AI Prompts for Healthcare: Clinical Notes & Patient Comms

AI prompts for healthcare work when you draft from clinician-owned facts: redacted visit notes, approved scripts, and a fence that blocks diagnosis, treatment plans, and medical advice the model invents. You leave with ROLE / TASK / FORMAT shapes for clinical note scaffolding and patient communication macros for GPT-5.5 Instant, GPT-5.6 Sol, Claude Sonnet 5, or Gemini 3.5 Flash, plus [UNKNOWN] and [CLINICIAN REVIEW] markers. This guide covers documentation structure and portal wording. It is not a diagnostic tool, triage engine, or source of medical advice. Soft tip: PromptMake /text can turn a rough documentation ask into a labeled scaffold at https://promptmake.net/text before you paste into your chat model.

Who AI prompts for healthcare help

You need first-pass structure on a progress note, a discharge instruction packet, or a patient portal reply, and you want the model to stay inside facts and scripts you supplied. The patterns fit clinicians who turn dictation bullets into SOAP-shaped drafts, care coordinators who rewrite visit-prep and follow-up messages from approved language, and clinic ops leads who want reusable macros that never invent a diagnosis or a dose. Medical scribes and documentation coaches who standardize section headers land here too, as long as a licensed clinician owns every clinical claim before the chart or message goes out.

This page is for note scaffolding and patient-facing wording. Do not use it to diagnose, suggest differentials, recommend drugs, interpret labs as disease, or build a symptom checker. Those asks turn a drafting clerk into fake clinical tooling. Treat the model as a writing assistant with hard limits. You supply SOURCE_PACK, SCRIPT_PACK, audience, and banned claims. The model proposes section order, clearer verbs, and consistent headers. You reject anything that invents a diagnosis, a medication, a dose, a prognosis, or a care plan.

HIPAA and your employer rules still govern protected health information (PHI). Check your organization's AI policy before you paste visit content into a consumer chat. Redact names, MRNs, dates of birth, phone numbers, addresses, and unique facts that would identify a person. Use Patient A / Visit Date labels. Keep the official chart in your EHR. Consumer ChatGPT, Claude, or Gemini is not a medical record system.

HIPAA and safety fences before you prompt

Name the honesty rule in the first block of every healthcare thread. Patients, boards, and payers grade your chart and messages, not the model's fluency. A polished SOAP note from a blank "write the note" prompt fails the moment a peer asks which symptom you documented and which claim the model invented. Put the fence in ROLE and repeat it in a REMINDER line so a long paste does not bury it.

Allowed artifacts: clinical note drafts from facts you pasted, discharge or referral letter wording from clinician-authored content, patient portal macros from SCRIPT_PACK you approved, tone and plain-language passes on text a clinician already wrote, and audits of a draft against SOURCE_PACK. Banned artifacts: differential diagnosis lists, symptom-to-disease mapping, drug or dose recommendations, triage scores, "what should I tell the patient about their condition," imaging or lab interpretation as diagnosis, and any prompt that asks the model to decide what the patient has.

Keep a SCRIPT_PACK outside the chat: house voice lines, approved after-visit phrases, portal reply templates, banned medical claims, escalation contacts for clinical questions, and phrases legal or compliance banned. Keep SOURCE_PACK as redacted visit bullets the clinician typed. Update both when policy or templates shift. Feed them into every prompt. The model should never be your only source of clinical content or your PHI store.

If your organization forbids consumer tools for PHI, stop and use the approved vendor or EHR-native scribe. The prompt shapes still apply inside that vendor. This article does not create a HIPAA-compliant workflow by itself. Your covered entity's policies, BAAs, and EHR controls do that work.

Core prompt pattern for notes and patient comms

Strong AI prompts for healthcare give five inputs before any tone request: the source pack, the script pack, the deliverable, the audience, and the honesty fence. The source pack holds redacted visit facts, vitals the clinician recorded, meds the chart already lists, and open questions marked with a question mark. The script pack holds approved patient-facing phrases and section headers your clinic uses. The deliverable names the artifact: SOAP draft, discharge instruction rewrite, portal reply, or visit-prep message. The audience is chart, patient portal, or referral recipient. The honesty fence forbids invented diagnoses, meds, doses, prognoses, and advice where SOURCE_PACK is silent.

