AI School · Level 3 · Lesson 5

Exercises with synthetic clinical cases

Léelo en español →

You will not write a single prompt here. The AI has already answered — a medication review and a breastfeeding question — and your job is the one you will be doing every day from now on: saying what is wrong. It is marked against a rubric written by a pharmacist, not against whatever a model happens to think that day.

Missing a serious finding is not a narrow pass: it is the failure the lesson exists to prevent. That is why spotting six things and skipping the major interaction does not pass, however high the percentage. It is exactly what happens at the counter.
Both cases are synthetic. They are not real patients, never have been, and they ship with the site. An exercise is where it is most tempting to use "a real one" — and it is exactly where you cannot (lesson 1).

The two cases

How it is marked

The mark comes from a rubric: every finding has its severity and its words. Spotting something serious is worth three times something minor, and there is one rule above the percentage — miss a serious one and you do not pass. It is the same rule the Level 1 exercise uses, with the same engine.

And the marking also shows you what you missed, with the reasoning. Knowing the score without knowing what is missing teaches nothing.

A full example, and what is wrong with it

Before the two cases, one worked through out loud. It is neither of them and spoils nothing — it is on a different topic — and it is here for the same reason as in the Level 1 exercise: what stops you the first time is not spotting the errors, it is not knowing what you are supposed to write.

The query is one of those that arrive by phone mid-morning, from a care assistant at a residential home:

"We have a resident who chokes on tablets. Can we crush them and give them with yoghurt? There are eight of them."

What it replies

Yes, tablets can generally be crushed to help swallowing in patients with dysphagia. The exception is prolonged-release forms, which must never be crushed because the whole dose is released at once. They are usually identified by letters in the name (retard, LP, XR, OD).

My practical recommendation: crush all the tablets together in a mortar, mix them with a little plain yoghurt and give them in one go. The yoghurt masks the bitter taste and greatly improves acceptance.

If any of the eight were prolonged-release, the doctor should be asked to change it to a liquid or dispersible form.

As for capsules, they can be opened and the contents sprinkled on food without any problem.

And now, what is wrong with that reply:

  1. It never asked which eight they are. (major) The whole reply is built on a list it has not seen. And the question has no answer without it: crushing is decided medicine by medicine, by reading section 6 of each information sheet — the previous lesson — not from a general rule. It is the failure that repeats in both cases of the exercise and the easiest one to spot once you know to look: the reply answers instead of saying a fact is missing.
  2. The letters are not how you identify them. (major) "They are usually identified by the letters" is the sentence that does the damage, because it sounds like a useful trick and turns a check into a visual inspection. There are modified-release forms with no letters in the name at all, there are enteric coatings that also must not be crushed and are not prolonged-release, and there are letters that mean something else. Anybody who keeps this sentence will crush wrongly in precisely the cases that carry no written warning on the box.
  3. "Crush all eight together in a mortar" is an invented procedure. And it sounds entirely sensible. But crushing and mixing eight drugs in the same mortar brings in two problems nobody asked for: residue from one dose into the next — a mortar does not come out clean and tomorrow's dose carries some of today's — and drugs that must be taken apart from the others or away from dairy. If any of the eight cannot go with yoghurt, this instruction puts it in yoghurt.
  4. Nobody has said who does the crushing, and some must not be crushed by hand. (moderate) The person doing it is a care assistant, eight times a day, every day. There are drugs whose powder must not be handled unprotected — cytotoxics, some hormonal agents, some immunosuppressants — and the risk here is not the resident's: it belongs to whoever crushes it and breathes it in every morning. It is the only one of the five failures that does not affect the patient, which is why almost nobody thinks of it. The question was missing a fact and the answer is missing a person.
  5. And "capsules can be opened without any problem" is false as a rule. (moderate) It says it last, in passing, like somebody closing a minor point — which is exactly where the most expensive error in this whole school lives. There are modified-release capsules and capsules with enteric-coated granules that must not be opened, and others whose contents must not be chewed even though they can be sprinkled. Section 6 of each sheet says so. "Without any problem" is the surplus part, and it is the part that will get repeated on the phone.

This is how the marking answer gets written: unordered, in your own words, not drafted. One that would pass looks like this:

"It never asked which eight, and without that there is no answer: you check them one at a time in section 6. The letters thing is false — there are modified-release forms with no letters, and enteric coatings. Crushing all eight together mixes residue and may put something with dairy that should not be. It says nothing about protecting whoever crushes them. And capsules cannot always be opened."

