How to Study for Step 2 CK with AI: A 2026 Guide

TL;DR

Step 2 CK is now the scored USMLE exam that residency programs actually see, and since May 7, 2026 it has been delivered as sixteen 30-minute blocks of up to 20 questions in a single 9-hour day. The content is almost entirely clinical vignettes that ask what to do next: diagnosis, the right test, and management together make up roughly half of the exam by the USMLE’s own content specifications. The best-evidenced way to prepare is still a full pass through a major question bank (UWorld or AMBOSS), calibrated with NBME’s Comprehensive Clinical Science Self-Assessments, built on top of the shelf-exam studying you are already doing during rotations. AI does not replace any of that. Where it helps is the work the banks were not built for: generating extra vignettes on the specific topics you keep missing, turning your own misses into spaced-repetition cards without the hour of card-making, rehearsing clinical reasoning out loud, and checking a management question against a cited source instead of a confident chatbot guess. This guide walks through the exam, the incumbents, where an AI layer like Neural Consult earns its place, and a rotation-by-rotation plan that gets you to dedicated without starting from zero.

Why Step 2 CK needs a different plan than Step 1

Step 1 went pass/fail in 2022, which made Step 2 CK the first numeric USMLE score most programs see. The USMLE describes Step 2 CK as a test of whether you can “apply medical knowledge, skills, and understanding of clinical science” to patient care under supervision, with an explicit emphasis on health promotion and disease prevention. Scores are reported on a 1–300 scale; the sample score report for the 2024–25 cohort of first-time takers from LCME-accredited schools gives a mean of 250 and a standard deviation of 15. The minimum passing score is now 218, up from 214 as of mid-2025. Passing is not the hard part for most examinees: the USMLE’s performance data for 2024–25 show 98% of 22,249 first-time MD takers from US schools passed, 96% of 5,025 DO first-takers, and 90% of 16,186 first-takers from non-US schools. The hard part is the number, and the number is decided by how well you reason through clinical vignettes under time pressure.

Three things about the exam should shape your plan.

The format changed in May 2026. For exams on or after May 7, 2026, Step 2 CK is divided into sixteen 30-minute blocks with no more than 20 questions each, with a minimum of 55 minutes of break time and a 5-minute tutorial, still in one 9-hour testing session. The older format was eight 60-minute blocks of up to 40 questions. Shorter blocks mean more frequent resets and more frequent chances to burn break time; practice in 30-minute, 20-question sets rather than hour-long blocks so the rhythm on test day is familiar.

It is scored by task, not just by organ system. The USMLE’s content specifications break the exam down by physician competency as well as by system. On the current outline, Patient Care: Diagnosis is 16–20% of items, Laboratory/Diagnostic Studies 13–17%, Mixed Management 12–16%, Pharmacotherapy 8–12%, Clinical Interventions 6–10%, Health Maintenance/Disease Prevention 5–10%, and Prognosis/Outcome 5–9%, with Systems-based Practice and Professionalism around 5–7% each. By discipline, Medicine is 55–65%, Pediatrics 17–27%, Obstetrics and Gynecology 10–20%, Psychiatry 10–15%, and Surgery 5–15% (the ranges overlap because many items are integrative, and USMLE notes that all percentages are subject to change). History-and-physical-exam items are explicitly not tested here; they live on Step 1 and Step 3. In practice this means the question you will see over and over is not “what is the diagnosis” but “what is the next best step”: which test, which drug, which disposition, which screening interval.

It is cumulative with your clerkships. The NBME subject exams you take at the end of each rotation draw on the same clinical-science content as Step 2 CK. Every shelf you study for is Step 2 CK studying, provided you keep what you learned rather than letting it decay between rotations. That is the biggest planning difference from Step 1: there is no clean preclinical-to-dedicated handoff, and dedicated for Step 2 CK is typically only a few weeks, so the work has to be spread across the year.

The core resources, and what each one is actually for

None of the tools in this section is optional in the sense that the AI layer below is optional. Pick your question bank, buy at least two self-assessments, and build everything else around them.