Paste those blocks near the top. Put format rules next. Repeat the fence at the end so it survives a long paste. GPT-5.5 Instant, GPT-5.6 Sol, Claude Sonnet 5, and Gemini 3.5 Flash follow labeled blocks. Vague "write a clinical note" prompts produce generic essays because the model has no fact pool and fills gaps with textbook disease talk. Aim for one artifact type per thread. Mixing a SOAP draft and a portal reply in the same pass muddies tone and invents clinical claims to fill empty sections. Keep SOURCE_PACK and SCRIPT_PACK outside the chat in plain text. Update them when the clinician edits the chart facts. Feed them into every prompt that touches the draft.

Role, task, format skeleton

Copy this skeleton and fill the brackets with your material:

ROLE: You are a healthcare documentation assistant for [specialty / clinic]. You draft from SOURCE_PACK and SCRIPT_PACK only. You never invent diagnoses, differentials, medications, doses, prognoses, lab interpretations as disease, or medical advice. You do not practice medicine. You produce a draft for a licensed clinician to review.

TASK: Turn SOURCE_PACK and SCRIPT_PACK into a [SOAP note draft | progress note | discharge instructions | referral letter | patient portal reply | visit-prep message] for [audience]. Keep every clinical fact tied to SOURCE_PACK. Keep patient-facing promises tied to SCRIPT_PACK. Mark gaps with [UNKNOWN]. Mark any line that needs clinician judgment with [CLINICIAN REVIEW].

FORMAT: Use the section headers listed under OUTPUT_SHAPE. Short sentences. No new ICD codes, drug names, or doses unless they appear in SOURCE_PACK. If a fact is missing, write [UNKNOWN] and do not guess.

AUDIENCE: [chart note | patient portal | referral recipient | after-visit summary]

SOURCE_PACK: [paste redacted clinician facts, listed meds already on chart, vitals recorded, open questions marked ?]

SCRIPT_PACK: [paste approved phrases, section headers, banned claims]

OUTPUT_SHAPE: [list required sections]

RULES: No diagnosis. No differential. No treatment plan the clinician did not supply. No "you likely have." No invented PII. This output is not medical advice and is not a signed chart entry.

REMINDER: Never invent diagnoses, meds, or advice. Use [UNKNOWN] for gaps and [CLINICIAN REVIEW] for clinical conclusions.

That reminder line stops confident fake medicine. Models love tidy Assessment sections. Your rule forces a gap marker the clinician fills after they decide the clinical content.

Honesty tokens and banned clinical claims

Build a short SAFETY_FENCE block you paste under RULES every time:

SAFETY_FENCE: Forbid new diagnoses, differential lists, drug names, doses, routes, imaging impressions as disease labels, lab "consistent with" claims, prognosis statements, and triage recommendations unless they appear verbatim in SOURCE_PACK. Prefer [UNKNOWN: clinical fact] over a confident guess. Prefer [CLINICIAN REVIEW] on Assessment, Plan, and any advice to the patient. If SOURCE_PACK lacks a fact needed for a section, write [UNKNOWN] and leave that subsection empty of invented content.

Use two tokens and keep them distinct. [UNKNOWN] means the pack has a hole: missing vital, missing listed med, missing instruction the clinician meant to add. [CLINICIAN REVIEW] means the text states a clinical conclusion, plan item, or patient instruction a human must own. Example rule: "Do not write an Assessment diagnosis. If OUTPUT_SHAPE includes Assessment, list only facts from SOURCE_PACK as bullets and mark the section [CLINICIAN REVIEW]."

For patient messages, require the model to copy only language from SCRIPT_PACK or clinician-authored DRAFT. Add: "Quote and reuse only phrases that appear in SCRIPT_PACK or SOURCE_PACK. If you cannot find an instruction, write [UNKNOWN: instruction] and stop inventing home care." That blocks fabricated "take this antibiotic" lines. For names and identifiers, paste only labels you chose (Patient A). Tell the model never to invent MRNs, phone numbers, or addresses.

