The MRCGP AKT is unusually unforgiving of an AI tutor's weak points: it turns on current UK guidance, it tests statistics as a distinct strand, and a fifth of it is UK-specific organisational knowledge. This is the implementation companion to our Pastest MRCGP AKT audit: where that piece tests whether the AI is trustworthy, this one shows how to use it so it sharpens reasoning rather than replacing it — and forces the currency check the exam demands.
The commit-first rule
Write your answer and a one-line rationale before Tutor mode opens, every time. On an exam that rewards current UK practice, the temptation to consult first and commit later is both strong and self-defeating: it converts retrieval practice into assisted reading. The rule is non-negotiable, and everything below assumes it.
Reusable prompts for a currency-dependent exam
Keep prompts generic; never paste proprietary question text. For a clinical miss: "Without the answer, name the discriminating feature between my choice and the key." For currency — the AKT-critical prompt: "What UK guideline supports this, and what is its date? Has it changed in the last two years?" For statistics: "Work this likelihood ratio (or NNT) step by step and show every step." For organisational content: "What is the specific UK framework here, and what is the source?" For transfer: "Give me one question on this concept in a different presentation." These push the tutor toward currency, appraisal reasoning and UK specifics — the three things the AKT most tests and an AI most easily fudges.
The statistics stress-test
Once a session, give the tutor a worked-numbers problem and make it show every step, then check the arithmetic. Critical appraisal is the domain where language models most often fail while sounding most convincing, and the AKT tests it as a distinct tenth of the paper. A tutor that hand-waves the intermediate steps, or produces a confident final figure you cannot follow, is one to distrust on the entire evidence strand — and you only find out by testing it on a problem where you already know the answer.
The verification habit that matters most here
For any clinical or organisational claim that would change how you answer, force a named, dated UK source, and check it. On an exam decided by currency, an undated answer is a guess in the costume of one. Route management questions through a citation-first system built on NICE, CKS, SIGN and NHS content — Ask iatroX returns the source attached, so the check takes a minute. Build this into every genuine miss and the largest category of avoidable AKT errors — confident, stale guidance — simply closes.
Misconception records and the weekly sample
Four lines per genuine miss: incorrect rule, corrected rule, one transfer question, one review date; for statistics misses, record the method, not just the number. Weekly, verify five random tutor outputs against primary sources (NICE, CKS, RCGP guidance, a statistics reference) and log discrepancies. The records are the durable asset; the audit keeps your trust calibrated as Pastest iterates its AI.
A seven-day pattern for busy trainees
Monday: 40 Pastest clinical questions, commit-first, reasoning prompts on misses. Tuesday: a statistics/appraisal block, every calculation checked step by step. Wednesday: a timed, unseen 40-question mixed block across all three strands in iatroX's free MRCGP AKT bank, no tutor. Thursday: an organisational-content session plus the weekly verification sample. Friday: 40 clinical questions, timed at one minute each, currency prompts on management items. Saturday: a full timed three-strand simulation, same-day review. Sunday: rest. Pastest explains and drills; iatroX measures and verifies; the non-clinical strands get protected, checked practice.
A worked statistics stress-test
The single most revealing five minutes you can spend on any AKT AI tutor is a statistics stress-test, because critical appraisal is the domain where language models most often fail while sounding most convincing. Give the tutor a worked-numbers problem — say, a screening test with a stated sensitivity, specificity and disease prevalence — and ask it to calculate the positive predictive value step by step, showing every stage. Then check the arithmetic yourself. What you are testing is not whether the final number is right; it is whether the tutor shows defensible working. A trustworthy response lays out the 2×2 logic, substitutes the numbers, and reaches an answer you can follow. An untrustworthy one produces a confident final figure with hand-waved steps, or — worse — quietly swaps sensitivity for specificity somewhere in the middle and still sounds authoritative. A tutor that cannot transparently work a likelihood ratio or a number-needed-to-treat is one to distrust on the entire evidence strand, whatever its clinical fluency, and you will only know by testing it on a problem where you already know the answer.
The currency trap, illustrated
The other AKT-specific failure worth a worked example is guideline currency. Pick a clinical topic where UK guidance changed in the last year or two, ask the tutor for the current first-line management, and check whether it is up to date or confidently behind. A model trained on older material answers fluently and wrongly, with no signal that it is out of date — and the AKT rewards precisely the current answer. This is why the workflow's verification step insists on a named, dated source for anything behaviour-changing: on an exam decided by currency, an undated answer is a guess in the costume of one. Route every management claim through a citation-first check, and the largest category of avoidable AKT errors — confident, stale guidance — simply closes. The two stress-tests together, statistics and currency, probe exactly the two places an AKT AI tutor is most likely to be both wrong and convincing.
Continue, supplement, switch or stop
Continue if the routine holds and your verification log stays clean — especially on the statistics and currency checks. Supplement with dedicated evidence and organisational practice whenever those strands lag, and with unseen timed blocks when explanations satisfy but measured performance is flat. Switch only on logged failures — repeated stale guidance, confident statistics errors. Stop all AI assistance in the final fortnight's simulations.
Three mistakes this workflow is designed to stop
First, skipping the commit step because the tutor is right there. Every consultation before you have written an answer and rationale converts a retrieval attempt into an assisted read — it feels efficient and quietly deletes the practice effect the AKT rewards. Second, accepting fluent, undated answers on management questions. The AKT turns on current UK guidance, so any behaviour-changing claim without a named, dated source is a guess in the costume of one; force the currency check every time. Third, trusting the tutor on statistics without making it show its working. Critical appraisal is the strand where a language model is most likely to be both wrong and convincing, so a confident final figure with hand-waved steps should lower your trust, not raise it. Run the statistics stress-test and the currency check as standing habits, because together they probe exactly the two places an AKT AI tutor most often fails while sounding most authoritative — and those two places are a disproportionate share of borderline failures.
Frequently asked questions
Is Pastest enough for MRCGP AKT on its own? Its bank and AI can anchor the clinical strand, but the evidence and organisational thirds need deliberate, verified practice, and the exam's dependence on current UK guidance means every AI answer needs the currency check this workflow builds in.
Which MRCGP AKT component does Pastest not reproduce well? No commercial AI tutor reliably handles statistics/appraisal reasoning and UK-specific organisational content under exam conditions — the two strands where confident wrong output is both most likely and most costly.
How should I verify Pastest AI answers for MRCGP AKT? Force a named, dated UK source on clinical and organisational claims, check it directly or via Ask iatroX, demand full working on statistics items, and audit five outputs weekly with a written log.
When should I stop using Pastest and move to mixed mocks? When all three strands meet their floors, accuracy is stable and pace is under a minute per item — the final two to three weeks, given to full timed simulation with the tutor closed.
How should I combine Pastest with iatroX without duplicating practice? Pastest for commit-first clinical drilling and reasoning prompts; iatroX (free for MRCGP AKT) for unseen adaptive measurement across all three strands, Socratic repair, and cited UK-guideline verification — the check that matters most on a currency-driven exam.
Editorial notes and references
Written by Dr Kolawole Tytler, NHS GP and founder of iatroX. Last checked 19 July 2026; the AKT format (from October 2025: 160 questions, 2 hours 40 minutes) and weightings are per the RCGP; Pastest features are vendor-published and evolving. Disclosure: iatroX operates a free competing MRCGP AKT bank and a Socratic Tutor. Corrections via the feedback route on iatrox.com. References: RCGP Applied Knowledge Test pages (rcgp.org.uk); Pastest AKT pages (pastest.com); related reading: the Pastest MRCGP AKT audit and why your Q-bank percentage is not your exam score.