The Clinical Question
A field guide for clinicians: how to ask clinical AI so it reasons well. The evidence, the SCOPE method, twelve worked rewrites and a quick-reference card.
Respocare Connect AI Team — Respocare (PTY) Ltd · Reg. 2018/411829/07 · Practice No. 9990900010775614. Licensed healthcare practice operating since 2018.
· 15 min read
A field guide for clinicians · Edition 1.0, September 2026
The models are now strong reasoners. The evidence shows that capability does not reach the patient automatically. It passes through a question, and the question sets the ceiling. This guide is about asking well.
For doctors, nurses and allied health professionals using clinical AI. Twenty minutes to read; seconds per question to apply.
Download the printable guide (PDF)
Eleven pages, including the quick-reference card for your workstation.
01 · Why this guide exists
The machine can reason. The gap is in the asking.
Across the best recent trials, AI models working alone have outperformed the people using them. The capability was present. It did not survive the interaction.
In a 2024 randomised trial, 50 physicians worked through complex diagnostic cases. Those given GPT-4 alongside their usual resources scored a median of 76% against 74% for those without it, a difference that was not statistically significant. The model on its own scored 16 percentage points higher than the physicians using conventional resources.
A 2026 preregistered trial in Nature Medicine found the same gap at scale. Alone, the models identified the relevant condition in 94.9% of scenarios. The 1,298 people using those same models managed it in fewer than 34.5% of cases, no better than a control group. When researchers read the conversations, they found people often gave the model incomplete or incorrect information.
| Study | Finding | What it means for you |
|---|---|---|
| Goh et al. (JAMA Netw Open 2024 · 50 physicians) | Model alone +16pp; physicians with the model 76% vs 74% | Access to a strong model is not the same as benefiting from it |
| Bean et al. (Nature Medicine 2026 · 1,298 people) | Model alone 94.9%; people using it under 34.5% | Incomplete information in produces poor reasoning out |
| Johri et al. (CRAFT-MD) (Nature Medicine 2025 · 2,000 cases) | Significant accuracy drops when exam questions became real conversations | Benchmarks overstate performance on messy, real exchanges |
| Chen et al. (npj Digit Med 2025 · 5 models) | Compliance with illogical drug requests reached 100% | Models tend to accept the premise of your question |
| Goh et al. (Nature Medicine 2025 · 92 clinicians) | Model assistance improved management reasoning by 6.5pp | Deliberate engagement closes the gap. Better performers spent longer |
The principle
A model can only reason toward the goal your question sets. The quality of the answer is decided before the model begins.
This is not a failing peculiar to AI. Clinicians have always known that the history is most of the diagnosis. The same holds here.
02 · How your question travels
What happens between pressing enter and reading the answer
Clinical AI works through stages, and your wording steers every one of them.
Your question → Retrieval → Reasoning → Draft → Your review
| Stage | What the system does | What your wording controls |
|---|---|---|
| Question | Interprets your intent from your words alone. It cannot see what you meant but did not say. | The goal, the scope and the standard of evidence |
| Retrieval | Decides which parts of the record and which sources to pull. | A narrow question retrieves narrowly. An unbounded one lets the system choose what matters |
| Reasoning | Connects what it retrieved. Compares, trends, checks for conflicts. | A buried assumption becomes the starting point of the reasoning |
| Draft | Writes an answer shaped to what it believes you asked. | Length, structure, and whether uncertainty is shown or smoothed over |
Three mechanisms behind a poor answer
1. It answers what you asked, not what you meant — Every decision you leave unstated is made for you, silently. Most disappointing answers are faithful answers to vague questions.
2. It tends to agree with the premise — In one study, models wrote misinformation built on a false drug premise in up to 100% of cases, despite knowing the facts. Explicit permission to refuse sharply reduced this.
3. Small changes in wording change the answer — Prompt style alters accuracy, and the same question does not always produce the same answer. Precision reduces the variance.
Why agentic systems raise the stakes
An agentic assistant retrieves, reasons, drafts, and should decline or escalate when evidence runs out. A flawed question can therefore produce an answer that is fully sourced and still addresses the wrong thing. Sources protect the answer. Only you can protect the question.
03 · Division of labour
You are not a prompt engineer. You are the one who knows the question.
You should not need to write elaborate instructions. A well-built clinical system carries the context. What it cannot carry is your intent.
| The system should carry | Only you can supply |
|---|---|
| Which patient, and their full record | The clinical question you actually need answered |
| Retrieval of the relevant documents and results | The period that matters |
| Allergies, medications and critical results, surfaced by default | The decision this answer will inform |
| A source behind every clinical statement | What you already suspect, and want tested |
| Declining when the record does not support an answer | Permission to be told you are wrong |
The rule of thumb
Brief the AI the way you would brief a capable registrar who has never met the patient. The difference is that the registrar would ask what you meant. The model will often guess.
A note on safety-critical facts
No clinician should have to remember to ask whether a patient will be harmed. A trustworthy system surfaces allergies, interactions and critical results whatever the question. If a tool only mentions an allergy when you ask about allergies, it is relying on your memory, and that is a property worth testing before you rely on it.
