This workflow is for GP trainees using AKTRevision's AI tutor who want it to sharpen reasoning rather than hand over answers. It targets the applied-knowledge core of the MRCGP AKT — the 80% clinical domain plus the evidence and organisational slices — by turning the tutor into a Socratic questioner. The principal limitation to respect: an AI tutor is a generative tool, so it can be confidently wrong, and every high-stakes output must be checked against official UK sources before you trust it.
What AKTRevision offers right now
Vendor-reported, last checked 19 July 2026 — verify on aktrevision.co.uk:
- Question bank: over 5,000 exam-style questions with 400+ fact sheets, spanning single best answer, multiple best answer, extended matching and free-text formats.
- Mocks: full-length 160-question papers in the current AKT format.
- AI tutor: a built-in tutor, vendor-reported at 30 messages/month with any paid membership, with an optional unlimited add-on (vendor-reported around £5/month).
- Learn courses: structured teaching (vendor-reported as 12 courses, 610 bite-size skills across 122 topics).
- Adaptive elements: spaced repetition and confidence calibration to resurface weak areas ("a question bank that learns you back," in the vendor's words).
- Price: vendor-reported at around £12.99/month (bank), £9.99/month (Learn) or £19.99/month (combined); built by the team behind SCA Revision.
Because the tutor is metered, prompt quality matters twice over: you want maximum reasoning per message.
The exam anchor
The AKT is 160 single-best-answer questions in two hours forty minutes, four sittings a year at Pearson VUE, weighted roughly 80% clinical / 10% evidence-based practice / 10% organisational (RCGP, 19 July 2026). The exam rewards higher-order application of current UK guidance at roughly one minute per item. An AI tutor helps only if it trains that applied reasoning against UK sources — so every prompt below is designed to test understanding, not to fetch a fact.
The one rule that makes the tutor work: commit before you open it
The anti-answer-leak rule: never open the tutor until you have written down your chosen answer and a one-line rationale. If you ask "what's the answer to this?" first, the tutor becomes a spoiler and you learn nothing about your own reasoning. Commit, then interrogate. Everything that follows assumes you have already committed an answer.
Prompt taxonomy
Group your prompts by the cognitive job: mechanism, discriminating features, option elimination, guideline/threshold verification, counterfactuals, and retrieval testing. The 25 prompts below map to the six commonest error types. They are reusable templates — fill in your own clinical variables and never paste proprietary question text into the tutor.
25 reusable prompts, grouped by error type
Error type 1 — Mechanism gaps (you guessed without understanding why).
- "Without giving me the answer, explain the underlying mechanism that links [finding] to [condition], then ask me to predict one other feature I'd expect."
- "Ask me a question that checks whether I understand why [first-line treatment] works in [condition], not just that it is first-line."
- "I think the mechanism is [my explanation]. Challenge it with one probing question rather than confirming or correcting it."
- "Give me a physiological reason this diagnosis could be wrong, and ask me to defend or revise my choice."
Error type 2 — Discrimination failures (you confused two similar conditions). 5. "Ask me one question that would reliably distinguish [condition A] from [condition B] in a UK primary-care setting." 6. "List the two most commonly confused alternatives to my answer and ask me which feature rules each in or out." 7. "Without confirming, ask me what single piece of history or examination would change my diagnosis from [A] to [B]." 8. "Pose a near-identical vignette where the correct answer flips, and ask me to explain what changed."
Error type 3 — Poor option elimination (you didn't rule out distractors). 9. "For each of the five options, ask me to state one reason it could be wrong before I commit." 10. "I ruled out [option] because [reason]. Test that reasoning with a follow-up question." 11. "Ask me which distractor is the 'best trap' here and why a rushed candidate would pick it." 12. "Without revealing the key, ask me to rank the options from most to least likely and justify the ranking."
Error type 4 — Guideline/threshold errors (you missed the current UK rule or number). 13. "Ask me to state the specific NICE or CKS threshold that applies here, then tell me to check it against the current guidance and report the date." 14. "Do not tell me the target. Ask me what monitoring interval current UK guidance recommends and how confident I am." 15. "Ask me which UK guideline governs this decision and what changed at its last update — I will verify independently." 16. "Prompt me to name the relevant SmPC/eMC caution for [medicine] before I finalise, without supplying it."
Error type 5 — Counterfactual blind spots (you didn't consider what would change the answer). 17. "Ask me: what one change to this patient's age, comorbidity or medication would change the correct management?" 18. "Give me a 'what would have to be true' question for my answer to be correct." 19. "Ask me which red-flag feature, if present, would override my current plan." 20. "Pose the same scenario in a different jurisdiction's system and ask why the UK answer differs."
Error type 6 — Retrieval and pacing failures (you knew it but couldn't retrieve it under time). 21. "Quiz me on this topic with three rapid recall questions, 20 seconds each, then ask what I missed." 22. "Ask me to reconstruct the management algorithm for [condition] from memory, then question the weakest step." 23. "Two days from now, ask me this same reasoning question cold and tell me to log whether I retrieved it." 24. "Ask me to explain this concept as if teaching an F2, then probe the part I over-simplified."
