This workflow is for GP ST3s using SCA Prep's AI voice roleplays and case generator who want a disciplined record–review–repeat loop rather than an endless stream of new cases. It addresses one job: turning AI consultation practice into calibrated, transferable performance across the RCGP domains. Its principal limitation is that an AI patient and an AI grade approximate, but do not replace, a trained role-player and a human examiner.
What SCA Prep offers for the MRCGP SCA right now
Vendor-reported figures, captured at the last check. Verify anything price- or count-sensitive on the product page.
| Item | What SCA Prep reports (last checked 19 July 2026) |
|---|---|
| Format | An AI SCA platform: an AI tutor (several modes, including WhatsApp), an SCA case generator, AI voice roleplays and a full mock exam |
| Case volume | "100+ realistic practice cases" plus unlimited AI-generated cases across the twelve domains, with patient information, notes and mark schemes and PDF export (vendor-reported) |
| AI roleplay | Voice-based AI patients with distinct personalities, designed to understand various accents including IMG candidates, that may withhold information, interrupt or respond to empathy; twelve-minute timed consultations |
| Feedback / scoring | Marking across the three RCGP domains with pass bands (Clear Fail, Fail, Pass, Clear Pass) and examiner-style coaching; an optional review of your own recorded real consultations (with consent) |
| Price | Tiered at last check — a revision tier around £14.95/month, an AI-actors tier around £34.95/month with a monthly credit allowance, and a three-month intensive around £95 with a larger credit block and a full mock; credits valid around a year; money-back window noted. Verify current pricing and credits |
| SCA components addressed | All three domains, with a full twelve-station mock at exam timing |
SCA Prep is closer to a full synthetic simulator than the record-and-review or course tools in this series: it generates cases, voices the patient and returns a banded score. That makes the record–review–repeat loop below natural, and it pairs with the companion SCA Prep SCA simulator audit rather than repeating it.
The exam you are actually preparing for
The SCA is twelve simulated remote consultations of twelve minutes each — 144 minutes total — sat in ST3, with nine diets a year and a fee of about £1,207. Each is examiner-judged across three domains: Data Gathering and Diagnosis; Clinical Management and Medical Complexity; and Relating to Others.
Cases are blueprinted across twelve Clinical Experience Groups, and the RCGP is explicit that not every group appears in every diet, that a case can span several groups, and that no single case tests all three domains — they are covered across the whole exam. The exam cases are unseen, and the score is a human judgement. An AI simulator that generates unlimited cases is powerful for volume, but that same abundance is a trap if you never preserve unseen material or calibrate the AI's grade against a human. The loop below is designed to prevent both.
Step 1 — Build a balanced case set with the generator, not a random stream
SCA Prep's generator will happily produce cases forever. Constrain it. Build a grid of the twelve groups against four variables — acuity, patient age band, complexity or comorbidity, and communication challenge — and generate deliberately to fill it.
| Clinical Experience Group | Vary acuity | Vary age | Vary complexity | Vary communication |
|---|---|---|---|---|
| Under 19 | Routine to urgent | Infant to adolescent | Safeguarding | Parent present |
| Reproductive & sexual health | Screening to acute | Teen to peri-menopausal | Contraception + comorbidity | Confidentiality |
| Long-term condition | Stable to decompensating | Working-age to elderly | Polypharmacy | Adherence |
| Older adults | Routine to acute frailty | 70s to 90s | Frailty, end of life | Carer, sensory loss |
| Mental health | Low risk to crisis | Any | Physical–mental overlap | Risk assessment |
| Urgent / unscheduled care | Emergent | Any | Red-flag exclusion | Time pressure |
| Health disadvantage / diversity | Any | Any | Language, homelessness | Interpreter, beliefs |
| Undifferentiated disease | Any | Any | Diagnostic uncertainty | Managing not-knowing |
| Prescribing | Routine to urgent | Any | Interactions, monitoring | Deprescribing |
| Investigation / results | Routine | Any | Incidental findings | Uncertain news |
| Professional dilemma | Any | Any | Ethics, capacity | Disclosure, complaint |
Generating to a grid turns "unlimited cases" from a distraction into coverage. The same discipline underlies the blueprint-coverage matrix framework: volume is not coverage.