04 · The method
SCOPE: the anatomy of a precise clinical question
A precise clinical question has five parts. Most weak questions are missing three of them. You rarely need all five in writing, but you should have considered each.
| Element | Ask yourself | Example | |
|---|---|---|---|
| S | Subject | Which patient, and which part of the record? | Her renal results and current medication list |
| C | Clinical question | What single thing do I need to know? | Has her kidney function changed? |
| O | Over what period | Since when does it matter? | Over the last six months |
| P | Purpose | Which decision does this inform? | Before I review her metformin dose |
| E | Evidence standard | How should uncertainty be handled? | Cite each result and tell me what is missing |
Before
Tell me about her kidneys.
Over the last six months, how has her renal function changed, and which of her current medications are relevant to it? I am reviewing her metformin dose. Cite each result, and tell me if any expected results are missing from the record.
Why each element earns its place
| Element | What goes wrong without it |
|---|---|
| Subject | The system draws from the wrong part of the record, or everything at once |
| Clinical question | You receive a summary when you needed an answer |
| Over what period | Old data crowds out recent change, or recent data hides a long trend |
| Purpose | The answer is accurate but irrelevant to the decision in front of you |
| Evidence standard | Gaps are smoothed into confident prose instead of being flagged |
05 · Worked rewrites
Twelve everyday questions, asked well
Each rewrite applies SCOPE to a common clinical task. Adapt the pattern rather than copying the words. All patients are illustrative.
| Task | Weak | Precise |
|---|---|---|
| Pre-consult preparation | Anything I should know? | Since his last visit in March, what has changed that bears on his heart failure management? Include results, medication changes and admissions, with sources. |
| Record summary | Summarise this patient. | Summarise her active problems and current medications for a colleague covering my clinic. Flag anything unresolved from the last three visits. |
| Allergies | Confirm no allergies. | List every documented allergy and adverse reaction with dates and sources. State explicitly if no allergy record exists. |
| Medication review | Are her meds OK? | Review her current medications for interactions and for doses needing adjustment given her latest eGFR. Show the evidence for each concern. |
| Result trends | Why did his HbA1c get worse? | How has his HbA1c trended over the past year, and what changed in his medication or documented adherence over the same period? |
| Referral | Write a referral. | Draft a cardiology referral focused on his recent chest pain episodes. Include relevant ECG and troponin results, and list what the cardiologist will need that is not in the record. |
| Discharge letter | Do the discharge summary. | Draft a discharge summary for her GP covering the admission diagnosis, key results, medication changes with reasons, and the follow-up plan. Mark anything I need to confirm. |
| Handover | Give me a handover. | Prepare a handover for a colleague who has never met her: active problems, current treatment, pending results and what to watch for overnight. |
| Treatment question | Is this the right treatment? | What in the record supports continuing this treatment, and what argues against it? What information is missing that would change the picture? |
| Guideline check | What does the guideline say? | Which current guideline recommendations apply to a patient with these characteristics, and where does her situation fall outside them? Cite the guideline sections. |
| Outstanding items | Anything outstanding? | List all investigations, screening and follow-ups that were planned or deferred in the past year and not completed. |
| Patient letter | Explain this to the patient. | Draft a plain-language letter explaining her new diagnosis of pre-diabetes, what it means, and the agreed next steps, at a reading level suitable for a general audience. |
06 · What to avoid
Five question traps
Each trap feels natural to type. Each quietly lowers the quality of the answer.
1. Presumption "Why did X get worse?" — The model reasons from an assumption it never checks. Fix: ask what the record shows before asking why.
2. Leading "Confirm there is no…" — Invites agreement over accuracy, and blurs the difference between "none" and "none recorded". Fix: ask the open question and request sources.
3. Unbounded scope "Tell me everything" — The system silently decides what matters. Fix: name the period and the purpose.
4. Compound Three questions in one — Later parts are answered thinly or dropped altogether. Fix: one clinical question per request, then follow up.
5. Demand for certainty "What is the diagnosis?" — Pushes the model past what the record supports, and invites a confident guess. Fix: ask for evidence for and against, and what is missing.
The subtle one: "none" versus "none recorded"
"Does she have any allergies?" and "Is there an allergy record?" are different questions. A patient with no documented allergies may simply never have been asked. A precise question keeps that distinction visible, and a good answer states it.
07 · Asking for honesty
Phrases that make uncertainty visible
The Evidence element of SCOPE is the one most often left out, and the one that most protects patients. These phrases can be added to almost any question.
| Add this | What it does |
|---|---|
| Cite the source for each clinical statement. | Makes every claim checkable against the record |
| Tell me if the record does not contain this. | Stops gaps being filled with plausible guesswork |
| List what information is missing. | Turns omission, the invisible error, into something you can see |
| If my assumption is wrong, say so. | Gives explicit permission to disagree, which counters agreement bias |
| What would change this conclusion? | Exposes how fragile the answer is |
| Separate what is documented from what is inferred. | Distinguishes record from reasoning |
Follow-up questions are where the value is
The first answer is a draft. The most useful clinicians in the trials engaged longer, not shorter. Three follow-ups cover most needs.