The jurisdiction and dating prompt (use on every clinical item): 25. "State plainly whether your answer reflects current UK MRCGP AKT practice and UK guidance, and give the approximate date of the guidance you are relying on. If you are not sure it is UK-specific or current, say so."
The misconception record
For every corrected error, write four lines: (1) the wrong rule you were using; (2) the corrected rule in your own words; (3) one transfer question that tests the same idea in a new context; (4) a review date two to five days out. This turns a single tutor exchange into a durable correction rather than a fleeting "ah, of course."
The weekly verification sample
An AI tutor can hallucinate. Once a week, take a small random sample — five to ten tutor outputs — and check them independently against NICE, CKS, SIGN or the relevant SmPC/eMC entry. Log any discrepancy. If you find a wrong or out-of-date answer, note the topic and treat that area with extra caution. This is the audit-an-AI-tutor discipline — grounding, leakage, hallucination and retention — applied in miniature, and it is the difference between using AI and trusting it blindly.
A worked seven-day plan for a busy trainee
Give AKTRevision one job: reasoning and retrieval on your weak topics.
- Monday: 25 AKTRevision items on a weak topic; commit answers first, then use two or three prompts from the relevant error type.
- Tuesday: build misconception records for Monday's errors; run the jurisdiction prompt on any guideline item.
- Wednesday: an unseen iatroX block on the same topic — never a replay — to measure transfer on questions you have not seen.
- Thursday: 25 mixed items at exam pace; use retrieval prompts (21–24) on anything shaky.
- Friday: the weekly verification sample; log discrepancies.
- Saturday: review all misconception records due; retest the transfer questions.
- Sunday: plan next week from the two error types that recurred most.
No proprietary-algorithm claim is needed: you learn and interrogate on AKTRevision, and you prove it cold on a separate bank.
Decision checklist: continue, supplement, switch or stop
- Continue if your verified tutor exchanges are producing corrected rules that survive to the unseen retest.
- Supplement with a second unseen bank for measurement, since a tutor you learn from should not also be your readiness gauge.
- Switch your approach if the tutor's answers repeatedly fail your weekly verification — the problem is trust, not effort.
- Stop buying more AI messages if you are not converting them into misconception records; unused reasoning is wasted spend.
Frequently asked questions
Is AKTRevision enough for MRCGP AKT on its own? For many candidates a bank of AKTRevision's size, with mocks and a working AI tutor, can serve as the primary resource — but "enough" still depends on your unseen first-attempt performance, not on message count or completion. The AI tutor adds real value for reasoning, provided you verify its outputs; the risk of relying on any single source is a blind spot it shares with you. Prove readiness on questions it has never taught you before concluding one platform is sufficient.
Which MRCGP AKT component does AKTRevision not reproduce well? No AI tutor reliably reproduces guaranteed-current organisational and guideline detail, because generative answers can lag or drift from the latest NICE, CKS or SmPC/eMC position. The clinical-reasoning training is strong; the weak point is trusting the tutor on precise thresholds, statutory rules and recent guideline changes without checking. Treat those items as verify-first, which is exactly what prompts 13–16 and 25 are for.
How should I verify AKTRevision AI answers for MRCGP AKT? Use the jurisdiction-and-dating prompt on every clinical item so the tutor states whether its answer is current and UK-specific, then run a weekly random sample of five to ten outputs against primary UK sources — NICE, CKS, SIGN and the relevant SmPC/eMC — and log discrepancies. Any answer that carries a specific number, threshold or legal duty deserves an independent check before you commit it to memory. Verification is a routine, not a one-off.
When should I stop using AKTRevision and move to mixed mocks? Shift to predominantly mixed, full-length, timed mocks in the final three to four weeks, once your weak topics are shrinking and your errors are pacing-related rather than knowledge gaps. Keep the tutor available for the occasional stubborn misconception, but by the closing weeks your practice should be dominated by timed papers that rehearse the real one-minute-per-item pressure.
How should I combine AKTRevision with iatroX without duplicating practice? Keep the jobs separate: learn and interrogate on AKTRevision's tutor and bank, then measure transfer on a free UK-core iatroX block using items you have never seen, so your readiness signal is uncontaminated. Do not mirror the same questions across both — the second bank exists to give an honest score, and your Q-bank percentage is only meaningful when it is unseen. One platform teaches; the other tests.
Editorial notes and references
Written by Dr Kolawole Tytler, NHS GP and founder of iatroX. Last checked 19 July 2026; AKTRevision's question counts, AI-message limits and prices are vendor-reported and change, so verify on aktrevision.co.uk. Disclosure: iatroX operates a competing MRCGP AKT bank with its own Socratic tutor; this article confines iatroX to the unseen-measurement job that a learning tutor should not also perform, and does not present it as superior to AKTRevision. Corrections are welcome via the feedback route on iatrox.com.
References: RCGP, Applied Knowledge Test, rcgp.org.uk/mrcgp-exams/applied-knowledge-test; AKTRevision, aktrevision.co.uk; iatroX, "How to Audit an AI Medical Exam Tutor"; iatroX, "Your Q-Bank Percentage Is Not Your Exam Score"; iatroX comparison hub.