Step 2 — Record the first attempt cold, and preserve unseen cases
Run each AI roleplay for the full twelve minutes without pausing, restarting or reading the mark scheme first. Let the AI patient interrupt or withhold; that is the point. The cold run is your baseline. Because the generator is effectively unlimited, deliberately ring-fence a preserved set — at least one full twelve-station mock, ideally two — that you generate but do not open until the final fortnight. Unlimited generation makes it tempting to treat everything as disposable; preserving unseen cases is what keeps a final mock honest.
Step 3 — Review by scoring twice, calibrated by observability
Score every roleplay twice: first take SCA Prep's banded grade and coaching, then score the same recording yourself against the three RCGP domains, ideally with a peer or trainer doing so independently. Log the disagreements.
Sort each AI comment by how observable the underlying thing is:
- Observable — clearly present or absent in the transcript: red-flag questions, a named-timeframe safety net, an explicit plan, checking understanding. The AI reads these most reliably; act on them once the clinical content checks out.
- Inferred — judgements the AI deduces from proxies: rapport, patient-centredness, shared decision-making. This is Relating to Others, and it is where an AI grade is weakest. Have a human confirm before acting.
- Generated — content the AI produces: a model plan, a suggested phrasing, the band itself. Check against current guidance and a human; the band is a starting hypothesis, not a verdict.
A voice AI adds two extra failure modes worth watching: transcription errors (it may mis-hear and mark you down for something you said correctly) and reward for verbosity (it may score fluent talking over efficient, safe consulting). Both live in the inferred and generated tiers and both need a human check. The underlying calibration method is set out in how to calibrate automated feedback before you trust the score.
Step 4 — Convert feedback into two observable behaviours
Choose two observable behaviours per case and park the rest. Swap "safety-net better" for "state a follow-up interval and one worsening trigger before closing". Swap "build rapport" for "reflect the patient's stated concern back in the first two minutes". Two behaviours tracked over a fortnight beat a ten-point list every time.
Step 5 — Repeat with deliberate variation
Do not re-run the same generated case. Regenerate the same clinical principle into a different group with a different agenda, comorbidity or time pressure. Shared decision-making learned in a statin case should be re-rehearsed in contraception, then in deprescribing for a frail patient. The AI makes this cheap; use that to vary the surface while keeping the principle fixed, which is exactly what transfer to an unseen exam case requires.
Keeping the clinical management current — where iatroX fits, and where it does not
An AI simulator can generate a confident management plan that is subtly out of date, because generated content is not guaranteed to track current guidance. Closing that currency gap is the single job iatroX does here. The boundary matters: iatroX is a question-bank and clinical-knowledge platform, not a consultation simulator. It does not voice a patient, grade your rapport or replace SCA Prep's roleplays and mock. It measures whether the knowledge under your consultations is current — unseen, SCA-style clinical MCQs and citation-first clinical answers grounded in NICE, CKS, SIGN, the SmPC/eMC and NHS content. Consult in SCA Prep; check the medicine in iatroX. And treat any score as a measurement, not a prediction: your Q-bank percentage is not your exam score.