UWorld Step 2 CK

UWorld remains the default bank for Step 2 CK, and the default is earned. Its Step 2 CK product page lists 4,250+ Step 2 and shelf-exam questions (the FAQ on the same page says “over 4,300”), with explanations that are the real product: a well-written UWorld explanation teaches the whole differential, not just the answer. Pricing at the time of writing is $349 for 30 days, $459 for 90 days, $519 for 180 days, and $579 for 360 days of QBank access, with the 180- and 360-day tiers adding a study planner and a one-time reset; QBank Plus adds videos for roughly $100 more per tier. UWorld’s self-assessment forms are sold separately at $50 each for two weeks of access. All tiers now include UWorld’s own AI assistant (“UAsk”), with a prompt cap that scales with the tier. Where UWorld falls short is the thing it was never designed for: when you have exhausted the bank’s items on a topic you are still weak on, there are no more. UWorld itself says a QBank average around 40% has historically been enough to pass, while advising you to aim for 45–50%.

AMBOSS

AMBOSS is the strongest alternative, and for many students a better second bank than a second pass of UWorld. Its yearly membership plus Qbank bundle is $448 on the US pricing page, which covers a pool of 13,900+ questions across USMLE, COMLEX and several specialty board exams and a library of 1,400+ articles; the Step 2 CK-specific count is approximately 3,500 according to a competitor comparison, since AMBOSS does not publish a per-exam number on its pricing page. The library is the differentiator. Every question links to the article it is drawn from, so a missed item on hyponatremia becomes a two-minute read of the algorithm rather than a 20-minute search. AMBOSS questions are widely described as closer in style to NBME items than UWorld’s, which is a reputation rather than a measurement, but it is a consistent one.

NBME self-assessments and the free official items

The NBME’s Comprehensive Clinical Science Self-Assessment (CCSSA) is the only practice material written by the same organization that writes the real exam. There are seven forms, each with four sections of 50 questions, priced from $62 per form, and each returns an equated percent correct, a performance breakdown by content area, and an estimated probability of passing. Treat them as calibration, not as study material: one form a few weeks before dedicated to set a baseline, then one each week or two during dedicated to track the trend. The USMLE also publishes free sample Step 2 CK items and an interactive testing experience on usmle.org; do those in the final week, because they are the only items that show you the exact interface and the newer formats (sequential item sets and abstract-based questions).

Anki and the rapid-review layer

Spaced repetition is the best-evidenced study method in this entire post, and the Step 2 CK version of the AnKing deck is how most students apply it. If you have been maintaining Step 2 cards through your rotations, keep going; the evidence that premade Anki decks correlate with Step scores is observational but consistent (we reviewed it in AI Flashcards vs. Manual Anki Decks). If you have not, do not try to start a 20,000-card deck six weeks out. Free rapid-review resources such as the Divine Intervention podcast are popular for a reason: they are organized around the “next best step” logic the exam rewards and they fit into commutes. None of these replaces questions.

Where an AI layer earns its place

The honest case for AI in Step 2 CK prep is narrow and specific. A general chatbot is a poor study tool for this exam for a documented reason: a 2024 study in JMIR found roughly 28.6% of references generated by a general model were fabricated, and a 2025 paper in Communications Medicine found that models repeated planted clinical errors in prompts up to 83% of the time. On an exam that is almost entirely about management decisions, a tool that confidently repeats a wrong management decision is worse than no tool. What follows is built on the assumption that the AI is grounded in sources, which is how Neural Consult is designed, and it is scoped to four jobs the banks above do not do.

Job 1: more vignettes on the topics you keep missing

By the end of a UWorld pass you know your weak areas; the bank just cannot give you more items on them. The Question Generator builds unlimited board-style vignettes by topic and difficulty, each with an explanation, so “ten more next-step-in-management questions on postpartum hemorrhage” is a request rather than a search. The relevant evidence here is specific to this exam: Neural Consult scored a perfect score on the USMLE’s published Step 1, Step 2 CK and Step 3 sample sets (119, 120 and 137 items respectively). That is a floor on accuracy with public items, not a prediction of anyone’s score, but it is the right floor for a tool that writes Step 2 CK-style questions. In a blinded study, raters scored AI-generated questions as equal or superior to NBME-style items, with explanations rated 4.4 versus 3.5 on a 5-point scale (p<0.001). The practical pattern: after each UWorld block, generate a short set on the two or three topics you missed, in 20-question, 30-minute sets to match the new format.

