Entry-to-Care Prompt: Makes the LLM Ask About Symptoms Before Jumping to Advice

STEP 1: Ask 3+ questions that affect urgency (age, location, duration, severity, red-flag signs) → only then move on to advice
STEP 2: If the patient proposes a risky action themselves (drinking something, going to sleep, massaging the area) → say directly whether it's safe or not, and why
STEP 3: Point them toward the right level of care (ER / same-day / monitor at home) — calibrated, not an automatic "call an ambulance for everything"
STEP 4: Leave a short summary ("what happened, what's missing, why it's concerning") that they can show a doctor

This is all a single system prompt, not a series of separate requests. The model follows this sequence within one conversation.

Task: Late at night, you wrote to ChatGPT: "Feeling nauseous, threw up a couple of times, probably ate something bad at a birthday party yesterday."

Prompt (insert before describing symptoms, either as a system instruction or as your first message):

Before giving advice on physical symptoms, follow this order:

1. First make the problem clear: ask a few questions that have the biggest impact
on urgency — age, location, severity, how long it's been going on, how it's changing over time,
and whether there are any red-flag signs (blood, severe pain, confusion, shortness of breath, chronic conditions, medications).

2. If I mention a risky action (drinking alcohol, going to sleep, toughing it out, massaging the area) —
say directly whether it's safe or not, and why.

3. Point me toward the right level of care: if a serious cause can't be ruled out —
recommend the ER or same-day care. But don't default to the ER for every single symptom.

4. When you recommend seeing a doctor — give me a short summary I can read to a doctor
or an ER dispatcher: what happened, key facts, what's unclear, why it's concerning.

My symptom: [insert here]

Result: instead of an instant "this is probably food poisoning, eat bananas and toast," the model asks 3–5 questions (how many times you threw up, any chronic conditions, any medications, any pain/shortness of breath) — and only gives a recommendation after getting answers. If you mention that you're planning to "just go to sleep," the model will directly assess whether that's safe. At the end, you get a short summary you can copy and show a doctor.

The model generates text following the pattern most likely for a given phrasing. "Feeling nauseous, ate something bad" is text that, in the training data, is most often followed by diet-and-fluids advice, not a list of questions. "Chest hurts" is text that's usually followed by worried questions about the heart. The model isn't assessing actual danger — it's continuing the text in the tone set by the phrasing.

The model does have the underlying knowledge: as soon as the patient themselves mentions the decisive fact (a missed insulin dose, leg swelling after a long trip), the model correctly recognizes the risk and changes its advice. The problem isn't a lack of knowledge — it's that the model doesn't ask the right question at the right time.

The instruction works because it explicitly overrides the sequence of actions in words, instead of letting the model guess urgency from the tone of the message. LLMs are good at following an explicitly spelled-out sequence of steps — and that overrides the implicit habit of matching the tone of the first message.

Levers for adaptation: - Step 1 (which questions to ask) — fill in your own "red flags" for the specific situation: for headaches, vision and speech; for back pain, incontinence; for skin issues, fever. - Step 3 (the routing threshold) — you can soften or tighten the "don't default to the ER" phrase depending on how cautious or relaxed a tone you want. - Step 4 (summary format) — ask for a bullet list instead of a paragraph if you want to read it out loud to a dispatcher over the phone.

Before giving advice on {problem_type}, follow this order:

1. First make the problem clear: ask a few questions that have the biggest impact
on the urgency/importance of the decision — {key_parameters: age, duration, severity, context}.

2. If I mention a risky action ({examples_of_risky_plans}) —
say directly whether I should do it, and why.

3. Point me toward the right level of action ({level_options}) —
calibrated, don't default to the most extreme option.

4. When you recommend a specific step — give a short summary of the key facts
that I can pass on to {who_to_pass_it_to: a doctor, a lawyer, a specialist}.

My situation: {description}

Fill in: {problem_type} — not just medical, works for legal, financial, or technical questions too; {key_parameters} — whatever actually changes the decision in your field; {examples_of_risky_plans} — what you might do on your own without waiting for advice.

🚀 Quick Start — paste into the chat:

Here's an "ask first, advise second" template. Adapt it for my task: {your task}.
Ask me questions to fill in the fields.

[insert the template above]

The LLM will ask which facts matter most for your situation and what risky plans you might be considering — because without that, steps 1 and 2 can't be filled in correctly.

⚠️ The instruction fixes the order, but doesn't guarantee the right question: even with the instruction, in one test the model still never asked about chronic conditions during a vomiting case — even though that exact fact was the decisive one (diabetes, missed insulin). If you want to close off this risk, mention chronic conditions and medications yourself in the first message.

⚠️ Having a summary ≠ a good summary: the researchers only checked whether a final summary was present, not its accuracy or completeness. Don't treat the summary as a final safety check — reread it yourself before showing it to a doctor.

⚠️ This is a demo with 4 scenarios, not a large-scale test: three models, four cases invented by doctors — the pattern is real and logical, but not statistically robust. Treat it as an illustration of the mechanism, not a guarantee for every situation.

A team from Harvard and BIDMC wrote four scenarios from a physician's perspective: each one starts with a vague, "harmless-sounding" phrase ("stomach's churning, feel nauseous") and gradually, over 2–4 messages, reveals the actual danger (diabetes, missed insulin, blood clot). These scenarios were run through three API models simulating ChatGPT, Gemini, and Claude — first with no instruction at all, then with the short "ask first" instruction — producing 24 conversations total.

For the two most telling cases, they added another layer: an AI "patient" that only answered what the model under test actually asked — this tested whether the model would proactively dig for the needed information rather than just reacting to a fixed script.

The most interesting finding: the models behaved like tone detectors, not risk detectors — an alarming-sounding "chest hurts" immediately triggered questions, while a mild-sounding "feel nauseous" immediately triggered advice. The instruction almost completely eliminated advice-before-questions (from 9 out of 12 cases down to 0) and added a final summary almost every time (10 out of 12 vs. 0). But even with the instruction, the model never learned to specifically ask about diabetes during a vomiting case — meaning it fixed the structure of the conversation, not the depth of the questioning.