Module 4 · clinical practice

Cardio prompt architecture

This is the practice module. You'll learn to build prompts according to the fixed formula role + context + task + format + uncertainty, and apply it to six typical cardio scenarios: differential diagnosis for chest pain, editing a discharge letter, patient information at B1 reading level, summarising ECG findings, providing imaging context, and creating teaching for residents.

What you'll learn: how a good clinical prompt is built, five live exercises (do each at least once), and which information you'd better not put in a prompt, regardless of the tool.

Cardio prompt architecture

Lesson 4.1: The five building blocks of a clinical prompt

Goal: one checklist you mentally run through before every prompt.

The five building blocks
1. Role, who is the model playing? (“You are an experienced cardiologist with attention for differential diagnosis …”) 2. Context, what does it need to know? (anonymised case, relevant guideline, earlier consideration) 3. Task, what is the exact requested output? (“Give a top 5 DD with pro/con arguments per item.”) 4. Format, what does the answer look like? (list / table / letter / SBAR / length) 5. Uncertainty, let the model explicitly indicate what it does/doesn't know and what assumptions it makes.

The fifth step in particular is often skipped. Without it, the model defaults to a “know-it-all” tone. Adding one line, “name two assumptions you made and one point you're not sure about at the bottom”, often makes the entire answer much more usable, and puts you in verification mode in advance.

Questions for lesson 4.1

1. Why would you use the five building blocks, role, context, task, format, uncertainty?

2. What's the most important effect of adding “name two assumptions and one uncertainty”?

Lesson 4.2: Provide source, don't ask for source

The most important prompt rule to avoid hallucinated guideline citations.

From module 1 you know: an LLM doesn't “know” the guideline, it has memorised text patterns. As soon as you ask “what does the ESC guideline X say about Y?”, the model fills the form of a guideline citation with content that sounds statistically plausible. Sometimes it's right, sometimes not.

The rule

Want reliable guideline content? Paste the relevant chapter yourself into the prompt. Then the model works with your source, instead of its vague memory.

Example: correct form, wrong content

A model may answer with something like: “See Table 7 of the 2023 ESC AF Guidelines, Section 6.4.1.” That sounds like a real citation, but such a table/section often does not exist. Check within 30 seconds: open the real PDF, search for the section number. Form looks right but content is wrong? Classic hallucination. Hence: provide the source, don't ask for the source.

Two variants of the same question, see what the difference does:

❌ Form A, asking for source
What does the ESC 2023 endocarditis guideline say about the indication for PET-CT in prosthetic valve endocarditis?
✅ Form B, providing source
Below is chapter 5.3 of the ESC 2023 endocarditis guideline (copied from the original): [paste here the exact text of the chapter] Summarise in 5 bullets which modalities are recommended for diagnostics in prosthetic valve endocarditis, and in which clinical situation. Quote each time the exact sentence from the chapter you base it on. If something is not explicitly in the chapter, write “not stated in this fragment”.

Questions for lesson 4.2

3. What's the best prompt strategy to prevent hallucinating guideline citations?

Lesson 4.3: Differential diagnosis for chest pain

Goal: a DD prompt that structures instead of forcing conclusions.

What an LLM must not do: give a definitive clinical conclusion. What it can do well: generate a broad DD, order pro/con arguments, and help you not forget any category. Here's how it goes into a prompt:

Cardio DD template
Role: you are a cardiologist in the ED with attention for broad differential diagnosis. Case (anonymised): [paste here: age range, sex, brief history, complaint presentation, vital signs, brief ECG finding, brief lab. NO name, BSN or unique identifiers.] Task: give a top 6 differential diagnosis, sorted from most to least likely in this presentation. Per item: - Pro arguments (from the case) - Con arguments (from the case) - One additional investigation that can quickly confirm or exclude this diagnosis Format: a table with 4 columns. Uncertainty: at the bottom, name which 2 pieces of additional information would change the DD most, and which 1 possibility you deliberately left out and why.