Clinical notes templates clinicians can reuse

Draft notes in this order when the visit is new: fact bullets into SOURCE_PACK, then section scaffold, then Assessment and Plan only after the clinician supplies those lines. A fact-first pass forces the model to organize what was said and done before it writes conclusions. Reuse the same SOURCE_PACK and SAFETY_FENCE. Change only TASK and OUTPUT_SHAPE. After each draft, run a short audit before anyone pastes into the EHR. Soft tip: if your rough ask is messy, shape the prompt once at https://promptmake.net/text, then fill SOURCE_PACK yourself.

Keep audience labels tight. A chart note can use clinical shorthand the clinician already uses. A patient-facing after-visit summary needs plain language from SCRIPT_PACK and still needs clinician sign-off. Ask the model for a chart scaffold first. Convert to patient language in a second pass only after the clinician approves the clinical content.

SOAP and progress note scaffolding

SOAP scaffold pattern:

TASK: From SOURCE_PACK, draft a SOAP-shaped note under [word cap]. Sections: Subjective (patient-reported items from SOURCE_PACK only), Objective (vitals, exam findings, and listed results from SOURCE_PACK only), Assessment ([CLINICIAN REVIEW]: list open clinical questions as bullets; do not name a diagnosis), Plan ([CLINICIAN REVIEW]: list only plan items present in SOURCE_PACK; mark missing items [UNKNOWN]). Ban differentials, new drug names, and "likely" disease language. If Assessment or Plan content is absent from SOURCE_PACK, leave bullets as [CLINICIAN REVIEW] placeholders.

Progress note pattern:

TASK: Draft an interval progress note. Sections: Interval history (SOURCE_PACK), Current meds as listed in SOURCE_PACK, Exam or findings from SOURCE_PACK, Pending items marked [UNKNOWN] if not supplied, Clinician plan block marked [CLINICIAN REVIEW]. Do not add diagnoses or meds missing from SOURCE_PACK.

Note audit pattern: "TASK: Audit NOTE_DRAFT against SOURCE_PACK and SAFETY_FENCE. List every diagnosis word, drug name, dose, and prognosis claim. Reply Keep, Soften, or Remove. Soften means the claim overreaches the pack. Propose a rewrite that uses [UNKNOWN] or [CLINICIAN REVIEW]. Never add a new diagnosis or medication."

Discharge, referral, and after-visit drafts

Discharge instructions pattern:

TASK: From SOURCE_PACK and SCRIPT_PACK, rewrite discharge instructions for patient reading level [grade]. Sections: What we did today (facts only), Home instructions (SCRIPT_PACK and SOURCE_PACK only), Medications as listed (no new drugs), When to seek care (only phrases from SCRIPT_PACK), Follow-up (dates and contacts from SOURCE_PACK). Mark missing lines [UNKNOWN]. Ban new diagnoses and new meds. Mark the whole packet [CLINICIAN REVIEW].

Referral letter pattern: "TASK: Draft a referral letter to [specialty] from SOURCE_PACK. Sections: Reason for referral in the clinician's words from SOURCE_PACK, Relevant history bullets from SOURCE_PACK, Current meds listed, Questions the referring clinician wants answered (from SOURCE_PACK only). Do not invent a working diagnosis. Do not invent labs. Use [UNKNOWN] for missing items. Mark clinical framing [CLINICIAN REVIEW]."

After-visit summary pattern: "TASK: Convert CLINICIAN_APPROVED_NOTE into plain language for the portal. Keep every clinical claim identical in meaning. Replace jargon with SCRIPT_PACK plain phrases when available. Do not add advice. Do not name a new condition. Flag any line that would need a new clinical decision with [CLINICIAN REVIEW]." A licensed clinician still signs the chart and the message.

Patient communication macros

Patient comms are the other half of AI prompts for healthcare. Notes stay in the chart. Macros cover portal replies, visit-prep reminders, and follow-up wording that must match SCRIPT_PACK. Without a script pack, the model invents "your test shows" and "start this medication" lines that create unsafe care and compliance risk. Paste your real approved phrases first: scheduling language, prep instructions your clinic already publishes, and "message your care team" redirects for clinical questions.

A good macro names channel, tone, required facts, and what the message must refuse. ChatGPT or Claude can turn SCRIPT_PACK into agent-facing templates and into a patient-facing draft a human can send after a two-minute check. You still decide whether this message is appropriate. The model drafts the wording after you name the macro type and paste the facts.

