The wrong question is which AI tools are best; the right question is which problem this year of medical school actually poses, because the problem changes annually and the stack should follow it. The other principle this guide runs on: deliberately small stacks, two or three tools per phase, chosen by function, since every additional subscription taxes attention, money and the switching costs treated at /blog/medical-ai-lock-in-who-owns-learning-history, and the students drowning in tools are rarely the ones passing comfortably. Year by year, problem first, tools second, all platform descriptions public and current at writing.
Early preclinical: the comprehension years
The problem: building understanding of mechanisms, structures and vocabulary from near zero. The stack: one comprehension layer, video-and-visual (Osmosis-style) or library-first (AMBOSS-style) by taste and subject mix, per the comparison at /blog/osmosis-ai-vs-amboss-ai-preclinical-students; one general learning mode, ChatGPT Study Mode, Gemini Guided Learning or Claude's Learning mode, for unlimited reformulation and attempt-first dialogue, chosen by the hour-long protocol at /blog/chatgpt-study-mode-vs-gemini-guided-learning-vs-claude-learning-mode-medical; and flashcards with genuine spacing for the vocabulary layer. The rule that makes the year work: every comprehension session ends in attempted recall, because the watching-and-reading habit feels like progress and banks almost none.
Later preclinical: the transfer years
The problem: converting mechanisms into vignette performance, the mechanism-to-application jump that ambushes second-year confidence. The stack tightens: the comprehension layer stays for new material; a question bank enters as the centre of gravity, run attempt-first with wrong answers interrogated rather than absorbed; and cumulative retrieval becomes non-negotiable, spaced return across everything taught so far, since the examinations sample the whole map while the timetable moves on. This is where upload-based systems, Neural Consult-style, earn consideration for institution-specific material, with the generated-question caveats and the sequencing logic at /blog/neural-consult-vs-iatrox-upload-curriculum-vs-blueprint-qbank.
Clinical years: the evidence-and-rehearsal years
The problem changes shape entirely: real cases generating real questions, presentations to survive, differentials to build, consultations to rehearse. The stack: a grounded clinical evidence tool for ward questions, OpenEvidence-style literature search or askiatroX's guideline-grounded answers by jurisdiction and taste, run inside the four-step loop at /blog/openevidence-for-medical-students-rotation-revision; a virtual-patient layer, Geeky Medics-style AI patients or institutional SimConverse access, for history-taking and difficult-conversation rehearsal, judged against the ten standards at /blog/ten-standards-realistic-ai-patient-simulation; and the question bank continuing underneath, because clinical years erode preclinical knowledge silently and cumulative retrieval is the maintenance contract. The confidentiality layer governs everything here: the never-paste rules travel onto every ward.
Finals: the calibration year
The problem: performing against a specific blueprint, under time, unaided. The stack contracts to its sharpest form: an exam-specific bank with adaptive targeting and full mocks at true format, UKMLA, USMLE, MCCQE or AMC per your track, which is the configuration iatroX is built as and the season it is built for; the tutoring layer for interrogating persistent errors rather than re-reading around them; and a planning layer, the mock-stability and weak-domain-floor dashboard from /blog/why-qbank-percentages-are-not-comparable replacing every magic-number anxiety. Comprehension tools retire to reference duty; the general learning mode narrows to concept repair; and every week's final act is the same, unassisted, blueprint-mapped performance, because that is the condition being purchased with all of it.
Frequently asked questions
How much should the whole stack cost per year?
Free layers first and fully, then one paid layer matched to the year's problem; most years are servable at one subscription, and finals justifies the exam-specific spend that earlier years do not.
What if my school provides tools centrally?
Provided tools take their functional slot and free your budget; they do not change the problems-by-year logic, and a provided tool in the wrong slot still leaves the slot empty.
When should the stack be reviewed?
Each academic year and at each phase transition, with one question: what is this year's binding problem, and does each tool still map to one? Anything unmapped is cancelled without sentiment.
Does the stack change for graduate-entry programmes?
It compresses: the comprehension and transfer years overlap, so the question bank enters earlier and the cumulative-retrieval rule starts from month one, with the same finals contraction at the end.
Where do virtual patients enter for schools with early clinical exposure?
With the exposure: the simulation layer maps to patient contact, not calendar year, and early-contact curricula justify the Geeky Medics-style layer from first year, run inside the close-the-loop workflow.
Should I keep tools I love that no longer map to a problem?
Affection is data about last year: keep one legacy tool at most as reference, cancel the rest, and let the annual review re-admit anything that earns a slot back.
What is the single most common stack mistake?
Carrying the comprehension-heavy configuration into finals: the year the problem becomes calibration, video hours and explanation chat are comfort spending, and the contraction to bank, mocks and tutoring is the highest-yield reorganisation on this page.