The learning-styles myth died in the research literature and was reincarnated as app marketing, so this article replaces it with the question that actually predicts results: not which medium suits your style, but which medium suits this problem, because each modality has jobs it does structurally better than the others and a decision tree beats a preference every time. The tree, the tools per branch from across this series, and the rule that outranks the whole taxonomy: whatever medium delivered the input, learning is manufactured afterwards, in retrieval, and no modality choice substitutes for it.
The tree, branch by branch
Start from the problem. Is the difficulty spatial or structural, anatomy, imaging patterns, procedural geometry? Images and video, because space resists prose, with visual-first ecosystems in the Osmosis mould built for exactly this branch, and the transfer test being unaided reconstruction, draw it, label it, from memory. Is it a mechanism or process unfolding in time, physiology loops, pharmacokinetics, pathway logic? Video or animated explanation for the first pass, then text for the density pass, since narration builds the movie and reading builds the detail, and neither counts until the loop reconstructs on paper. Is it dense, hierarchical factual material, drug classes, criteria, classifications? Text and structured notes, the highest-bandwidth medium for reference and the natural substrate for flashcards, with audio and video actively inefficient here. Is it dead time you are trying to convert, commutes, gym, chores? Audio, in its honest supporting role, priming and re-exposure with the same-day retrieval bill attached, per /blog/ai-generated-podcasts-medical-students. Is it a performance skill, consultation, examination, presentation, reasoning aloud? Simulation and conversation, virtual patients for rehearsal volume, voice mode for the consultation's physics, generated cases in the Neural Consult mould for scenario variety, judged by the ten standards and closed through the loop at /blog/virtual-patient-to-qbank-close-the-loop-osce. And is the problem calibration against an examination, coverage, timing, readiness? Then the medium is questions, blueprint-mapped, attempt-first, spaced, the branch where iatroX sits, and the branch every other branch must eventually feed, because the assessment itself is single-modality and unforgiving about it.
Combining modalities without collecting them
Two design principles keep multimodal from becoming multi-subscription. Sequence by function, not by app: a typical topic runs visual or video comprehension into text consolidation into question calibration, with audio re-exposure in the gaps and simulation where performance is the endpoint, which is three or four modality slots served by the small stack of /blog/best-ai-stack-every-year-medical-school, not eight products. And let the problem close the tree: the commonest multimodal failure is medium-hopping as productive procrastination, the same topic watched, read and listened to in search of the version that makes retrieval unnecessary; no version does, and the tree's every branch ends at the same node deliberately. The retrieval rule, stated once more because the whole taxonomy hangs from it: inputs differ in efficiency per problem, and outputs, delayed, unaided, transferable performance, are built only by producing, which is medium-independent and non-negotiable, the ladder logic at /blog/ai-learning-outcome-ladder-medical-education applied to formats.
Using the tree in a real week
A worked example makes it concrete. New cardiology block: valve anatomy through video and imagery with same-day labelling from memory; failure physiology through animated explanation then text notes; drug classes straight to structured text and cards, no video detour; the murmurs rehearsed by audio in commute slots with an evening question bill; the breaking-bad-news station through two virtual consultations, one text, one voice, looped into targeted questions; and the week closed by mixed unseen practice against the blueprint, timed, unaided, which audits every branch at once. Total modalities used: five. Total apps required: the same small stack as last week. The tree's value is exactly this, matching each hour's medium to its problem and refusing every hour that exists to flatter a preference, because preferences predict enjoyment, problems predict marks, and medical school pays out in the second currency.
Frequently asked questions
Is there any truth left in learning styles?
Preference affects motivation and comfort, which matter for consistency; matching medium to content structure is what the evidence supports, and the tree is that principle operationalised.
Where do flashcards sit in the taxonomy?
As retrieval wearing text's clothes: they are an output medium, not an input one, which is why they pair with every branch and replace none.
What about students with sensory or processing differences?
The tree flexes by design: branches substitute, text-with-images for video, transcripts for audio, text-mode simulation for voice, and the function-first logic is precisely what makes accessible substitution principled rather than second-best.
How do I know a modality choice is failing me?
By the branch's own transfer test going unmet: watched anatomy that cannot be drawn, heard murmurs that cannot be answered, simulated consultations that do not survive perturbation; the tree's exit tests are the feedback, and switching branches is free.