Clinical Prompt Library — Safe AI Prompts for Healthcare

🟢 Green Prompts — Safe to Use Now

Patient Education One-Pager

🟢 Green

"Write a patient-friendly explanation of [CONDITION] at a reading age of about 12. Cover: what it is, common symptoms, standard treatment options, and 3 questions patients should ask their doctor. No dosing, no individualised advice. One-page handout format."

✓Input: condition name only — never a real patient’s details.

!Before use: verify clinical accuracy against your local guidance.

✗Never for: individualised treatment instructions.

Patient CommsNursingGP

Journal Article Summary

🟢 Green

"Summarise this abstract for a practising [SPECIALTY] clinician: study design, population, primary outcome with effect size, key limitations, and what — if anything — should change in practice. Flag if the conclusion overstates the data. [PASTE ABSTRACT]"

✓Input: published text only — no patient data, no clinical records.

!Before use: read the primary outcome yourself before citing it anywhere.

ResearchCPD

Teaching Case Generator

🟢 Green

"Create a fictional teaching case for [LEARNER LEVEL] on [TOPIC]: presentation, 3 plausible differentials with discriminating features, the key teaching point, and 2 MCQ-style questions with explanations. Fully fictional — not based on any real patient."

✓Input: topic and learner level only — fully fictional, no real patients.

✗Never: “base this on a patient I saw” — even “changed” details may still be identifiable under GDPR.

EducationTraining

Infection Control Staff Notice

🟢 Green

"Write a clear, plain-English staff notice about [INFECTION CONTROL MEASURE — e.g. hand hygiene protocol, PPE requirements, isolation precautions]. Audience: ward/care home staff. Include: what to do, why it matters, and who to contact with questions. Short, no jargon."

✓Input: the policy topic only — no patient or staff personal data.

!Before use: check against your organisation’s current IPC policy before distributing.

IPCNursingCare Homes

Care Home Activity Programme Draft

🟢 Green

"Create a one-week activity programme for a care home serving residents with [GENERAL PROFILE — e.g. mixed dementia and frailty, mainly ambulatory]. Include: morning, afternoon and evening options. Balance: social, cognitive, physical and creative activities. Note any safety considerations per activity type. No resident names or personal details."

✓Input: general resident profile only — no individual names, histories or care plans.

!Before use: activities lead reviews suitability for actual resident group before implementation.

Care HomesActivitiesWellbeing

Family General Update Letter Template

🟢 Green

"Write a warm, professional letter template for a care home to send to families. Topic: [GENERAL UPDATE — e.g. seasonal activities, staffing changes, visiting policy update]. Tone: reassuring and clear. Leave [RESIDENT NAME] and [FAMILY NAME] as blank placeholders. One page."

✓Input: topic only — all names and personal details added manually after, never to the AI tool.

✗Never: include a resident’s actual health status, care plan details or personal history.

Care HomesFamily Comms

Clinical / Staff Meeting Agenda

🟢 Green

"Create a structured agenda for a [MEETING TYPE — e.g. ward governance meeting, care home MDT, nursing handover review] lasting [DURATION]. Include: standing items, agenda format, time allocations, and a space for any other business. No attendee names needed — I’ll add them."

✓Input: meeting type and duration — no staff or patient names.

!Before use: add your organisation’s standing agenda items before circulating.

AdminMDTCare Homes

Clinical Role Job Description Draft

🟢 Green

"Draft a job description for a [ROLE TITLE] in a [SETTING — e.g. community hospital, care home, GP practice] in [COUNTRY]. Include: key responsibilities, essential qualifications, desirable experience, and reporting line. Professional, plain English, no jargon."

✓Input: role details only — no applicant or employee personal data.

!Before use: HR or line manager reviews before advertising — AI may miss mandatory regulatory requirements.

HRAdminCare Homes

🟡 Yellow Prompts — Use With the Input Rules

Referral Letter Skeleton

🟡 Yellow

"Draft a referral letter from [REFERRING SPECIALTY] to [CONSULTANT SPECIALTY]. Reason: [REASON]. Relevant history: [DE-IDENTIFIED SUMMARY — no names, dates of birth, addresses or identifying details]. Specific question for the consultant: [QUESTION]. Formal, concise, one page."

