Where the fifteen minutes goes

Most people do not leave an appointment empty-handed because the doctor rushed them. They leave empty-handed because of what happened in the first ninety seconds.

You sit down. Someone asks what brings you in. And you start in the wrong place — with last night, because last night is what you remember, instead of with April, when it actually began. By the time you have circled back and corrected yourself twice, a third of the appointment is gone. Then the plan gets explained, and you nod. In the car park you remember the thing you came in about.

That is not a failure of intelligence. It is what happens to anybody who is anxious, sitting on the edge of an examination bed, summarising eleven weeks from memory while a stranger types.

The pressure is worst in exactly the places most South Africans are seen. The National Department of Health’s own National Guideline on the Management of Patient Waiting Time, approved in 2024, sets an aspired maximum of 120 minutes of waiting plus 60 minutes of service time at clinics and community health centres — three hours in the building for one visit. Whatever slice of that is you and a clinician facing each other is small, and it is the only part nothing else can replace.

You are not going to get better at remembering under pressure. You are going to arrive with it already written down.

That is the whole idea. Not a smarter chatbot. A page.

What AI is for here, and what it must never do

There is a clean split here, and the rest of this guide sits on top of it.

AI is genuinely good at organising information you supply. Give it months of scattered memory and it will put the events in order and produce a short, readable summary in the same shape every time. Afterwards, it will take the words on a discharge letter and tell you what they mean in ordinary English. Those are real jobs, and it does them well.

AI must not be used for three things. Not carefully, not as a second opinion, not out of curiosity.

1. It must not tell you what is wrong with you. No diagnosis, no shortlist of conditions, no “it’s probably one of these three”.
2. It must not help you decide whether to seek care. That decision belongs to you and a clinician, and it should always tip towards going.
3. It must not touch your medication. Not the dose, not the timing, not starting something, and above all not stopping something. If a chatbot has an opinion about a medicine you take, close it and phone the pharmacy or the rooms.

The reason is not that the technology is stupid. It is two more specific things.

It is confident when it is wrong. A language model produces the same calm, well-organised, reasonable-sounding paragraph whether it is right or badly wrong. There is nothing in the tone to warn you. A doctor who is uncertain sounds uncertain. A chatbot that is uncertain sounds exactly like a chatbot that is correct, and that difference is the whole danger.

It cannot examine you. It cannot take your blood pressure, listen to your chest, feel your abdomen, look in your ear, or send blood to a laboratory. Much of what a clinician concludes comes from things that never appear in your typing — how you walk in, what your skin looks like, what happens when they press somewhere.

The vendors say so themselves. OpenAI’s help pages for health features in ChatGPT state that it “is designed to support, not replace, medical care and is not intended for diagnosis or treatment”, and its usage policies prohibit “provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional”. When the company selling you the tool draws the line that firmly, believe it.

And the dangerous output is not the frightening one. It is the reassuring one. A paragraph saying this sounds like nothing much has handed you a reason to cancel, and a reason to cancel is the last thing you needed. Nothing a chatbot says should ever be why you did not go.

Prompt one: turning the mess into a timeline

You already have the information. That is the part people do not believe. It is in WhatsApp messages to your sister, in the photo you took of a rash, in the fact that you felt terrible at a wedding you can date exactly. It is not organised, so it does not feel like information. It is.

Do the brain-dump first, and do it badly on purpose. Open the notes app and type everything you remember about this problem, in any order, for about ten minutes. Repetition is fine. Half-sentences are fine. You are collecting, not writing.

Then hand the mess over. The guardrail sits in the first lines of the prompt, before anything else. That is deliberate. Do not remove it.

You are helping me organise notes for a medical
appointment. You are not a doctor and you are not
diagnosing anything.

Rules you must follow:
- Do not tell me what condition I might have.
- Do not suggest tests, treatments or medicines.
- Do not comment on whether this sounds serious.
- If I ask you to do any of those, decline.
- Use only what I tell you below. Invent nothing.

Here is everything I can remember, in no order:
[paste your brain-dump here]

Turn it into a one-page summary for a clinician, using
these headings and nothing else:

1. Main problem, in one sentence
2. When it started, and what has changed since
3. Timeline of events, oldest first, with dates
4. What makes it better, what makes it worse
5. What I have already tried
6. Conditions I already have, and medicines I take
7. What I want out of this appointment

Where a date or a detail is missing, write "not noted"
and list it at the end under "Ask me about".
Plain words only. No medical terms I did not use myself.
Keep the whole thing under 350 words.