Two deliberately built-in things:

  • An additional investigation per item. Forces the model from “naming diagnosis” to “how do you proceed?”. Usable in practice.
  • “Which information would change the DD most?” That's the question you'd also ask a good resident, and it keeps the conversation on “what don't we know” instead of “what's the answer”.

Teaching rule before you paste

End every DD prompt with: which two additional pieces of information would change this DD most? That forces the model to make uncertainty explicit and helps you see whether the reasoning holds.

Live exercise 1: DD for chest pain

Paste the whole template into the chat and supplement the case with this anonymised data:

Role: you are a cardiologist in the ED with attention for broad differential diagnosis. Case (anonymised): Man, ca. 50 years old, ex-smoker, hypertension. Since 90 minutes retrosternal pressure radiating to left arm, no relief after 2x sublingual nitro. BP 145/90, HR 95, SpO2 96%. ECG: T-wave inversions V2–V5, no ST elevations. First hs-troponin slightly elevated above 99th percentile. Not diabetic, no family history of premature CV. Task: give a top 6 differential diagnosis, sorted from most to least likely in this presentation. Per item: - Pro arguments (from the case) - Con arguments (from the case) - One additional investigation that can quickly confirm or exclude this diagnosis Format: a table with 4 columns. Uncertainty: at the bottom, name which 2 pieces of additional information would change the DD most, and which 1 possibility you deliberately left out and why.
Open in ChatGPT ↗

Points to consider when assessing: (1) is NSTEMI at the top with correct pro/con? (2) do aortic dissection, pulmonary embolism and pericarditis appear? (3) what does the model name as “what would change the DD most”, does that match what you'd want to know?

Lesson 4.4: Editing a discharge letter

Goal: have a letter's form and tone polished without shifting clinical content.

This is one of the high-frequency wins: turning a first draft or rough notes into a clean discharge letter for the GP. Important: your clinical outcomes, doses, follow-ups are in the prompt. The model rewrites form, not content.

Discharge letter template
Role: you are a cardiologist writing a short, professional discharge letter to the GP. Below are my raw points / dictation (anonymised): [paste here as bullets] Make a clean discharge letter from this in this structure: 1. Brief greeting + reason for admission 2. Main findings (in 1 paragraph) 3. Diagnosis/diagnoses 4. Management in the hospital 5. Medication at discharge (bullets with dose/frequency) 6. What the GP must do (bullets) 7. Follow-up / outpatient appointment Rules: - Do NOT change doses, numbers or medication names that are in my dictation. - Do NOT add clinical findings I did not mention. - Keep it concise: maximum 350 words. - Neutral, professional tone. No enthusiasm words. - End with “Kind regards, <my name>” (leave <my name> literally, I'll fill it in later).

Four rules you must always include

  1. Don't change doses, numbers, or medication names from your dictation.
  2. Don't add clinical findings you didn't mention.
  3. Fixed length (forces 'compression' instead of addition).
  4. Fixed tone (otherwise you accidentally get “the patient is in fantastic health”).

Questions for lesson 4.4

4. Which rule keeps the model from “accidentally changing” doses as much as possible?

Live exercise 2: Discharge letter from raw points

Paste the template with these anonymised raw points:

Role: you are a cardiologist writing a short, professional discharge letter to the GP. Below are my raw points / dictation (anonymised): - Woman, ca. 70 years old - Admission: heart failure decompensation in pre-known HFrEF, EF ca. 30% - Reason for admission: dyspnoea class III, total weight +4 kg in 2 weeks - Findings: pulmonary congestion, no leg oedema, lab Na slightly low, NT-proBNP markedly elevated - IV furosemide 80-40-40 mg pump started, then oral furosemide 40 mg morning - HFrEF therapy optimised: bisoprolol 5 mg, sacubitril/valsartan 49/51 BID, dapagliflozin 10 mg started - MRA not resumed due to hyperkalaemia 5.4 - Echo repeated: EF 32%, no new wall motion abnormalities - Diagnosis: decompensated HFrEF - Outpatient follow-up 2 weeks - GP: blood pressure + potassium + creatinine after 1 week Make a clean discharge letter from this in this structure: 1. Brief greeting + reason for admission 2. Main findings (in 1 paragraph) 3. Diagnosis/diagnoses 4. Management in the hospital 5. Medication at discharge (bullets with dose/frequency) 6. What the GP must do (bullets) 7. Follow-up / outpatient appointment Rules: - Do NOT change doses, numbers or medication names that are in my dictation. - Do NOT add clinical findings I did not mention. - Keep it concise: maximum 350 words. - Neutral, professional tone. No enthusiasm words. - End with “Kind regards, <my name>”.
Open in Claude ↗