✓Input: placeholders and de-identified summaries only. Identifiers are added after the draft, inside your clinical system.

!Before use: full clinical read-through — the AI has not seen the chart and cannot be trusted on completeness.

✗Never: paste chart exports or identifiable patient records into any consumer AI tool.

LettersGPHospital

Discharge / Aftercare Instructions Rewrite

🟡 Yellow

"Rewrite these discharge instructions at a reading age of 12, without changing any clinical content. Keep all medication names, doses, and follow-up timing exactly as written. Flag anything ambiguous rather than guessing. [PASTE DRAFT — de-identified, no patient name or DOB]"

✓Input: de-identified draft instructions only — remove patient name and DOB before pasting.

!Before use: line-by-line check that no dose, medicine name or follow-up timing changed. This is the known failure mode.

DocumentationNursing

Care Home to GP Letter Skeleton

🟡 Yellow

"Draft a letter from a care home to a GP regarding a resident’s [CONCERN — e.g. change in behaviour, medication review request, wound assessment request]. Background: [DE-IDENTIFIED CLINICAL SUMMARY — no name, DOB or address]. Request: [SPECIFIC ASK]. Formal, one page."

✓Input: de-identified clinical summary only — add resident name/DOB in your care system after drafting.

!Before use: senior nurse or care manager reviews before sending.

✗Never: paste care plan records or medication administration logs into the AI tool.

Care HomesLettersGP Liaison

Ward Handover Note Structure

🟡 Yellow

"Create a SBAR-format handover note structure for a [WARD/UNIT TYPE] patient with [DE-IDENTIFIED CLINICAL SITUATION — no name, DOB or MRN]. Situation, Background, Assessment, Recommendation. Flag any urgent actions clearly."

✓Input: de-identified clinical situation only — all patient identifiers added in your clinical system.

!Before use: handover clinician verifies clinical accuracy before using in any handover.

NursingDocumentationSBAR

Medico-Legal / Insurance Report Skeleton

🟡 Yellow

"Draft the structure and neutral wording for a [REPORT TYPE — e.g. insurance medical report / occupational health summary] based on this de-identified case summary: [SUMMARY — no identifiers]. Flag any section where clinical judgement or examination findings must be inserted by the clinician. Formal register."

✓Input: de-identified summary only — these reports carry legal weight.

!Before use: the AI drafts structure; the clinician supplies every substantive finding.

✗Never: let the model assert clinical findings you did not make yourself.

LegalReports

Care Plan Section Draft

🟡 Yellow

"Draft the [SECTION — e.g. personal care, nutrition, mobility, communication] section of a person-centred care plan for a resident with [DE-IDENTIFIED NEEDS SUMMARY — no name, DOB or care home]. Use person-first language. Include: current status, goals, interventions, and review date placeholder."

✓Input: de-identified needs summary only — resident name and identifiers added in your care system.

!Before use: key worker or senior carer reviews and personalises before the care plan is finalised.

✗Never: paste existing care plan records with resident identifiers into the AI tool.

Care HomesCare PlanningNursing

Incident Report Draft

🟡 Yellow

"Help me structure an incident report for the following de-identified event: [WHAT HAPPENED — no staff names, no resident names, no dates that could identify an individual]. Include: description of the event, immediate actions taken, contributing factors, and recommendations. Factual, no speculation."

✓Input: de-identified event description — names and exact dates added in your reporting system.

!Before use: line manager or governance lead reviews before submission. AI cannot assess causation.

✗Never: use AI output as the final submitted incident report without human sign-off.

Care HomesGovernanceIncidents

Nursing Assessment Documentation Helper

🟡 Yellow

"Structure a nursing assessment note for a patient presenting with [DE-IDENTIFIED PRESENTING COMPLAINT AND OBSERVATIONS — no name, DOB or ward]. Use standard nursing assessment headings: Airway, Breathing, Circulation, Disability, Exposure. Flag any gaps where I need to add clinical findings."

✓Input: de-identified clinical observations only.

!Before use: the nurse completes all clinical findings before this enters the patient record.