Keep portal work in its own thread when the case is sensitive. Paste SCRIPT_PACK, the redacted PATIENT_ASK, and the macro type. Ask for the patient-facing draft and an internal checklist as two separate outputs so the patient never sees internal severity tags or chart jargon you did not approve.

Visit prep and follow-up messages

Visit-prep pattern:

TASK: From SCRIPT_PACK and VISIT_FACTS, draft a visit-prep message for [channel]. Sections: Appointment time and location from VISIT_FACTS, What to bring (SCRIPT_PACK only), Prep steps (SCRIPT_PACK only), How to reschedule (SCRIPT_PACK). Ban clinical advice. Ban "this visit will diagnose." If a prep step is missing, write [UNKNOWN: prep] instead of inventing fasting or medication-hold instructions.

Follow-up reminder pattern: "TASK: Draft a follow-up reminder. Include date/time from VISIT_FACTS, reason for visit in non-diagnostic wording from SCRIPT_PACK (example: 'follow-up as planned'), and portal reply instructions. Do not state a diagnosis. Do not interpret prior results. Mark any clinical content [CLINICIAN REVIEW]."

Portal replies and clarification macros

Portal reply pattern:

TASK: Draft a patient portal reply under [N] words for PATIENT_ASK. Tone: clear and calm. Use only SCRIPT_PACK for clinical boundaries. If PATIENT_ASK asks for a diagnosis, dose change, or urgent triage, refuse to invent an answer and use the SCRIPT_PACK redirect: message the care team / call the clinic / seek emergency care if SCRIPT_PACK includes that line. Never invent a diagnosis or medication change.

Clarification macro: "TASK: Rewrite DRAFT for plain language at reading level [grade]. Keep every instruction identical. Flag any sentence that introduces a new clinical claim. Ban disease labels not in SOURCE_PACK. Output: Revised message | Flags."

Urgent-symptom fence: "If PATIENT_ASK describes severe symptoms, do not diagnose. Output the SCRIPT_PACK emergency redirect only, plus [CLINICIAN REVIEW]. Do not give home treatment steps beyond SCRIPT_PACK." Pair that fence with human routing. The model drafts words. The clinic owns the clinical response path.

Step-by-step healthcare prompt workflow

Use one chat thread per visit label and artifact: Visit A SOAP scaffold, Visit A portal reply, Visit A discharge packet. Dumping every patient into one long thread mixes facts and invents shared clinical claims. The loop below keeps a stable SCRIPT_PACK while you swap only this visit's SOURCE_PACK. You spend free ChatGPT, Claude, Gemini, or PromptMake runs on structure, then clinician time on chart review and send.

Keep SOURCE_PACK in a local file the clinician controls: redacted bullets, listed meds, vitals, and open questions. Keep SCRIPT_PACK as approved patient phrases. The model should never be your EHR, your formulary, or your diagnostic brain. If your organization forbids consumer tools for PHI, stop and use the approved vendor. The prompt shapes still apply inside that vendor.

Step 1: Redact PHI and build the source pack

List what you may say. Example: "Patient A (redacted). Visit for follow-up. Subjective bullets the clinician typed. Vitals as recorded. Meds as listed on chart. Plan items the clinician already decided: continue current meds as listed, return in 4 weeks. Ban: invented diagnosis, new drugs, dose changes, prognosis." Ugly notes beat a polished fake Assessment.

Mark every clinical conclusion as clinician-supplied or absent. In the prompt, tell the model to keep absent items as [UNKNOWN] or [CLINICIAN REVIEW]. Guessing a diagnosis to fill Assessment creates a chart a peer cannot trust and a patient safety problem.

Step 2: Outline the artifact, then draft

Name the artifact and audience before you paste long notes. Ask for an outline first when the note type is new:

TASK: From GOAL and AUDIENCE, propose an outline with headers only. Do not draft body text yet. Flag any header that needs facts missing from SOURCE_PACK. Refuse if GOAL asks for a diagnosis, differential, or treatment plan the pack does not contain.