Job 2: turn your misses into spaced repetition without making cards

The biggest reason students abandon Anki during clinical year is not the reviews; it is the card-making. The Flashcard Hub generates spaced-repetition decks from your own uploaded material (rotation lectures, your shelf notes, a page of guidelines) and exports to Anki, so the cards you review are about the things you got wrong this week, scheduled by the same algorithm. An independent medRxiv preprint from Brown’s medical school tested a similar workflow with a different tool: 74% of first-year students reported time savings from AI-generated flashcards and 61% from AI summaries, with no significant difference in exam scores versus conventional materials. Read that as what it is: a preprint, about a similar tool rather than Neural Consult, showing that AI-generated cards save time without costing performance. The hours it saves are hours you can spend on questions, which is where the score gains are.

Job 3: rehearse the reasoning, not just the recognition

Step 2 CK has no patient in front of you, so a patient simulator is not a Step 2 CK tool in the way a question bank is, and we would rather say that plainly than oversell it. What the Case Simulator does is make you generate the differential, choose the workup and commit to a plan out loud with a voice-enabled AI patient, then score you against a rubric. The reason that matters for this exam is that recognition and generation are different skills: picking the right next step from five options is easier than producing it, and students who can produce it reliably pick it faster under a 90-second-per-question clock. It also keeps your reasoning warm on rotations where you see few of the bread-and-butter presentations the exam loves. One or two cases a week during the clinical year is enough; during dedicated, questions should take almost all of the time.

Job 4: settle a management question with a citation, not a vibe

Management questions are where question banks disagree with each other, with your attendings, and with last year’s guideline. AI Medical Search returns cited answers, including PubMed-indexed sources, so when UWorld and AMBOSS give different first-line agents you can see which guideline each one is following and when it was published. The AI Lecture Notebook does the same for your own material: upload the rotation’s lectures and ask it what your clerkship director actually said about the topic. For a fuller comparison of grounded search against a general chatbot, see Neural Consult vs. ChatGPT for Medical Students.

A rotation-by-rotation plan that ends in a short dedicated

The students who do best on Step 2 CK treat the clinical year as the study period and dedicated as the taper. A workable version of that looks like this.

During each core rotation. Do the UWorld (or AMBOSS) questions for that rotation’s subject as shelf prep, 20 to 40 a day on most days, timed, in 20-question sets. After each set, generate a short follow-up set on what you missed and let the Flashcard Hub turn the explanations into cards. Keep the cards from previous rotations in review; a Surgery shelf does not stop Medicine content from decaying. One Case Simulator encounter a week in the rotation’s specialty keeps the generation skill alive. By the end of the clinical year, this alone gets you through most of a bank with a card deck that is entirely about your own weak spots.

Four to eight weeks out. Take a CCSSA form at standard pacing to get a baseline and a content-area breakdown. The breakdown is the plan: whatever two or three content areas are lowest get the bulk of your generated questions and your reading in AMBOSS or UWorld’s library. If your school schedules a dedicated period of only two or three weeks, this is the point to start reducing clinical-year commitments where you can.

Dedicated (commonly two to four weeks). Finish the bank or complete a second pass of incorrects, at 60 to 80 questions a day in mixed, timed, 20-question blocks. Use a self-assessment every week or two (CCSSA forms, then UWorld’s forms if you want more) and plot the trend; a single form is noise. Use AI Medical Search only to resolve conflicts, not to browse. In the final week, do the free official sample items on the USMLE interface, stop generating new material, and sleep.

The specific daily counts above are common practice rather than evidence-based prescriptions; the structure (shelf prep as Step 2 prep, calibration before dedicated, timed mixed blocks in dedicated) is what matters. If you want the same approach written out for Step 1, our Step 1 dedicated schedule guide uses the same logic with the preclinical content.

Recommended stack

The bottom line

Step 2 CK rewards clinical reasoning practiced at volume, under time, across the whole clinical year. UWorld or AMBOSS plus NBME self-assessments remain the core, and nothing in this post should be read as a reason to skip a full pass through a bank. The AI layer is for the parts of that plan that break down in practice: running out of questions on the topics you most need, not having time to make cards, letting reasoning go stale on quiet rotations, and settling management disagreements with a source. Neural Consult does those four things with a grounded-retrieval architecture, a perfect score on the public USMLE sample sets, and a question study that puts its explanations ahead of the comparison items; Education Technology Insights listed it among its Top AI-Powered Medical Education Tools for 2026. More than 75,000 students use it alongside the resources above; if the Step 2 CK plan here looks like yours, it fits in the gaps rather than replacing anything.

Try Neural Consult for your Step 2 CK prep. For a head-to-head on the primary bank, read Neural Consult vs. UWorld for Board Prep, and for the evidence behind AI study tools in general, see Do AI Study Tools Actually Work?