Check literally: are the doses unchanged? Has nothing been added that you didn't mention? Is the tone professionally neutral (no “patient has recovered”-type phrasing)?

Lesson 4.5: Patient information at B1 reading level

Goal: explain a complex diagnosis clearly, without knowledge risk.

LLMs are strikingly good at simplifying. For a simple leaflet for a patient with new atrial fibrillation, a well-composed prompt is enough to get a first draft, that you'll clinically correct in a few minutes.

What B1 means

Average secondary education level, short sentences (max 15 words), active voice, no jargon without explanation, examples from everyday life. An LLM picks up this level quickly when you ask explicitly, but without specification you often get C1 level (too complex for a leaflet).

Patient information template
Role: you are a cardiologist writing a short patient leaflet. Topic: [fill in: e.g. newly discovered atrial fibrillation in adults, first outpatient appointment] Target group: adults with basic reading skills. Write at B1 level: - Short sentences (max 15 words) - Active verbs - No jargon without direct explanation - No English/medical terms that aren't generally known Structure (in this order): 1. What is <topic> in 3 sentences 2. How it feels or doesn't feel (what you may notice) 3. Why we treat it 4. What our agreements are (medicines, follow-ups, lifestyle) 5. When you must call immediately (3 alarm signals) 6. Space for questions, invitation to write down your questions Limitation: max 400 words. One topic, so no DD or rare complications. Uncertainty: at the bottom, in 1 sentence, name which medical detail you deliberately left out because it would be confusing in this leaflet.

Questions for lesson 4.5

5. How do you consistently lower the language level of an AI-generated leaflet?

Live exercise 3: Leaflet for new AFib

Role: you are a cardiologist writing a short patient leaflet. Topic: newly discovered atrial fibrillation in adults, after first outpatient appointment at which a DOAC has been started and rate control has been agreed. Target group: adults with basic reading skills. Write at B1 level: - Short sentences (max 15 words) - Active verbs - No jargon without direct explanation - No medical terms that aren't generally known Structure (in this order): 1. What is atrial fibrillation in 3 sentences 2. How it feels or doesn't feel 3. Why we treat it 4. What our agreements are (medicines, follow-ups, lifestyle) 5. When you must call immediately (3 alarm signals) 6. Space for questions, invitation to write down your questions Limitation: max 400 words. One topic. Uncertainty: at the bottom, in 1 sentence, name which medical detail you deliberately left out because it would be confusing in this leaflet.
Open in Gemini ↗

Assess: sentence length, jargon-free terms, are alarm signals clinically correct? Is what it 'deliberately left out' truly confusing in this setting (e.g. ablation options at first leaflet) or would you yourself have mentioned it?

Lesson 4.6: Imaging and ECG, what to put or not to put to the LLM

Goal: realistic expectation of a text LLM with ECG and echo/CMR.

Today's text LLMs (ChatGPT, Claude) are not ECG interpretation models. Separate AI systems exist for that (PMcardio, Cardiologs, etc.), trained on millions of ECG signals. What a text LLM does do well:

  • Summarise an ECG report in understandable language.
  • Develop differential diagnosis after you have named the ECG findings in text.
  • Order an echo report by abnormalities + severity.
  • Create patient explanation for an imaging outcome.

What it shouldn't do: directly interpret an ECG image as basis for clinical management. Multimodal models (GPT-4 vision, Claude 3.5/4 with image) do see an ECG graphically, but make errors that a specialised ECG AI system or cardiologist wouldn't. For teaching or a second opinion that thinks along, fine, for management: not.

