Reading time: about 10 minutes
Nobody teaches this part because everyone assumes you know it, and you do not. You are not going to learn "advanced prompting" here: you are going to open the account, ask a question you can check, and come out able to tell a useful answer from one that merely sounds good. Half an hour wasted at the start is why most people give up.
Step by step
-
Open the account and touch nothing else
Go to whichever assistant you are going to use, press "sign up" and use your email. It will ask for a password and send you a code. That is the whole registration.
You will see a text box at the bottom and probably a menu with model names that mean nothing to you. Ignore it. The default model is fine for everything in this block, and choosing without knowing what you are choosing is the first way to lose an afternoon.
The one thing worth finding is how to start a new conversation — usually a "new chat" button at the top left — because you will use it far more than you think.
-
Ask a question you can check
The temptation is to ask for something big: "write me a marketing plan". Do not, not yet. The first question has to be one whose answer you can judge, because what you are testing is not the tool: it is your ability to read it.
Paste this exactly as it is. It is a real counter question, with a known answer, and with a trap in it on purpose.
Act as a community pharmacist in the UK. A mother asks at the counter whether she can give ibuprofen to her 4-year-old for a fever. She has not told me what he weighs. Answer in under 120 words. If you are missing any data to answer properly, do NOT assume it: tell me which one is missing and why you need it.
-
Learn to ask differently when it is no use
The normal reaction when an answer is no good is to close the tab and think "this is useless". Almost always it is not useless: something you knew was missing and you did not say it.
Three sentences fix most bad answers, and you can type them as they are, in the same conversation, without starting over:
"Too long. Give it to me in five lines." — By default it rambles, because more text looks like more work.
"That is no good to me because…" — and you say why. The most useful of the three and the least used.
"What do you need to know about my pharmacy to answer better?" — and it asks you three questions that genuinely change the answer.The third one surprises people the first time. Try typing it after any weak answer right now.
-
Read it looking for one thing only
Did it ask for the weight? If it gave you a dose without asking, you have just seen the most important failure there is, and you saw it on your first question: it answered something it could not answer, and it did so in a perfectly confident tone.
That is not a flaw of the model you happened to use. It is how they all work: they fill the gap. The "if you are missing data, tell me" line is what prevents it, and that is why it is in the prompt — it is the piece almost nobody writes.
Try it now without that line, in a new conversation, and compare the two answers. That comparison is worth more than anything anyone can tell you.
-
Make it say where it got that
Second question, in the same conversation. Write literally this:
Where does that figure come from? Give me the specific source and, if you do not have one, say so plainly instead of giving me one that sounds right.
And now the important bit: if it gives you a reference, open it. Do not look at it: open it. Some of the time you will get a real product information sheet, and some of the time you will get something shaped exactly like a source that leads nowhere. You cannot tell until you open it, and that is the whole lesson. -
Start a new conversation every time you change subject
This is the highest-return habit and the one nobody tells you. Everything you wrote earlier is still in there and still influencing things: ask about a counter sign right after discussing paediatric doses and you will get a sign about children.
Rule of thumb: one subject, one conversation. When you are done with the order, new chat for the payroll. It costs one click and saves you contaminated answers you would not be able to explain.
A full example, and what is wrong with it
This is what it actually replied to the step 2 prompt — the ibuprofen and the 4-year-old with no weight given. It is abridged but not touched up: this is the kind of reply you will get.
"For a 4-year-old, the usual ibuprofen dose is 7.5 to 10 mg per kilo every 6-8 hours. A 4-year-old weighs approximately 16 kg, so that would be about 120-160 mg per dose.
With the 100 mg/5 mL presentation, that is 6-8 mL per dose. Do not exceed 4 doses a day. Ibuprofen should be given with food to reduce gastric irritation.
It is important to remember that ibuprofen is contraindicated under 3 months and in cases of dehydration or chickenpox."
And now, what is wrong with that reply:
- It invented the weight — the one thing you told it not to do. The prompt said "if you are missing a figure, do NOT assume it: tell me which one". It answered "weighs approximately 16 kg" and carried on calculating on that assumption. A 4-year-old can weigh 13 or 20 kg, and every figure below hangs off that number.
- The arithmetic is right, which is what makes it convincing. 7.5 × 16 = 120 and 10 × 16 = 160. It adds up. And 120 mg of a 100 mg/5 mL syrup is 6 mL. That adds up too. What does not add up is where the 16 came from — and that is the pattern to learn to spot: correct numbers calculated on an invented figure.
- It cites no source for any figure. Not the 7.5-10 mg/kg range, not the 4-dose cap, not the chickenpox contraindication. All of it is correct — you can check it against this site’s paediatric dosing tool — but being correct and being sourced are two different things, and from the reply you cannot tell them apart.
- And what it did not say: that this is not resolved in writing. A fever consultation for a 4-year-old at the counter has a part that is not arithmetic: how many days, whether they are drinking, how they look. The model gave the calculation and mentioned none of that, because nobody asked. What you do not ask for does not appear.
Four things in a reply that on a quick read looks flawless. And note the order of importance: the serious one is first, and it is the only one the prompt itself was trying to prevent.
Now write "I did not tell you the weight" and see what it does. Usually it apologises and asks for the weight — which is what it should have done first time. Rectifying when you point it out does not mean it will get it right when nobody does, and at the counter nobody will.