A seven-day plan for a working ST3
SCA Prep does one job — generate, voice and band; iatroX does another — measure unseen knowledge. No proprietary predictive-algorithm claim is made for either, and credit use is kept deliberate.
| Day | SCA Prep (record + review) | iatroX (knowledge job) |
|---|---|---|
| Mon | Generate two cases in under-covered groups from the grid | 15 unseen MCQs in those topics |
| Tue | Cold-run both AI roleplays; do not read the mark scheme first | Check any management you doubted |
| Wed | Score both twice; log observable / inferred / generated disagreements | — |
| Thu | Regenerate one shaky principle into a new group | 15 mixed unseen MCQs |
| Fri | Use the AI tutor to close one specific knowledge gap | Recheck the one guideline you got wrong |
| Sat | Sit one preserved, unseen mock station end-to-end | 20-item timed block |
| Sun | Review the band-vs-human log; refill grid holes | Log recurring knowledge gaps |
Three mistakes this workflow is designed to stop
Chasing volume over coverage. Unlimited generation invites hundreds of similar cases and no safeguarding or dilemma practice; the grid stops that.
Trusting the band. A generated pass band is a hypothesis; adopting it without a human check trains you towards the AI's scoring quirks, including any reward for verbosity.
Never preserving unseen cases. If everything is disposable, you reach the exam with no honest final mock and a false sense of readiness.
Exit standard and the continue / supplement / switch / stop decision
Ease off when three signals hold across unseen cases: consistent rather than one-off performance; agreement between SCA Prep's band, your own scoring and a human reviewer, especially on the inferred tier; and no recurrent safety-critical omission.
Decide on gaps, not novelty or sunk credits. Continue if your grid is filling and unseen performance is rising. Supplement with live human role-play if Relating to Others stays weak, because that is where an AI grade is least reliable. Switch the primary tool only if it does not cover the groups you keep failing after an honest audit. Stop adding platforms once you have one consultation source, one human reviewer and one knowledge check. The comparison hub lines the options up if you are still deciding.
Frequently asked questions
Is SCA Prep enough for MRCGP SCA on its own? For consultation volume and a record–review–repeat loop it is a strong single tool, with AI voice patients, banded feedback and a full mock. But an AI grade needs human calibration and its generated content needs a currency check, so treat it as the practice engine within a stack rather than the whole preparation.
Which MRCGP SCA component does SCA Prep not reproduce well? Relating to Others is hardest for an AI to grade, because it infers rapport and shared decision-making from proxies and can reward fluent talking over safe, efficient consulting. Calibrate that domain with a human before you trust the band.
How many unseen SCA Prep cases or stations should I preserve for final MRCGP SCA calibration? Because generation is effectively unlimited, deliberately ring-fence one to two full twelve-station mocks — roughly twelve to twenty-four cases — that you generate but never open until the final fortnight, weighted to your weakest groups, so your final signal is a transfer test rather than recall.
When should I stop using SCA Prep and move to mixed mocks? Once single-case behaviours are stable and the band broadly agrees with a human reviewer, move to full, mixed, timed mocks in the final two to three weeks to test stamina and switching between unrelated cases.
How should I combine SCA Prep with iatroX without duplicating practice? Keep the jobs separate: SCA Prep for the consultation and its scoring, iatroX for unseen clinical-knowledge measurement and current-guidance checks, never re-answering seen items. This is the two-Q-bank rule in practice: add unseen material, not duplicates.
Editorial notes and references
Written by Dr Kolawole Tytler, NHS GP and founder of iatroX. Last checked 19 July 2026; platform figures are vendor-reported and captured at that date — counts, features, credits and prices change, so verify current details on the product page. Disclosure: iatroX operates a competing question bank and clinical-knowledge platform; its role here is confined to unseen MCQ measurement and current-guidance checks, jobs SCA Prep's simulator does not claim to do. Corrections are welcome via the feedback route on iatrox.com.
References: RCGP, Simulated Consultation Assessment — overview, preparing, and case content (rcgp.org.uk/mrcgp-exams/simulated-consultation-assessment); SCA Prep (scaprep.co.uk); iatroX MRCGP SCA bank (iatrox.com/mrcgp-sca); "Your Q-Bank Percentage Is Not Your Exam Score" (iatrox.com/blog/qbank-percentage-not-your-exam-score).