The practical division of labour

You or an ECG AI system interpret the ECG. The LLM then helps with language: summary, DD, patient information, letter.

Questions for lesson 4.6

6. What's a sensible role division between LLM and cardiologist around an ECG?

Live exercise 4: Order echo report for MDT

Paste this prompt into Claude Sonnet 5.5:

You are a cardiologist preparing an MDT valve surgery meeting. I give you raw, unordered echo findings from a fictional patient. Turn these into a structured MDT block of at most 180 words, in this fixed order: 1. One-sentence key question for the MDT. 2. LV (function, EF, dimensions, wall motion). 3. Valves (focus on the pathological valve, grade, mechanism, LVOT/annulus if relevant). 4. RV and pulmonary pressure. 5. Other relevant (atria, pericardium, aorta). 6. Conclusion in one sentence, without treatment advice. Rules: - Use only information from the input. If something is missing, write explicitly "not stated" and do not invent values. - Do not explain abbreviations; this is for cardiologists. - No treatment proposal, no surgery indication, no risk score. Raw findings: Male 74 yrs, history of HT, DM2. TTE today. LV normal dimensions, EF visually 55-60%, no regional wall motion abnormalities. Aortic valve tricuspid, heavily calcified, Vmax 4.6 m/s, meanPG 52 mmHg, AVA 0.8 cm2 by continuity, DI 0.22. LVOT 2.1 cm. Mild AI. MR grade 1, degenerative appearance. RV normal function, TAPSE 21 mm. PASP estimated 38 mmHg. LA volume indexed 42 ml/m2. No pericardial effusion. Ascending aorta 38 mm. Referrer question: severity of AS and suitability for intervention.
Open in Claude ↗

Paste into Claude Sonnet 5.5. Check that the output keeps exactly the six blocks and stays within 180 words. Then change one finding in the raw results (for example EF to 35% or AVA to 1.4 cm2) and see whether the key question and conclusion move with it. Note one sentence of what you would still adjust for your own MDT format.

Lesson 4.7: Summarising literature and guidelines

Goal: get an ESC guideline or paper into an MDT/CAT format, with source.

Deliberately combine here three things from earlier modules: provide source (module 4.2), question at the bottom (module 2.4), and uncertainty explicit (module 4.1).

Guideline summary template (with source)
Role: you are a scientifically trained cardiologist summarising a guideline chapter for an MDT. Below is chapter [X.Y] of the [guideline name, year], copied verbatim: [paste here the complete chapter] Task (at the bottom): Make a summary in this layout: 1. Main question of this chapter (in 1 sentence) 2. What is the general recommendation? (class / level of evidence if stated) 3. Sub-recommendations, bullets, each with the exact sentence from the text you cite 4. Differences from previous version (only if stated in the text, otherwise “not stated”) 5. Practical implication for our department, in 2 sentences Strict: - Cite ONLY what is in the above text. - If something is not explicitly stated: write “not stated in this fragment”. - Do NOT invent table or paragraph numbers. Uncertainty: at the bottom, name which two questions your MDT should ask to safely implement this chapter.

Live exercise 5: Summarise your own guideline chapter

Grab a chapter of an (openly available) ESC guideline you need today. Copy the entire chapter and paste it into the template. Then test for contradiction: ask the model explicitly “which of your bullets might a fellow cardiologist dispute?”

Open in Gemini ↗

For long guidelines (60+ pages): Claude (1M context) offers a smoother route than ChatGPT. For loose chapters, both work fine.

Lesson 4.8: Creating teaching for residents and trainees

Goal: quickly generate a case + 5 learning questions for teaching moments.

Cardiologists provide a lot of teaching. Making a good case + multiple choice questions takes 30–60 minutes. With a good template, a first draft is done in 5 minutes, that you then clinically refine.