Forty seconds and five lines. No need to explain why or to cite anything: you have to point. And notice what decides the pass — both findings marked major are in there. An answer twice as long that misses either one does not pass, whatever percentage it scores, because spotting three details and missing that the question had no answer is the whole failure.

And if your pharmacy is not like that

If you have come from the Level 1 exercise. The two here are harder and it helps to know why: there the errors were in what was said and here half of them are in what was not asked. It is a change of gaze, not of knowledge. If you get stuck, start with one question — "what would I need to know before answering this?" — and compare it with what the model took as given. That is usually where the major finding is.
If you are a technician or an assistant. You will find the framing errors before anybody else, because they are the counter ones: "it never asked what else they take", "that is not the pharmacy's to decide", "you cannot say that sentence on the phone". That is half the marks and it needs no dose memorised. The fine clinical detail costs more, and that is fine: the rubric does not score your degree, it scores whether you spot a missing fact.
If you are going to use it with your team. Have everybody do it separately before talking. The moment somebody says out loud what they spotted, the rest can no longer know whether they would have spotted it — and the interesting part is not who found more, it is what one person saw and the others did not. Fifteen minutes later, one case talked through out loud teaches more than the month's training session, because it surfaces what each person looks at first.
If it has been a long time since you did clinical work. Do not revise first: do it as you are. What is measured here is not what you remember, it is whether you accept an answer that cannot be given, and that keeps better than people think. If the marking turns up a figure you did not know, so much the better — that is the specific lesson that was needed, and it is worth more than revising a whole syllabus just in case.
If you work with care homes or do a lot of blister packing. The crushing case above will look familiar, and that is worth using: you get that call for real several times a month and almost always in a hurry. What changes the outcome is not knowing the exceptions by heart — there are too many — but having decided the sentence you answer the phone with: "tell me which ones and I will tell you which can and which cannot, give me ten minutes". It turns a general answer into a one-by-one check, which is the only kind that counts, and it puts it in writing too.

When it does not work first time

I read the reply and it looks correct to me.
Normal: they are written to look that way, and none is nonsense. Change the question — instead of "what is wrong?", walk the reply asking "where does this come from?": from a fact it was given (was it given?), from a rule or an information sheet (does it say that?), or from nowhere. And look separately at the last sentence, which is where these models slip in the actionable part with nothing behind it.
I spotted three things and it failed me.
Then you missed a major finding, and the marking tells you which. It is deliberately uncomfortable: spotting six details and missing the one that changes the decision is not a narrow pass, it is the failure this exists to prevent. What helps is not repeating the case until you pass — you will end up knowing it by heart — but looking at why that one got past you: did you skim the end? did you trust it because the rest was right? did you assume a fact?
I wrote a lot and scored badly.
The rubric scores findings, not words: three paragraphs explaining one error are worth the same as one line pointing at it, and while you were writing those paragraphs you were not looking at the rest. Try it the other way — one line per finding, unordered, and when nothing else comes to mind, reread the whole reply from the top. That is usually where the missing one turns up, and it is almost always in the first paragraph or the last.
I passed both and I still do not trust my own judgement.
Here you knew something was wrong because we told you in the title, and at the counter nobody warns you: that is the whole difference and it explains the feeling. The practice that pays is not repeating these cases, it is applying the same reading to the next real reply an AI gives you — where each statement comes from, what it did not ask, what it decides that is not its to decide. Three goes with a real answer beat ten cases here.
What if my answer is right and the rubric does not list it?
It can happen and it does not subtract: the rubric lists what you have to spot in order to pass, not everything that can be spotted. If what you saw is a genuine error, good — it usually comes from knowing something about the case that whoever wrote it did not assume, and it is exactly the judgement this trains. What is worth checking is whether your finding took the place of a major one: seeing everything except the thing that decides is precisely the pattern the marking looks for.
I went blank and wrote nothing.
It is commoner than it looks and it has a mechanical fix: write first what the question is missing, without looking at the reply. In both cases — and in the example one — there is a fact nobody supplied, and naming it is already a finding, often the major one. Then go back to the reply and check only whether that missing fact changes anything that has been claimed. That usually produces half the marking answer without having had to remember anything.

Before you call Level 3 done

← Previous: checking against the SmPC Back to the school →