AI Prompts for Doctors
Ten clinical prompt templates for SOAP notes, referral and prior authorization letters, PICO searches and patient explanations, with a PHI masking check.
The interactive tool loads here when JavaScript runs.
About this toolWhat is AI Prompts for Doctors?
Physicians lose hours every week to the writing that surrounds care: turning dictation into a SOAP note, answering portal messages, arguing a prior authorization with a payer, rewriting a discharge plan so a patient can follow it. This kit gives you ten prompt templates for that paperwork. Pick a task, fill in a few fields, and the page assembles a structured prompt you paste into ChatGPT, Claude, Gemini, Copilot or the assistant your hospital has approved.
Every prompt casts the model as a drafting assistant that works only from the facts you supply, marks gaps as [MISSING] instead of inventing findings, and copies doses and values exactly. A privacy check scans what you typed for names, dates of birth, MRNs, phone numbers and other identifiers and masks them in one click, so the excerpt you paste is closer to de-identified. The page runs in your browser and sends nothing anywhere.
- Ten templates covering notes, letters, payer appeals, evidence searches and patient education
- Structured prompts with role, context, task, constraints and output format
- PHI scan and one-click masking of names after titles, MRNs, dates, phone numbers, emails and ages over 89
- Reading-level and language options for patient-facing text
- Favourites, search and a Markdown export of the whole kit for your team
- Runs locally in the browser with no account and no upload
Updated
How to use AI Prompts for Doctors
- Type your specialty and setting, for example outpatient cardiology, and choose your region so the prompt uses local spelling and conventions.
- Search or browse the templates and pick one, such as SOAP note from dictation or Prior authorization or appeal letter. Star the ones you use daily.
- Fill in the fields with de-identified notes, or press Try an example to see a finished prompt built from fictional data.
- Run the privacy check, mask any identifiers it finds, then copy the prompt and paste it into your approved assistant.
- Review the draft against the check-the-output list before anything goes into the chart or to a patient.
Worked examples
Prior authorization appeal
Input
Denial reason pasted from the payer letter, plus treatment history with dates and stop reasons
Output
Appeal letter, a criterion-by-evidence checklist and a Not yet documented list
Mapping each denial criterion to a line in the chart is what reviewers look for. The Not yet documented list tells you what to add to the record before sending, rather than letting the model argue around a gap.
Discharge instructions at grade 6
Input
Pneumonia discharge plan with antibiotic course, analgesia, follow-up and return precautions
Output
Plain-language sheet with a medicines table and a separate When to get help now section
Check the medicines table against the discharge medication list line by line; that is where a model is most likely to drop or merge an item.
Background and insight
Large language models write fluent clinical prose and also produce confident errors: a plausible lab value that was never measured, a guideline that does not say what the model claims, a dose with the wrong unit. The templates are built to make those errors visible. Each prompt asks for a list of gaps or unverified claims next to the draft, separates the physician's assessment from anything the model suggests, and forbids citations that are not in your material.
Privacy rules shape what you can paste. Under HIPAA a covered entity may share protected health information with a vendor only under a business associate agreement (45 CFR 164.502(e) and 164.504(e)), and HHS guidance from 2016 treats a cloud service that stores ePHI as a business associate even if it cannot read the data. Most consumer chat apps do not sign one. The Safe Harbor method in 45 CFR 164.514(b)(2) lists 18 identifiers to remove; the alternative is an expert determination. In the UK and EU, health data is special category data under GDPR Article 9.
The masking helper is pattern based. It catches common formats such as "Mr. Alvarez", "MRN: 4471203", 03/14/1957, phone numbers, emails and ages over 89, which are Safe Harbor identifiers, but it cannot recognise a name written without a title or a rare diagnosis that identifies someone in a small town. Read the masked text before you copy it.
Common use cases
Clinic documentation
Turn dictated bullet points into a SOAP note in your preferred style, with gaps such as missing allergies flagged for you to fill.
Payer paperwork
Draft prior authorization requests, appeals and peer-to-peer talking points that quote each policy criterion and point to the evidence.
Patient communication
Explain results, reply to portal messages and write condition handouts at a reading level patients can act on.
Evidence questions
Frame a bedside question as PICO with MeSH terms and a Boolean search string, then read the studies yourself.
Tips and limitations
- Paste the smallest excerpt the task needs. A referral rarely needs the whole problem list.
- If your organisation has an approved assistant covered by a BAA, use it for anything that cannot be fully de-identified.
- Treat the differential checklist as a memory aid. It can leave out the right diagnosis, and it never replaces examining the patient.
- Ask the model to list every number it used, then compare that list with the source. Unit slips (mg versus mcg) are the classic error.
- Codes, guideline statements and drug interactions suggested by a model must be verified in your coding tool, the guideline itself or a drug reference.
Frequently asked questions
Can I paste a patient's chart into ChatGPT?
Not unless your organisation has approved that tool for patient data. In the US that means a business associate agreement under HIPAA, which consumer chat apps generally do not provide. Remove identifiers first, paste only the excerpt you need, or use your approved assistant.
Does the de-identification check make a note HIPAA compliant?
No. It masks common identifier formats from the Safe Harbor list, such as MRNs, dates, phone numbers and titled names, but it cannot find a name written without a title or rare details that identify someone. You remain responsible for checking the text.
Will the prompts make the AI diagnose or prescribe?
They are written to stop that. The role tells the model it drafts for a physician and does not make clinical decisions, and the differential template asks for a checklist with supporting and opposing features, not a diagnosis or probabilities.
How do I stop the model inventing lab values or doses?
Every prompt tells it to use only the facts provided, write [MISSING] where something is absent and copy values exactly. Still compare each number in the draft with your source before signing.
Can the PICO template find studies for me?
It builds the question, MeSH and free-text terms and a search string. Run the search in PubMed or the Cochrane Library yourself; models often describe studies inaccurately from memory.
Is anything I type stored or sent to a server?
No. The prompt is assembled in your browser. Only your starred favourites are saved, in this browser's local storage. Nothing reaches an AI service until you paste the prompt there yourself.
Which assistant works best with these prompts?
The prompts are plain text with headed sections, so they work with ChatGPT, Claude, Gemini, Copilot and most hospital-approved assistants. The open-in buttons only open the site; you paste the prompt.
Is there an in-depth guide to AI Prompts for Doctors?
Yes. “AI for Doctors and Nurses: Safe Prompts for Notes, Letters and Teaching” covers the background, how it works, common pitfalls and worked examples. Read the full guide.
Related Tools You May Like
Free, in-browser tools that pair well with this one.