To turn AI viva feedback into a study plan, stop collecting comments and start converting them into behaviours. Every note from the model should become three things on paper: the domain you missed, the exact behaviour you want next time, and the date you will retest it. 'Needs more structure' is not revision. 'State likely diagnosis, severity, and first priorities in the first 20 seconds' is.
That shift matters because spoken exams reward repeatable performance under pressure. If your plan only says revise cardiology or practise communication, you will stay busy without getting much better.
Why this matters
Across current spoken clinical assessments, examiners score recurring domains and skills rather than a vague overall impression. Official examples differ in format, but the pattern is the same: the MRCGP Simulated Consultation Assessment (SCA) links feedback to marking domains; the MRCP(UK) Part 2 Clinical Examination (PACES23) expects candidates to meet the standard across seven skills; MRCOG Part 3 samples five domains across its tasks; and the MRCEM OSCE uses repeated short stations with standardised scenarios.
So the value of AI feedback is not the score it gives you. It is the pattern it reveals. RCGP's current SCA feedback guidance is explicit that selected feedback should guide future preparation, particularly when the same issues recur.
Your job is to find those repeats early. Then rehearse the smallest change that would have improved the answer.
Build an AI viva feedback study plan from domains, not comments
Start with a simple rule: never store feedback as a loose sentence. Tag it.
For each AI comment, write down:
- the exam domain or skill it belongs to
- whether it is a knowledge, structure, prioritisation, or communication problem
- the replacement behaviour you want to hear next time
- the drill that will train it
- the retest date
A short feedback log works better than a long transcript archive. One page is enough.
| AI feedback comment | What it usually means | Revision action |
|---|---|---|
| You rambled before answering | Weak opener, no hierarchy | Rehearse 60-second openings: diagnosis, severity, immediate priorities, next steps |
| You missed red flags | Unsafe or incomplete data gathering | Build three-question red-flag prompts for common presentations and drill them aloud |
| Your management plan was vague | You know the topic but do not state decisions clearly | End every answer with treatment, escalation, and safety net |
| You sounded formulaic | You are using stock phrases instead of responding to cues | Practise listening for concerns, summarising, and answering the actual question |
| Differential too broad and unfocused | Poor prioritisation rather than poor memory | Rehearse one-sentence prioritised differentials with a why-this-is-top explanation |
If a comment cannot be linked to a domain and a drill, ignore it for now. The takeaway: useful feedback is specific enough to change tomorrow's performance.
Separate knowledge gaps from performance gaps
This is where many trainees waste weeks. The AI says you were unsafe, unfocused, or unclear, and you respond by opening a textbook.
Sometimes that is right. Often it is not.
Four diagnostic questions before you revise
Ask yourself:
- Did I truly not know the medicine?
- Did I know it, but say it too late?
- Did I know it, but fail to prioritise it?
- Did I know it, but communicate it badly?
A 67-year-old with crushing chest pain is a good example. If the AI says, 'Good differential, but immediate actions came too late,' that is not mainly a cardiology reading problem. It is a prioritisation problem. Your drill is the first 30 seconds of every acute case: assess acuity, state immediate concerns, call for help if needed, start first-line actions, then expand.
Likewise, in a GP-style consultation, a strong factual answer can still fail if you do not explore the person's agenda or give a clear safety net. That is not fixed by reading more NICE guidance alone. It is fixed by practising the opening and closing phases of the consultation until they become deliberate.
Read for genuine gaps. Drill for everything else.
Turn one weak domain into a drill
A useful rule is one weak domain, one narrow drill, five repetitions.
Say the AI feedback after a mock consultation is: good rapport, but you failed to explain uncertainty and did not safety-net clearly. Do not write revise headache. Write this instead:
- Domain: management and communication
- Likely cause: you spent too long gathering data and rushed the close
- Drill: four 8-minute closing drills on headache, back pain, cough, and abdominal pain
- Success metric: final minute always includes likely diagnosis, what you are ruling out, what the patient should do if things worsen, and when they should seek urgent help
- Retest: fresh case within 48 hours
The same logic works in hospital vivas. If a PACES23-style consultation keeps attracting the comment that your differential is broad but unfocused, your drill is not another full mock. It is ten one-minute summaries where you state the top diagnosis, two alternatives, and the feature that makes each more or less likely.
If an emergency medicine station repeatedly exposes chaotic starts, practise only the opening moves. A resuscitation answer should sound calm, structured, and safety-first before it sounds clever.
Make the weekly study plan
Most trainees do better with a small repeatable cycle than with a heroic timetable. Build your week around three jobs: audit, drill, retest.
