81% of physicians used AI professionally in 2026, more than double the 38% who reported using it in 2023Source ↗
Ambient AI scribes saved 7,260 Permanente physicians an estimated 15,791 hours of documentation time across 2.5 million patient encounters in their first yearSource ↗
The average number of AI use cases per physician rose from 1.1 in 2023 to 2.3 in 2026, led by research summaries and clinical documentationSource ↗
More than 40% of US physicians reported logging in to OpenEvidence daily for point-of-care clinical decisions as of mid-2025Source ↗
writingClaudeChatGPT

Appeal letters that overturn prior authorization denials

Prior authorization is the paperwork physicians hate most — 75% say denials have increased over five years, and 61% worry payers' own AI is driving more of them. Writing the medical-necessity appeal is the bottleneck, and drafting it with AI turns a 30-45 minute letter into a 5-minute review, which is why both physicians and patient-advocacy tools have adopted it fastest.

Prompt
You are a physician writing a medical-necessity appeal that a utilization reviewer can approve quickly. Draft an appeal letter for this denied service: {{treatment}}.

De-identified clinical summary: {{clinical_summary}}
The denial letter states: {{denial_reason}}

Requirements:
- One page maximum, professional and factual — no outrage, no pleading.
- Open by identifying the service and requesting reconsideration, quote the denial language exactly, then rebut it point by point.
- Use only the clinical facts I provided. Do not invent history, exam findings, lab values, or failed therapies. Where payer criteria typically require a fact I did not give you (documented step therapy, symptom duration, functional impairment), insert [NEED: description] so I can pull it from the chart.
- Reference a clinical guideline only if I included it in the summary; otherwise write [CITATION: suggest a guideline for me to verify] instead of inventing one.
- State the clinical consequence of further delay in one concrete sentence.
- Close by requesting a peer-to-peer review with a same-specialty reviewer within the required timeframe.
- After the letter, list the attachments this appeal should include.

Fill in your details and the prompt updates live — then copy.

communicationChatGPTClaude

Patient portal replies you review instead of write

In Basket volume has grown so much that health systems now pipe GPT-4 draft replies directly into Epic. The UC San Diego Health trial found drafts did not shorten reply time but reduced the burden of starting from a blank box — and a separate analysis found 35-45% of erroneous drafts were sent completely unedited, which is exactly the failure mode a disciplined prompt and review habit prevents.

Prompt
You draft replies to patient portal messages for a busy {{specialty}} clinic. Draft a reply I can edit and send.

Patient's message (de-identified): {{patient_message}}
What I want to communicate: {{my_answer}}

Rules:
- Under 120 words, written at a 6th-8th grade reading level: warm, direct, no medical jargon without a plain-word explanation.
- Communicate only the clinical content I gave you. Do not add advice, reassurance about symptoms, medication changes, or interpretations I did not state.
- If my answer does not fully address something the patient asked, list it under "UNANSWERED" below the draft instead of improvising.
- Never include a drug dose unless I stated it.
- If the message describes any potentially urgent symptom, open with a clearly worded safety-net line ("If X worsens or you develop Y, call 911 or go to the ER").
- End with a concrete next step: what the patient should do and when to expect follow-up.
- No "as an AI" language — this is sent from my clinic account under my name.

Fill in your details and the prompt updates live — then copy.

writingClaudeChatGPTGemini

Discharge instructions patients can actually follow

Most discharge instructions are written well above the recommended 6th-grade reading level, and patients act on what they understand, not what they were handed. In a blinded JMIR survey, patient raters preferred GPT-generated ED discharge instructions, and creating discharge instructions and care plans is now one of the top physician AI uses in the AMA survey (30%).

Prompt
You write discharge instructions for patients leaving a {{setting}}. Create a plain-language instruction sheet for: {{diagnosis_and_treatment}}. Follow-up plan: {{follow_up}}.

Structure, in this order:
1. "What happened today" — the diagnosis and what we did, in everyday words, two or three sentences.
2. "Taking care of yourself at home" — 4-6 short bullets covering activity, diet, and wound or symptom care as relevant.
3. "Your medications" — list only medications I named; for each, write "take exactly as prescribed" rather than inventing doses or schedules.
4. "Get help right away if" — specific warning signs, each on its own line, separating when to call the office versus go to the ER or call 911.
5. "Your follow-up" — who, when, and what to bring.

Rules:
- 6th-grade reading level. Short sentences. Define any unavoidable medical term in parentheses.
- Use only the clinical facts I provided; write [CONFIRM: item] wherever a standard instruction depends on something I did not state.
- Under 400 words so it prints on one page.
- Then produce the same sheet in {{language}}, and flag any phrase that may not translate precisely so our interpreter can verify it.

Fill in your details and the prompt updates live — then copy.

analysisClaudeChatGPT

Evidence summaries before you change what you prescribe

Summarizing medical research is the single most common physician AI use — 39% in the AMA's 2026 survey — and citation-grounded tools like OpenEvidence report daily use by over 40% of US physicians. The failure mode is a chatbot answering from memory with invented citations, so the working pattern is to paste the actual abstracts or guideline text and force the model to stay inside them.

Prompt
You are an evidence synthesis assistant for a practicing {{specialty}} physician. I will paste source material — abstracts, guideline excerpts, or full-text sections. Answer my clinical question using ONLY that material.

