Real patients don't arrive with the chapter attached, and the practical question that follows from that idea is the one this article answers: if the chapter heading is doing half the learner's work, what capabilities should an AI medical tutor actually train? Not "medicine", which is a library, and not "exam technique", which is a by-product. Seven specific, nameable competencies, each trainable, each measurable, and each served differently by the current generation of tools. The philosophical case is made at /blog/real-patients-dont-arrive-with-the-chapter-attached; this is the practical companion.
The seven competencies
One: problem representation. Can the learner compress a messy case into one clinically useful sentence, the age, the context, the trajectory, the pivotal finding? This is the master skill; experts do it automatically, novices drown in undifferentiated data, and it is trainable by the simple discipline of being made to produce the sentence, case after case.
Two: differential generation. Can they produce plausible hypotheses without being handed five options? The SBA format quietly removes this competency from practice, which is why high scorers can freeze in front of real ambiguity; training it means open-ended commitment before any options appear.
Three: discrimination. Given competing hypotheses, can they identify which findings actually separate them, the feature that makes one diagnosis less likely, not just the features consistent with the favourite? This is the competency question banks train best, and only if the review asks why, not just what.
Four: next-step reasoning. Which test, question or examination finding would most change management? Information has value only insofar as it moves decisions, and learners systematically over-order and under-think; the training is being asked, repeatedly, what would you do with the result?
Five: uncertainty. Can they state their confidence and their live alternatives out loud? Calibration, knowing when you are probably right, is the safety-critical competency, and it only trains against feedback: committed confidence, checked against outcomes, adjusted.
Six: management movement. Can they get from diagnosis to action, sequenced, prioritised, guideline-aware? Diagnostic elegance without management traction is a written-exam skill; the wards, and increasingly the exams, want the whole arc.
Seven: reflection. When they chose the wrong path, can they say why, the anchor, the missed clue, the discarded alternative? This is the competency that compounds all the others, because errors only teach when their mechanism is named.
What current tools train, honestly mapped
Question banks train discrimination under exam conditions, superbly, and competencies two and four only weakly, since options are supplied and next steps are pre-selected. Spaced repetition trains retention of whatever was learned, agnostic to which competency produced it. Virtual patients, Geeky Medics' and Quesmed's territory, train information-gathering and communication, the consultation-shaped competencies, and touch representation. Socratic tutors train the reasoning core directly: representation, discrimination, next-step logic, uncertainty and reflection, by interrogating the learner's own failed attempt, which is the design brief of the iatroX Tutor specifically. Study planners train nothing directly and determine what gets trained, which makes them quietly decisive. And clinical reference AI trains none of the seven, correctly, because its job is practice support, not pedagogy. No single tool covers the seven; the mapping is the point, and the stack view is at /blog/quesmed-osce-ai-vs-geeky-medics-ai-patients-vs-iatrox-tutor.
Where every current tool remains weak
Honest gaps, named: competency two, unprompted differential generation, is under-trained everywhere the SBA format dominates, including by us, and the tools closing it fastest are open-commitment formats, virtual patients and daily diagnosis games included. Competency six's sequencing is largely still taught by wards, not software. And competency seven's reflection is scaffolded by tutors but authored, necessarily, by the learner; no system should write it for them. An honest buyer's question for any AI tutor demo: which of the seven does this train, which does it merely decorate, and which does it quietly let me skip?
The ideal learning environment, in one paragraph
It would assess all seven continuously rather than at exam time; open every case with commitment before options; interrogate errors at the mechanism level; retest corrections at intervals; schedule the mix by measured weakness rather than learner preference; and hand the consultation-shaped competencies to simulation while keeping the reasoning core in dialogue. Every element exists today, distributed across the tools above; the environment that integrates them is the race the category is now running.
Frequently asked questions
Which competency should a struggling student fix first?
Representation: it is upstream of everything, cheap to train, and its absence masquerades as knowledge failure. The five-minute daily version is at /blog/five-minute-clinical-reasoning-daily-habit-new-doctors.
Can these seven be assessed in exams?
Increasingly they are: modern SBAs target discrimination and next steps, SCAs and OSCEs sample the consultation-shaped set, and script-concordance formats reach uncertainty. The exam lag is real and closing, which is one more reason to train the competencies rather than the format.