Use that outline to decide what to paste next. Partial facts get [UNKNOWN] subsections. Missing clinician decisions stay out of Assessment and Plan until the clinician writes them. Then run the full skeleton with OUTPUT_SHAPE matching the outline you approved.

Step 3: Audit clinical claims, then clinician review

Take the draft into a second message:

TASK: Audit DRAFT against SOURCE_PACK, SCRIPT_PACK, and SAFETY_FENCE. For each diagnosis word, drug, dose, prognosis, and patient instruction, reply Keep, Soften, or Remove. Soften means the claim overreaches the pack. Propose a safer rewrite. Never add new diagnoses or medications.

Optional scaffold: open https://promptmake.net/text, describe "healthcare clinical notes and patient portal macros with HIPAA-aware redaction, [UNKNOWN] gaps, and [CLINICIAN REVIEW] flags; no diagnosis," generate once, then paste your SOURCE_PACK into the returned structure. Guests get about three runs per day; free accounts get about five. Use a run to shape the prompt, then finish in your chat model.

Save the winning prompt next to the artifact name and model label so the next visit reuses the same wrapper. Paste into the EHR or portal only after a licensed clinician checks every clinical claim and owns the note or message.

Mistakes that wreck healthcare AI drafts

Mistake 1: Asking the model to "write the note" or "figure out the diagnosis" with no SOURCE_PACK. The model invents disease labels that sound real and fail the first peer review. Paste controlled clinician facts instead. Never ask for differentials or symptom-to-disease maps.

Mistake 2: Allowing invented meds, doses, and prognoses. If you did not type the plan item, ban it in FORMAT. Fake drug lines create patient harm. Require [UNKNOWN] or [CLINICIAN REVIEW].

Mistake 3: Treating the draft as medical advice or a signed chart entry. Put "Draft for clinician review. Not medical advice. Not a signed note." in ROLE and in the header the model must output. A human still signs the work.

Mistake 4: Using GPT-5.5 Instant or Gemini 3.5 Flash as the final clinical safety check. Fast models fit outlines and first-pass section order. Route the claim audit to GPT-5.6 Sol, Claude Opus 5, or Gemini 3.1 Pro, then still have the clinician read the chart.

Mistake 5: Pasting unredacted PHI into a consumer chat without your organization's AI policy check. Redact. Use Patient A labels. Keep sensitive fields in the EHR.

Mistake 6: One mega-prompt that asks for a SOAP note, a differential, and a patient advice letter together. Split artifacts. Refuse diagnosis tasks. Reuse SOURCE_PACK; change TASK and FORMAT.

Mistake 7: Asking how to diagnose from a symptom list, how to triage without a clinician, or how to "replace" a visit. Refuse that use. Prompt for documentation structure and approved macros only.

Mistake 8: Skipping the audit after a long paste. Narrative notes hide invented Assessment lines. Prompt for Keep / Soften / Remove against SOURCE_PACK every time.

Model notes for healthcare prompts (mid-2026)

ChatGPT often defaults to GPT-5.5 Instant for fast chat. Instant fits note outlines, section scaffolds, and portal macro first drafts when you already locked SOURCE_PACK and SCRIPT_PACK. Keep prompts short: ROLE, TASK, FORMAT, SOURCE_PACK, SCRIPT_PACK, RULES. Skip long chain-of-thought slogans.

GPT-5.6 Sol fits harder edit passes: claim audits against SOURCE_PACK, contradiction checks between Subjective and Objective bullets, and drafts that must not invent Assessment content. Give goal, constraints, and format. Drop "think step by step" padding on reasoning-class models.

Claude Sonnet 5 handles long SOURCE_PACK pastes and tidy SOAP tables well. Claude Opus 5 fits careful audits when the note must not invent plan items. Claude Fable 5 is the widely released top tier when your workspace offers it; check your plan. Haiku 4.5 fits short portal macro reshuffles when latency matters more than a deep audit.

Gemini 3.5 Flash fits volume work: many outline passes and plain-language rewrites from an updated SCRIPT_PACK. Gemini 3.1 Pro fits hard reasoning over long context when you paste a thick fact pack and need conflict flags against SAFETY_FENCE. Hedge on exact menu names in each vendor UI. They shift. Re-check the model picker when you open a new thread.

