AI 健康提示：Claude 與 ChatGPT 的安全實用用途

Ready-made prompts for the parts of health that are reading, organising and preparing rather than clinical: turning a leaflet or letter into plain language, organising what to raise at an appointment, writing down what you were told afterwards so you can check it, keeping track of appointments and repeat prescriptions, and structuring general wellbeing habits. They work in Claude, ChatGPT or any AI chat. They are not medical advice and do not replace a qualified professional.

No, and the reason is specific rather than a formality. A model agrees with the direction you asked from: ask whether something could be serious and you get a well-written case that it could be, ask whether it is probably nothing and you get reassurance, from the same facts. If you are anxious you will phrase it the anxious way and the answer will read like expert confirmation. It also cannot examine you, has no vital signs, no bloods and no history. Symptoms go to a person with a licence.

No. A number without your history and an examination is not an interpretation, and reference ranges are exactly the kind of confidently specific detail a model gets wrong in a clean, authoritative format. What it can safely do is explain the terms on the page in plain language so you understand the result well enough to discuss it, with an explicit instruction not to tell you what it means for you.

No. Medicines, doses, changes and interactions go to your pharmacist, who does this accurately, for free, and knows what else you are taking. This is one of the clearest lines on the page and it applies to over-the-counter medicines too.

A consumer AI chat is a third-party service, not a clinic, and once something is pasted it is pasted. Take out names, dates of birth, record numbers and addresses before you paste anything real, and consider whether you need to include the condition itself, since most of these prompts work on a description with the identifying detail removed. If the task is understanding a public leaflet, there is nothing to redact.

Five things. Turning dense information into plain language so you can discuss it. Organising what to raise in a short appointment, as questions rather than answers. Writing down what you were told afterwards and flagging what you are not sure you heard correctly, which is the half almost nobody does. Health admin, which has no clinical content at all: appointments, repeat prescriptions, letters, claims. And general wellbeing structure such as sleep, movement and breaks, kept general.

Yes. They are plain prompts, so they work in Claude, ChatGPT, Gemini or any AI chat, and the guardrails written into them travel with the text. The one thing worth keeping is the instruction that the rules override anything you ask later in the conversation, because that is what stops the chat drifting into diagnosis twenty messages in.

If the next thing you want to type would be answered by a person with a licence, ask that person. Chest pain, signs of a stroke, difficulty breathing, severe bleeding or any symptom that frightens you means contacting your local emergency services rather than a chat window. Anything involving a child's illness goes to a professional. If you are struggling with your mental health, that is a conversation for a professional or a helpline in your country, and it is worth having sooner than feels necessary.