Verify — "Show me the source for the statement about her potassium."
Narrow — "Focus only on changes since her admission in June."
Challenge — "What is the strongest argument against that conclusion?"
A caution on pushing back
Challenge the answer to test it, not to win it. Research on model sycophancy shows that when a user objects, a model may retract a correct answer and endorse a false one. If an answer changes the moment you disagree, ask for the evidence behind both versions.
08 · After the answer
Asking well is half the job. Reviewing well is the other half.
You remain the author of the record. A structured pass takes less time than writing from scratch and catches the errors that matter.
One of the largest manual evaluations of AI clinical note generation, covering almost 13,000 annotated clinical note sentences, found a hallucination rate of 1.47% and an omission rate of 3.45%. Missing content was more than twice as common as invented content, and it is harder to spot because nothing on the page signals a gap.
| Check | Ask of every clinically material statement |
|---|---|
| Sourced | Can I trace this to the record or the encounter? |
| Faithful | Are the negations, doses, values, dates and laterality exactly right? |
| Complete | Is anything clinically relevant missing, especially allergies and current medications? |
| Current | Is this still true now, or has it been superseded? |
What a well-built system should do with an imperfect question
| When the question is… | A trustworthy system should… |
|---|---|
| Ambiguous | Ask what you mean, rather than guess |
| Built on a false premise | Say so, rather than reason from it |
| Beyond what the record supports | Decline, and state what is missing |
| About a patient with safety-critical facts | Surface them, whether or not you asked |
Refusal is a feature
In a clinical tool, declining to answer is not a limitation. It is the highest expression of safety.
A system that always has an answer is a system that sometimes invents one.
09 · Keep this page
Quick-reference card
Print this page and keep it by your workstation.
| S | Subject | Which patient, which part of the record |
|---|---|---|
| C | Clinical question | One specific thing you need to know |
| O | Over what period | The time window that matters |
| P | Purpose | The decision this answer informs |
| E | Evidence standard | Sources for every claim; say what is missing |
Avoid
- Presumption · "Why did X worsen?"
- Leading · "Confirm there is no…"
- Unbounded · "Tell me everything"
- Compound · three questions in one
- Certainty · "What is the diagnosis?"
Add
- Cite the source for each statement
- Tell me if the record lacks this
- List what is missing
- If my assumption is wrong, say so
- What would change this conclusion?
Before you sign
Sourced · traceable to the record Faithful · doses, negations, sides, dates Complete · allergies and medications present Current · still true now
The AI drafts. The record verifies. The clinician decides.
References
The evidence behind this guide
-
Goh E, Gallo R, Hom J, et al. Large language model influence on diagnostic reasoning: a randomized clinical trial. JAMA Network Open. 2024;7(10):e2440969. doi:10.1001/jamanetworkopen.2024.40969
-
Bean AM, Payne RE, Parsons G, et al. Reliability of LLMs as medical assistants for the general public: a randomized preregistered study. Nature Medicine. 2026. doi:10.1038/s41591-025-04074-y
-
Johri S, Jeong J, Tran BA, et al. An evaluation framework for clinical use of large language models in patient interaction tasks. Nature Medicine. 2025. doi:10.1038/s41591-024-03328-5
-
Chen S, Gao M, Sasse K, et al. When helpfulness backfires: LLMs and the risk of false medical information due to sycophantic behavior. npj Digital Medicine. 2025;8:605. doi:10.1038/s41746-025-02008-z
-
Goh E, Gallo RJ, Strong E, et al. GPT-4 assistance for improvement of physician performance on patient care tasks: a randomized controlled trial. Nature Medicine. 2025;31:1233–1238. doi:10.1038/s41591-024-03456-y
-
Asgari E, Montaña-Brown N, Dubois M, et al. A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. npj Digital Medicine. 2025;8:274. doi:10.1038/s41746-025-01670-7
-
Wang L, et al. Prompt engineering in consistency and reliability with the evidence-based guideline for LLMs. npj Digital Medicine. 2024.
About
Respocare Connect AI
Respocare Connect AI is an agentic clinical assistant built by Respocare (PTY) Ltd, a licensed South African healthcare practice established in 2018. It is designed to carry patient context, retrieval and sourcing so that clinicians can focus on the question, and it is built to clarify, challenge or decline rather than guess. Clinical validation is ongoing.
Respocare Connect AI is a clinical assistant. It is not a medical device, is not diagnostic, and is not a substitute for clinical judgement. All output requires independent clinician review. This guide is educational and applies to clinical AI generally.
Respocare (PTY) Ltd · Reg. 2018/411829/07 · Practice No. 9990900010775614 · respocareconnectai.com · support@respocareconnectai.com
Built by Respocare (PTY) Ltd — a licensed South African healthcare practice operating since 2018.