Prompting split that holds: Instant and Flash get RTF plus a short sample of your house note format when you need matching structure. GPT-5.6 Sol, Opus 5, and Gemini 3.1 Pro get goal + constraints + format, with an explicit "never diagnose" rule, a [UNKNOWN] token, and a [CLINICIAN REVIEW] flag. All need your SOURCE_PACK in the message. None replace a human clinician, an EHR, or a HIPAA-governed workflow.

Build healthcare prompts with PromptMake /text

Write the rough ask in plain words: artifact type (SOAP scaffold, discharge rewrite, portal macro), safety fence, and whether you need a claim audit. Open https://promptmake.net/text and generate a structured prompt once. Expect labeled sections you can fill with SOURCE_PACK and SCRIPT_PACK.

Edit clinical facts and approved phrases yourself. PromptMake cannot know your visit or your formulary. Paste the filled prompt into Instant or Flash for drafts, or GPT-5.6 Sol / Opus 5 / Gemini 3.1 Pro for audits. Keep free-tier runs for scaffolding, not five synonym retries of the same weak "write a medical note" ask.

Workflow that sticks: redact PHI → SOURCE_PACK + SCRIPT_PACK → PromptMake scaffold → fill clinician facts → fast model draft → reasoning model claim audit → licensed clinician review → chart or send only after that review. Store one template per artifact so you do not rewrite ROLE and RULES from scratch each visit.

FAQ

What are the best AI prompts for healthcare in 2026?

The best AI prompts for healthcare lead with ROLE and a safety fence, paste a redacted SOURCE_PACK and an approved SCRIPT_PACK, then demand FORMAT for one artifact: SOAP scaffold, progress note, discharge instructions, referral letter, or patient portal macro. Add a second audit prompt that marks Keep, Soften, or Remove against those packs. Match GPT-5.5 Instant or Gemini 3.5 Flash for drafts and GPT-5.6 Sol, Claude Opus 5, or Gemini 3.1 Pro for claim audits, then still have a licensed clinician own every clinical line.

Can AI write clinical notes from scratch?

The model can draft structure and wording from facts a clinician supplies. It should not invent diagnoses, differentials, medications, or plans. Start from a redacted SOURCE_PACK the clinician typed, and treat blank-slate "write my medical note" prompts as high risk for fiction that fails peer review. A licensed clinician still signs the chart.

Are AI prompts for healthcare safe under HIPAA?

Consumer chat tools are not a HIPAA compliance solution by themselves. Your covered entity's policies, BAAs, and approved vendors decide what you may paste. Redact PHI, use Patient A labels, and prefer organization-approved tools when required. This guide's fences help documentation quality. They do not replace a HIPAA program or an EHR.

How do I stop AI from inventing a diagnosis?

State the ban in ROLE and again in SAFETY_FENCE plus a REMINDER line. Forbid differentials, new disease labels, and "likely" claims, and require [CLINICIAN REVIEW] on Assessment and Plan. Follow with an audit that lists every diagnosis word against SOURCE_PACK. Then have the clinician write or approve clinical conclusions offline.

Can I use AI for patient portal messages?

Yes, when you draft from SCRIPT_PACK and redacted PATIENT_ASK, and a human reviews before send. Use macros for scheduling, prep, and clarification. If the patient asks for a diagnosis, dose change, or urgent triage, use your SCRIPT_PACK redirect and escalate to a clinician. Never let the model invent home treatment beyond approved phrases.

Should I use GPT-5.5 Instant or GPT-5.6 Sol for healthcare prompts?

Use Instant for note outlines, section scaffolds, and first portal macros from packs you already locked. Use GPT-5.6 Sol when you need a careful claim audit, conflict checks between sections, or a draft that must not invent Assessment content. Run the same packs through both only when you measure quality for a recurring template.

Can PromptMake help with AI prompts for healthcare on the free tier?

Yes. PromptMake /text turns a rough documentation or portal ask into a labeled prompt you can aim at ChatGPT, Claude, or Gemini. Guests get about three generations per day; registered free users get about five. Fill in your own redacted SOURCE_PACK and SCRIPT_PACK, then paste into Instant, Flash, or a reasoning model for the audit. PromptMake does not practice medicine, does not store PHI as a chart, and does not diagnose.