Clinical question: {{clinical_question}}
Patient population I care about: {{population}}
Source material: {{source_material}}

Output format:
1. Bottom line — 2-3 sentences answering the question as directly as the sources allow.
2. What the evidence says — key findings with effect sizes and confidence intervals as reported, naming which pasted source each point comes from.
3. How this differs from prior or common practice — only if the sources address it.
4. Limitations — study design, population mismatch with my patients, and any conflicts of interest noted in the sources.
5. What the sources do NOT answer — list the parts of my question that are not addressed, rather than filling gaps from memory.

Hard rules: cite nothing beyond what I pasted. Do not supply doses, thresholds, or recommendations from your training data — if the sources omit them, say so. Flag any drug dosing that appears in your summary so I can verify it against the label.

Fill in your details and the prompt updates live — then copy.

planningClaudeChatGPTCopilot

Pre-visit briefs that make short slots feel longer

A complex patient arrives with years of chart history and outside records, and you have 15 minutes. Chart summarization is now one of the top physician AI uses (28% in the AMA survey), and EHR vendors are building it in — but a structured prompt over a de-identified chart extract gets an organized visit plan today, before your system rolls anything out.

Prompt
You are helping me prepare for a clinic visit. Build a pre-visit brief from this de-identified chart extract.

Visit reason: {{visit_reason}}
Appointment length: {{visit_length}}
Chart extract (problem list, medications, recent labs and notes — de-identified): {{chart_extract}}

Format, one page maximum:
1. One-line patient snapshot (age band and key conditions, exactly as given).
2. Active problems ranked by what most needs attention today, one line each with the latest relevant data point.
3. Medication flags — anything in the extract suggesting nonadherence, duplication, interactions, or missing monitoring labs. Frame each as a question to verify, not a conclusion.
4. Gaps and overdue items explicitly documented in the extract (screenings, vaccines, labs).
5. Suggested visit agenda — the top 3 things to address in the time available, with a one-line rationale each.
6. Open questions — anything ambiguous or contradictory in the extract, marked [CLARIFY].

Rules: use only what is in the extract. Never infer diagnoses, results, or dates that are not written there. If the extract looks incomplete for the visit reason, say exactly what is missing instead of guessing.

Fill in your details and the prompt updates live — then copy.

automationChatGPTClaude

Complete SOAP notes from dictated shorthand

Documentation is where physician AI has delivered its clearest win: Permanente's ambient scribes saved nearly 16,000 hours across 2.5 million encounters, with 84% of physicians reporting better patient communication. If your organization has an ambient scribe, use it — this prompt covers the same job for physicians without one, turning post-visit shorthand into a complete, signable note instead of an evening of pajama-time charting.

Prompt
You are a medical scribe converting my dictated shorthand into a complete clinical note. Visit type: {{visit_type}}.

My shorthand (de-identified): {{shorthand}}

Produce a SOAP note:
- S: chief complaint in the patient's words if I gave them, relevant history, pertinent positives and negatives I mentioned.
- O: vitals, exam findings, and results exactly as dictated — never normalize or expand an exam I did not describe.
- A: assessment as dictated, with differential only if I dictated one.
- P: plan as dictated — orders, prescriptions, referrals, counseling, follow-up interval, and return precautions.

Rules:
- Use only what is in my shorthand. Where a standard element for this visit type is missing (review of systems, time spent, counseling documentation), write [VERIFY: element] — do not fill it with boilerplate. A templated normal exam I did not perform is a false record.
- Past tense, factual, no editorializing about the patient.
- Expand ambiguous abbreviations; keep standard ones.
- After the note, list what a coder would still need to support the E/M level this visit type usually bills, based only on what is present or missing above.

Fill in your details and the prompt updates live — then copy.

Common questions from physicians

Is it a HIPAA violation to use ChatGPT with patient information?

With identifiable patient information on a consumer account, yes — OpenAI does not sign a Business Associate Agreement for standard ChatGPT, and removing the name alone is not de-identification when dates, rare diagnoses, and locations remain. The compliant patterns are: fully de-identify per the Privacy Rule before pasting, or use a BAA-covered option such as ChatGPT Enterprise, the API inside a compliant app, an EHR-integrated tool, or the NPI-verified ChatGPT for Clinicians.

Can I rely on AI for diagnoses or treatment decisions?

No — the AMA's framing is 'augmented intelligence' for a reason. Citation-grounded tools like OpenEvidence are widely used as fast reference at the point of care, but the diagnosis, the order, and the liability remain with the licensed physician, and general chatbots still fabricate citations and misstate doses. Verify anything that changes management against the primary source or drug label.

Who is liable if an AI-drafted note or patient message is wrong?

You are. A signed note and a portal reply sent under your name are your medical record and your medical advice, regardless of what drafted them — and studies show a meaningful share of erroneous AI drafts get sent unedited. Clear liability frameworks are physicians' top regulatory ask in the AMA survey precisely because the law has not caught up; until it does, assume full responsibility and read every word.

My hospital hasn't rolled out AI tools — can I just use my own account?

Check your organization's AI policy first; many systems now prohibit unapproved tools for anything clinical because unsanctioned 'shadow AI' on personal accounts is their biggest data-leak risk. De-identified, non-clinical drafting on a personal account is generally lower risk, but anything touching real patient data should wait for a sanctioned, BAA-covered tool — and 85% of physicians tell the AMA they want a say in what gets adopted, so ask.

Related professions