHS HIPAA guidance on AI and health information · the foundational compliance framework
- AMA: Augmented Intelligence in Health Care policy · professional society guidance for physician AI use
- Anthropic prompt engineering documentation · official prompt design guide
- FDA: Artificial Intelligence and Machine Learning in Software as a Medical Device · regulatory framework for clinical AI
- Anthropic Trust Center and BAA inquiries · for HIPAA compliance review with Claude Enterprise
You might also like
- AI Prompt Library · the full library this post pulls from
- Claude for Doctors · the Claude-specific medical playbook
- AI for Doctors · the broader playbook
- How to Edit AI Out of Your Writing · the cleanup pass before any patient-facing communication
- Prompt to Workflow: The AI Ladder · graduate prompts into saved workflows
- Best AI Prompts for Lawyers · for the legal-adjacent work in appeals and documentation
- Best AI Prompts for Email Writing · for the volume of communication physicians produce
Two ways to go further
The AI Prompt Library
1,000+ ready-to-use prompts for Claude, ChatGPT, and Gemini. Stop staring at a blank box.
Get it for $39 →2-Hour Live AI Crash Course
A private, beginner-friendly session across Claude, ChatGPT, Gemini, and the wider landscape.
Book for $125 →