A workable template looks like this:
- Audit session, 20 to 30 minutes: review the last three to five AI transcripts and tally recurring errors
- Drill session, 30 to 45 minutes: practise one micro-skill at a time, such as acute openers, prioritised differentials, explaining risk, or safety-netting
- Retest session, 30 to 45 minutes: do fresh cases under exam timing and look only at whether the target behaviour appeared
- Repair session, 20 minutes: read around factual gaps exposed by the mock, then return to spoken practice
If you are time-poor, use a 60/30/10 split:
- 60% on repeated high-risk performance errors
- 30% on factual gaps
- 10% on polish, phrasing, and fluency
That weighting is deliberate. Repeated safety, structure, and prioritisation errors usually cost more marks than a rare missed detail.
Keep one running dashboard with six columns:
- case topic
- domain missed
- exact behaviour to improve
- drill used
- retest date
- result
You should be able to glance at the page and know your next session. If the dashboard is longer than one page, your plan is too complicated.
Use AI, but don't outsource judgement
One current college guidance on generative AI in training says AI can widen access to feedback and support learning, but it is not a substitute for supervisor or expert feedback. It also says outputs should be challenged, verified, and fact-checked. That is the right mindset for viva revision too: use AI as a prompt engine, mock examiner, and pattern spotter, not as final truth.
Useful ways to use it include:
- asking for three highest-yield corrections rather than a wall of feedback
- requesting a domain-based score sheet for each answer
- generating five fresh stems that test the same weakness
- rewriting your answer as a tighter 45-second opener
- playing the role of a difficult patient, parent, or examiner
Less useful ways include chasing an overall score, asking it to tell you whether you would definitely pass, or copying its polished phrasing without understanding the clinical reasoning.
Be careful with confidentiality and exam security. GMC guidance says doctors have ethical and legal duties to protect patients' personal information, and RCGP guidance on AI use in training says registrars must stay within data protection and local organisational requirements. Separately, the RCGP misconduct policy for the SCA explicitly prohibits distributing or uploading exam content and sharing it outside the immediate educational sphere. In practice, that means you should use de-identified or fictional cases, avoid pasting real patient material into public AI tools, and never upload recalled exam cases or questions.
The practical takeaway is simple: protect patients, protect the exam, and verify what the model tells you.
Common mistakes
- keeping transcripts but never tagging patterns
- revising whole specialties when the real problem is answer structure
- treating every AI comment as equally important
- correcting only knowledge gaps and ignoring unsafe prioritisation
- reusing the same case until you memorise it instead of retesting on fresh stems
- practising with AI alone and never checking yourself against a trainer, peer, or consultant
- uploading identifiable clinical details or recalled exam content
Practice workflow
A simple debrief loop works well:
- Run one or two timed cases with AI.
- Mark them against the same domains every time.
- Pick one weakness only.
- Do a short drill five times.
- Re-test on a new stem.
- Write one sentence after the session: next time I will do this earlier, more clearly, or more safely.
Try to keep one human calibration point each week, even if brief. AI is good for volume. A trainer or peer is better for nuance, realism, and spotting habits that do not show up well in text.
If you are close to the exam, narrow the plan further. In the final stretch, you want fewer targets, more repetition, and cleaner answers.
Summary
- Turn every AI comment into a domain, a behaviour, and a retest date.
- Prioritise recurring safety, structure, and communication errors before rare factual misses.
- Use narrow drills, not vague intentions like revise more respiratory.
- Re-test on fresh cases so you measure improvement, not memory.
- Use AI critically and safely: verify its claims, protect confidentiality, and keep a human debrief in the loop.
References
- https://www.rcgp.org.uk/mrcgp-exams/wpba/generative-artificial-intelligence
- https://www.rcgp.org.uk/training-exams/mrcgp-exam/mrcgp-recorded-consultation-assessment
- https://www.rcgp.org.uk/mrcgp-exams/simulated-consultation-assessment/toolkit
- https://www.rcgp.org.uk/mrcgp-exams/simulated-consultation-assessment/marking-and-results
- https://www.rcgp.org.uk/mrcgp-exams/simulated-consultation-assessment/feedback-statements
- https://www.rcgp.org.uk/mrcgp-exams/misconduct-for-akt-sca
- https://www.mrcpuk.org/sites/default/files/documents/Candidate%20guide%20notes.pdf
- https://www.mrcpuk.org/sites/default/files/documents/PACES23%20Consultation%20scenario%20writing%20guidance.pdf
- https://rcem.ac.uk/exam-regulations-policies/
- https://www.rcog.org.uk/careers-and-training/exams/mrcog-our-specialty-training-exam/mrcog-part-3/mrcog-part-3-format/
- https://www.gmc-uk.org/professional-standards/the-professional-standards/confidentiality/ethical-and-legal-duties-of-confidentiality