Now the part nobody does, and the part that matters most: read the output line by line and correct it. The model will have guessed at an order. It may have merged two events into one, or turned “a few weeks” into a date you never gave it. You are the fact-checker, and the only one who can be. A tidy page with a wrong date on it is worse than no page at all, because it looks reliable. Fix it by hand, then read the whole thing aloud once.

What a good handover page looks like

A good handover page has a shape you can recognise from across a desk. One side of one page. Headings. Oldest first. Real dates instead of “a while back”. Ordinary words. And at the bottom, the two things people always leave out — what you want from the appointment, and what you are unsure about. The details below are invented to show the format, not drawn from anybody real.

FOR: my appointment, Tuesday 9 September
ONE PAGE - please keep this copy

MAIN PROBLEM
Headaches most mornings for about seven weeks, worse
and longer than the ones I used to get.

WHEN IT STARTED AND WHAT HAS CHANGED
Started early July, about twice a week. Now five or
six mornings out of seven. They last around three
hours instead of one.

TIMELINE
Early July  - first bad one, woke me before dawn
Mid July    - started keeping a note; 2-3 a week
1 August    - new job, much longer screen hours
Since then  - most mornings, and getting longer
19 August   - first time I felt sick with one

BETTER / WORSE
Better: dark room, sleeping later on weekends
Worse: mornings, bright screens, skipping breakfast

WHAT I HAVE ALREADY TRIED
Something from the pharmacy shelf most mornings for
the last three weeks. Helped at first, less now.

CONDITIONS I HAVE AND MEDICINES I TAKE
[your list here - what, how much, how long]

WHAT I WANT FROM THIS APPOINTMENT
To know whether this needs a test, and what would
mean I should come back sooner rather than later.

ASK ME ABOUT
Whether my sleep changed before or after the
headaches started - I genuinely cannot remember
which came first.

Print two copies if you can: one to hand over, one to read from. If there is no printer, have it open on your phone with the brightness up and auto-lock off.

Hand it over at the start, not the end, with one sentence: “I wrote this down so I don’t waste your time — the short version is at the top.” That sentence does a lot of work. It signals you are not going to talk for nine minutes, and it gives the person permission to read rather than interview.

Prompt two: the questions you would have forgotten

“Any questions?” is the hardest question in medicine. Everybody says no. Then everybody thinks of four on the way home.

The questions worth asking are mostly not about what it is. They are about what happens next, what the options are, and what should bring you back — which means you can prepare them, because they do not depend on a diagnosis.

Here is the summary I am taking to my appointment:
[paste your corrected one-page summary]

Give me the questions a well-prepared patient would
ask about this, so I do not go blank in the room.

Rules:
- Do not answer any of the questions.
- Do not guess at a diagnosis or mention conditions.
- Do not suggest which tests or treatments are right.

Group them under these headings:
- Understanding what is happening
- What tests or next steps might be involved
- What the plan is and how long it should take
- What would mean I must come back sooner
- Cost, referral and authorisation

Short enough to read aloud. Mark the three I should
ask first if the appointment runs out of time.

Cut the list to the three marked ones plus anything you genuinely care about, and write those at the bottom of your handover page by hand. Ten questions on a phone screen do not get asked. Three on a page you are already holding do.

If you keep only one, keep this: what would mean I should come back sooner? It turns a vague “come back if it gets worse” into a threshold you can recognise at eleven at night.

You are also allowed to ask more than you think. The National Patients’ Rights Charter, developed by the National Department of Health and reproduced in the Health Professions Council of South Africa’s ethical guidelines, says everyone has the right to “full and accurate information about the nature of one’s illnesses, diagnostic procedures, the proposed treatment and risks associated therewith, and the costs involved”, and the right, on request, “to be referred for a second opinion to a health practitioner of choice”. Cost is in there explicitly. So is the second opinion. Neither is rude to raise.

Prompt three: the ten minutes after you walk out

Do not drive yet. Sit in the car and write down what was said while it is fresh — typed, or as a voice note. What the plan is, what happens next, and every word you did not understand. Five minutes here stops the whole appointment evaporating by Thursday. Then, at home, use the model for the one job it is unambiguously good at: translation.

I have come out of a medical appointment. Explain the
words below in plain English so I understand what was
said to me.

Rules:
- Do not tell me what is wrong with me.
- Do not say whether the plan is right or wrong.
- Do not suggest anything different from what I was
  told, and do not comment on any medicine.
- Where something is unclear or missing, tell me to
  ask the rooms rather than filling the gap yourself.

