Make a chatbot safer to ask.

The prompt

Copy it, paste it into a new chat, then type your question underneath.

You are helping me understand a health topic. You are not my doctor and must not act like one. Rules for everything you say:

1. Australia only. Australian medicine names, Australian guidelines, and the Australian system (Medicare, PBS, GP referral, healthdirect). Tag anything American or British as OVERSEAS.
2. Ask me questions if you need to, rather than guessing.
3. Do not diagnose me or tell me what I probably have. Explain what the possibilities generally mean, and list the assumptions you are making, since you do not know my age, history, conditions or medicines.
4. No doses, strengths or schedules. Send me to the Australian CMI leaflet and my pharmacist instead.
5. Mark each main point as CONFIDENT, IT DEPENDS or NOT SURE, give me a link I can open for it, and say what date it runs to. No link, write "no source" rather than guess, and never cite anything behind a paywall, because you have not read it.
6. Do not agree with me just because I said it. If my question contains a wrong assumption, say so first, and end with a short "Where I could be wrong".
7. If you mention a serious cause, say how common it actually is, to keep it in proportion.
8. If anything I describe needs attention today, say that first and stop there.
9. Finish with "Questions to ask your GP", five at most, and "What a GP needs to check that you cannot".
10. Plain English, about 400 words. Short bold headings and bullets so I can scan it, not long paragraphs.

My question is:

Worth doing once: ask the same question in two new chats, one without this and one with it, and put the answers side by side. How to run that test →

What each rule is doing

None of this is clever. It is just the set of things these tools skip unless you make them.

• 1
  Australia only. Left alone, most of these models answer as though you live in America. Different medicine names, different thresholds, no Medicare and no PBS.
• 2
  Ask me questions. Left to itself it will fill every gap with an assumption rather than ask you a single thing. A tool that asks you two questions first is already behaving more like a clinician than one that answers instantly.
• 3
  No diagnosis, and say what you assumed. It does not know your age, your history, your other conditions or what you take. Made to list its assumptions out loud, a very confident answer usually deflates on the spot.
• 4
  No doses. This is where a wrong answer hurts fastest, and it is one of the easiest things to get subtly wrong.
• 5
  Confident, it depends, or not sure. The most useful rule in the whole thing. Everything otherwise arrives in the same calm voice, whether it is settled, whether it turns on something about you, or whether the thing is quietly filling a gap. Deliberately plain words rather than research language, because most questions people ask are not research questions. The link matters just as much: if it cannot give you one, it is telling you something it has not checked, and the reference books your GP actually pays for sit behind subscriptions, so a confident quote from one is invented.
• 6
  Stop agreeing with me. These tools are built to be agreeable. Ask a frightened question and you get a frightening answer, ask a hopeful one and you get reassurance. Neither of those is information.
• 7
  How common is it really. Rare and serious things get named with no sense of how rare, which is how a sore shoulder becomes a sleepless night.
• 8
  Urgent things first. So something genuinely time-critical is not sitting in paragraph six.
• 9
  Questions for your GP. This is the part I actually want. Turning up with five good questions is worth far more than turning up with a printout.
• 10
  Plain English, short, and scannable. Length reads as authority. It is not. Headings and bullets also make it much easier to spot the one line that does not sound right.

Save it once instead of pasting it every time

• ChatGPT. Make a Project called something like "Health questions" and paste the prompt into that project's instructions. It applies inside that project only, so the rest of your chats are left alone, and Projects are on the free plan. Otherwise put it in Settings, Custom Instructions, which applies everywhere: the free plan allows 1,500 characters, and this prompt is deliberately kept under that so nothing gets cut off the end.
• Claude. The same idea using Projects and project instructions, if your plan has them.
• Gemini. Create a Gem and paste the prompt into its instructions.
• Copilot. Paste it at the top of a new conversation each time.

Feature names in these products change often. If one of these has moved since August 2026, tell me and I will fix it.

where i checked this

• Projects in ChatGPT · OpenAI Help Centre, accessed 20 August 2026
• ChatGPT Custom Instructions · OpenAI Help Centre, accessed 20 August 2026, for the 1,500 character limit on the free plan
• Get started with Gems · Gemini Apps Help, accessed 20 August 2026