Case + learning questions template
Role: you are a cardiology trainer. You're making teaching for second-year residents. Topic: [fill in: e.g. interpretation of first-hour hs-troponin in NSTE-ACS] Make below: 1. A short realistic case (~150 words, anonymised, first presentation) 2. Five multiple choice questions (a/b/c/d), in ascending difficulty: - Question 1: factual / knowledge - Question 2: ECG / lab interpretation - Question 3: management / clinical decision - Question 4: pitfall / typical error - Question 5: explanation to patient 3. Per question: the correct answer + 2–3 sentences of feedback why this is correct and why the other options aren't, suitable to show a resident. Strict: - Use only mainstream clinical standards (no exotic rare diagnoses) - Do NOT give guideline citations with paragraph/table numbers (to prevent hallucination) Uncertainty: at the bottom, name which question you as trainer would probably rewrite, and why.

Questions for lesson 4.8

7. Why would you explicitly put “give no guideline citations with paragraph/table numbers” in a teaching prompt?

Lesson 4.9: What you don't put in your prompt

Goal: a short red list to remember every working day.

For the extensive privacy explanation with the traffic-light model: see module 5. Here already the shortest summary, to be a rule of action already for the exercises in this module:

Not in a public chat (ChatGPT.com, Claude.ai, Gemini, DeepSeek chat)

  • Patient name, BSN, date of birth, exact address or contact details.
  • Full text of a record note or letter with the above data.
  • DICOM files, original ECG trace files, photos with a recognisable patient.
  • Combinations of rare diagnosis + location + timestamp that make the patient re-identifiable.

What is usually OK (at your own responsibility)

  • Anonymised cases: age range, sex, complaint, findings, without identifiable details.
  • Your own raw points / dictation without patient names.
  • Entirely fictional cases.
  • Your own letters from which you have first removed everything personal.

For “own environment” (hospital-approved Copilot/Claude Enterprise/Azure OpenAI with DPIA), wider rules often apply, ask ICT and the Data Protection Officer.

Lesson 4.10: Let AI interview you for a DD helper

Goal: in five to ten minutes get a plan for a small tool that supports your differential-diagnosis thinking, without taking it over.

You can ask an LLM to build a tool straight away. Better: briefly say what you want (for example support for DD formation), and ask the AI to interview you. That forces boundaries: no diagnosing, but structure, checklists and marking missing information. Use only fictional or anonymised case elements.

Talking is allowed. ChatGPT Voice/Live works on the free tier with a 24-hour limit (often a lighter Live variant). Gemini Live is often free with limits. Claude has less voice. If Voice fails, type. The prompt works for both.

Live exercise 6: Interview for a DD supporter

Open Claude or ChatGPT. Paste. Answer one question at a time. Then ask for the plan; only build once the plan is right.

I am a cardiologist and want a small tool that supports differential diagnosis (DD) and organising clinical reasoning. It must not state a diagnosis or prescribe policy. Only structure, mark missing info, and a clean DD list with one-sentence rationale per item. No patient data; only fictional or anonymised case elements. Do the following, in this order: 1. Say in one sentence what you will do: you will interview me for 5-10 minutes to make the plan, then deliver a short build plan. 2. Ask me one question at a time. Wait for my answer. Maximum 8 questions. 3. Cover: which setting (clinic, ED, MDT), which input fields, which output structure, how you enforce "not stated", how you stop the model inventing policy, whether v1 is a chat prompt or a local HTML artifact. 4. After the questions: plan of max 200 words (goal, non-goals, fields, output, safety rules). 5. Ask whether I want you to build the prompt template or HTML artifact next. No exclamation marks. No marketing tone. Start now with step 1.
Open in Claude ↗ Open in ChatGPT ↗

Check that the plan explicitly says “no diagnosis, no policy”. If Voice is available: talk for 5-10 minutes and then ask for the written plan. Only build after that.

Take-home from module 4

Five building blocks

Role + context + task + format + uncertainty. One line of uncertainty makes every answer 30% more usable.

Provide source

Not “what does the guideline say?”, but: here is chapter X, only cite what's here. Anti-hallucination rule #1.

LLM = language

The LLM is a language machine for letters, summaries, leaflets, DD structure, teaching. ECG/echo interpretation stays with cardiologist or specialised AI system.

Interview first, then build

For a DD helper: let the AI question you for 5-10 minutes, check the plan (no diagnosis/policy), only then build.

← Module 3 On to module 5 →