NursingDocumentationABCDE

Complaint Response Letter Draft

🟡 Yellow

"Draft a compassionate, professional response to a complaint about [DE-IDENTIFIED COMPLAINT SUMMARY — no complainant name, no resident name, no identifying dates]. Acknowledge the concern, describe the investigation approach, and outline next steps. Tone: empathetic and clear. Leave [NAME] as a placeholder."

✓Input: de-identified complaint summary only — names and specifics added manually after drafting.

!Before use: manager or PALS lead reviews — never send without sign-off and legal check if required.

Care HomesLettersComplaints

🔴 Red Prompts — These Need a Governed Protocol, Not a Prompt

“Summarise this patient’s chart and write my note”

🔴 Red

The most-used clinical AI workflow in Europe — and the one you cannot run ad-hoc.

Why it’s red: Identifiable health data flowing to a third-party AI model is processing special-category data under GDPR Art. 9 without a lawful basis. Hallucinations flow into the legal medical record. From August 2026, your organisation carries deployer duties for exactly this workflow.

This needs: de-identification, a processing agreement, human oversight, logging, and a validated protocol — not a prompt alone.

Governed version → GovernedAI.io

Diagnosis Support (“What’s the most likely diagnosis?”)

🔴 Red

Why it’s red: Ad-hoc diagnostic AI use puts an unvalidated system in a role EU law treats as high-risk. If a tool is intended to support diagnosis, it may be medical-device territory (MDR / CE marking). No consumer chatbot is validated for your differential.

Use governed, certified tools — or your own clinical judgement.

Governed version → GovernedAI.io

Medication Decision Support

🔴 Red

Why it’s red: Asking AI to suggest dosing, drug interactions, or prescribing decisions puts an unvalidated model in a safety-critical role. AI hallucinations in medication outputs carry direct patient harm risk.

Use validated clinical decision support tools (BNF, Micromedex, local formulary) — not a general-purpose chatbot.

Governed version → GovernedAI.io

Safeguarding Risk Assessment

🔴 Red

Why it’s red: Safeguarding decisions affect a person’s safety and liberty. AI cannot assess the nuance, relationships, or context that a trained safeguarding lead can. Errors in either direction — false positives or missed risks — carry serious consequences.

Safeguarding assessments must be completed by a trained professional under your organisation’s statutory framework. AI has no role here without a governed, validated, human-in-the-loop protocol.

Care Homes Governed version → GovernedAI.io

Mental Capacity Assessment Support

🔴 Red

Why it’s red: Mental capacity assessments under the Mental Capacity Act / assisted decision-making legislation are legal processes requiring trained professionals and direct observation. An AI tool cannot assess capacity — and generating documentation that implies it did creates serious legal and ethical risk.

The assessment must be conducted in person by a trained clinician. Documentation support requires a governed protocol — not an ad-hoc prompt.

Care Homes Governed version → GovernedAI.io

The 5 Safe-Use Rules

Before you use any prompt above — green, yellow or red — these five rules apply every time.

No identifiers, ever

No patient name, date of birth, address, or detail so rare it identifies by uniqueness. If the patient could recognise themselves from the text, it’s personal data under GDPR — and processing it in a consumer AI tool has no lawful basis.

De-identify before, personalise after

Draft with placeholders. Add patient-specific details only inside your clinical or care system — after the AI has done its work.

You are the validation step

Every output gets a full clinical read before it touches a patient, a record, or a colleague. AI does not know the patient. You do.

Nothing decision-making

No ad-hoc diagnosis, triage, medication decisions, or capacity assessments. EU law treats these as high-risk AI for a reason.

Assume drift

The model changed since last month. A prompt that worked reliably in January is not validated in July. Re-read every output as if it’s the first time.

Know your organisation’s policy

Since February 2025, your employer is legally required to train staff on AI use (EU AI Act, Art. 4). If they haven’t — that’s a conversation worth starting. GovernedAI.io Article 4 training →

Clinical AI Prompts That Won’t Get You
or Your Patients in Trouble

The 5 Safe-Use Rules

Your clinicians found this site because
they’re already using AI.

Clinical AI Prompts That Won’t Get Youor Your Patients in Trouble

The 5 Safe-Use Rules

Your clinicians found this site becausethey’re already using AI.

Clinical AI Prompts That Won’t Get You
or Your Patients in Trouble

Your clinicians found this site because
they’re already using AI.