What was said or written (names, ID numbers and
scheme numbers removed):
[type it out here]

Give me:
1. Each medical word, and what it means in plain
   English
2. What the plan appears to be, in the order it
   happens
3. A short list of questions to phone the rooms about
4. Anything I have clearly misheard or half-written,
   flagged as uncertain

What you get back is a glossary and a list of phone calls. That is all it is for. The moment the translation starts to feel like a second opinion, you have crossed the line this guide is built around. If the plain-English version and your memory disagree, the model does not break the tie. The rooms do. Phone them — receptionists field that call constantly.

One caution before you paste: do not photograph a letter and upload the picture. Referral letters and discharge summaries carry your full name, identity number, scheme membership number and often a barcode, all in the corner you were not looking at. Type out the part you need instead.

The clinic, the rooms, and the medical aid

Where you are seen changes what the page is doing for you. It does not change whether it is worth writing.

The written timeline is worth the most in the rushed public-clinic visit — the opposite of what people assume. Where time is short and the person in front of you has never met you, a page they can read in forty seconds is the biggest thing you can bring. Take your clinic card and file number too, so nobody spends your minutes hunting for a folder.

The medical aid questions worth asking out loud

If you belong to a scheme, four procedural questions save a great deal of trouble later. Ask them in the rooms, and ask the scheme on the number on your card.

• “Does this need pre-authorisation, and who gets it — you or me?” The Council for Medical Schemes states that schemes “can make a benefit conditional on you obtaining pre-authorisation or joining a benefit management programme”. Some practices handle it. Some assume you will. Find out which.
• “Is this a Prescribed Minimum Benefit condition, and how do I register it?” The Council describes PMBs as benefits schemes must cover by law — 26 chronic conditions, medical emergencies, and 271 specified diagnoses linked to particular treatments — and says that even where your yearly benefits are depleted, the scheme must still cover diagnosis, treatment and ongoing care for a PMB condition.
• “Is this provider a designated service provider on my plan?” The Council notes that where a scheme appoints a designated service provider and you voluntarily use a different one, the scheme may charge you the difference — and that a scheme cannot charge a co-payment on a PMB where you follow its formulary and protocol.
• “Which codes will be billed, and what does my plan pay for them?” Ask before, not after. The rooms can usually tell you in a sentence.

Use a chatbot to help you phrase those questions. Do not use it to answer them. It does not have your plan document, it does not know your scheme’s rules for this year, and it will produce a tidy, confident, invented answer if you let it. The scheme is the only authority on what the scheme pays, and when real money is at stake, get it from them in writing.

What to take out before you paste

A consumer chatbot is somebody else’s computer. What you type goes to a server you do not own, sits in a conversation history, and depending on your settings may be retained or used to improve the product. That is not a scandal — it is how the tools work — but it is worth knowing before you paste your health history into one.

South African law treats this category seriously. The Information Regulator’s guidance note on processing special personal information lists health among the categories of special personal information, and explains that section 26 of POPIA prohibits processing it except under the exceptions in section 27. That law leans on organisations holding other people’s information rather than on you writing about yourself — but it shows how seriously the category is taken, and it applies directly the moment you do this for a parent, a partner or a child.

Strip these out. They add nothing to the output.

• Full names — yours and everyone else’s. Write “the patient”, “my mother”, “my son”.
• Identity and passport numbers, always.
• Medical scheme name, membership number and dependant code.
• Your doctor’s name, the practice name and the practice number.
• Your home address, your workplace, your child’s school.
• Photographs or scans of any document. The letterhead, the barcode and the corner block carry all of the above at once.

Keep the symptoms, the dates, the order of events, an age range, and the conditions and medicines that matter clinically. That is what makes the summary useful. The names and numbers add nothing to the output and everything to what you are handing over.

Three settings worth ten minutes

Turn off model training in your chatbot’s data controls, if the tool offers it. Delete the conversation once you have your page — OpenAI notes that information already in your ChatGPT history “remains until you delete those conversations”. And do none of this on a work account or a shared family device, where a colleague or an administrator may reasonably have access.

To be fair to the vendors, some of this is handled for you: OpenAI states that connected medical records and Apple Health information “are not used to train our foundation models or target ads, regardless of the model-training setting”. That covers connected health data in that one feature. An ordinary conversation where you paste symptoms still follows your normal settings, which is why the settings are worth checking.

And if you are building this page for someone else, ask them first. It is their information, not yours, even when you are the one who books the appointments and drives them there.

Next in this pillar: Your Body Has Data, on what to do with the numbers your devices already collect. New here? Start here.