Practical AI training for everyone
Module 1
7 min read
Medical school produces more reading than any human being can absorb at the pace it arrives. The honest problem is not that you read too slowly; it is that most of what you read, you will not remember in six months unless you engage with it differently than a straight read-through. AI is genuinely good at the first step of that: turning volume into structure.
A bad summary prompt produces a shorter version of the chapter that reads fine and teaches nothing, because it keeps the prose and drops the structure your brain actually needs: the causal chain, the classic exceptions, the thing that makes this topic show up on an exam. Ask for the structure explicitly.
A summary prompt that survives contact with an exam
'Summarize this chapter on diabetic ketoacidosis for USMLE-level understanding. Structure it as: the underlying mechanism in one paragraph, a numbered list of the classic presenting features with why each occurs, the key lab findings and what distinguishes DKA from HHS, and three things students commonly confuse about this topic. Do not just shorten the prose: teach the causal chain.'
Notice this only works because you supplied the chapter, or the notes, or a source you already trust. Asking cold for 'a summary of DKA' with no source pulls from wherever the model's training data happened to land, and for anything with a number in it, that is exactly where you do not want to guess.
It is easy to generate questions and easy to generate bad ones: questions where the wrong answers are obviously wrong, so getting it right proves nothing except that you can eliminate silly options. That is recognition, not the reasoning an exam or a ward actually needs from you.
Asking for questions worth answering
'Write 5 vignette-style questions on DKA management, USMLE style. Each wrong answer must represent a specific, plausible clinical reasoning error, not an obviously wrong choice. After each question, explain why the correct answer is right and, separately, why a student might pick each wrong answer and what that would reveal about their gap.'
That last clause is what turns a quiz into a diagnostic of your own understanding. If you picked the option built around 'confusing insulin timing with fluid resuscitation priority,' you now know precisely what to re-read, which is worth more than the point you missed.
A single AI-generated question set reviewed once behaves like any other passive resource: it feels productive and mostly is not. Ask for variations of the same concept spaced across sessions instead of a large batch reviewed once: 'give me 3 new questions on this same topic, different clinical scenarios, for me to try in two days.' Retrieval spaced over time is what actually moves things into long-term memory; cramming a big set once is not a substitute for it.
Try it now
Take one chapter or lecture you have not yet reviewed this week. Summarize it with the structured prompt above, then generate five vignette-style questions from it. Notice which wrong answer you almost picked; that is the most useful data point in the exercise.
A few quick questions. You'll see an explanation after each one.
Question 1 of 4
Why does asking for 'a summary of DKA' with no supplied source produce a risky result?