MedExamHelperThe AI Stack for New Residents: Documentation, Evidence, Differentials, Prescribing, Revision
Starting residency is the first time your time is no longer your own. The notes you wrote as a student will not survive a take. The textbook you half-read last night will not survive the ward round. So you do what the smart seniors do: you build a stack. Here is the stack worth building, the five jobs it must cover, and the demand you should make of every tool before you let it near a real patient.
The five jobs, in order of risk
| Job | Where it happens | Worst failure mode |
|---|---|---|
| Documentation | After the take, before handover | Note loses information your senior needed |
| Evidence | On the ward, in clinic | Acting on a citation that does not exist |
| Differentials | Bedside, on the phone | Anchoring on the first plausible cause |
| Prescribing | At the desk, with a senior | Decision support used as decision-making |
| Revision | After the shift, on the train | Re-reading instead of retrieving |
Build the stack in this order. Risk goes up the list; control goes down it. Documentation and evidence are where AI most often gets the credit when it is actually a competent human driving it. Prescribing and differentials are where the temptation to outsource thinking is highest. Revision is the one job you can fully own from day one.
Job 1: Documentation
Demand: the note must read as if you wrote it after a coffee, not as if a machine wrote it after midnight.
The right tool for documentation is the one that lets you dictate in your own words, produces a structured note that matches your local template, and refuses to invent findings. A good documentation tool will quote back only what you said, in the order you said it, and will flag the sections you have left empty. It will not hallucinate an examination you did not perform. It will not paste last week's free-text into today's clerking because the patient has the same surname.
You still own the note. The tool drafts; you sign. If a tool ever produces output you cannot defend at a coroner's inquest, throw the tool out. Documentation AI that you cannot audit is more dangerous than no AI at all, because the signature on the bottom is yours.
Job 2: Evidence
Demand: sources shown, sources live, sources you can open.
A tool that summarises the literature without showing its working is not an evidence tool. It is a confident stranger. The minimum acceptable behaviour is this: every clinical claim is followed by a link you can click, a year you can verify, and a study type you can name. If the tool cannot tell you whether the claim came from a randomised trial, a guideline, or a case report, you do not have evidence. You have a guess in a suit.
Treat evidence tools the way you treat a junior who is keen but unsupervised. Useful when they show their working. Dangerous when they don't. And never, ever, let them cite a paper you have not seen.
Job 3: Differentials
Demand: breadth before closure.
The single biggest cognitive error in early residency is premature closure: you hear "chest pain", you think "ACS", and you stop listing. A good differentials tool pushes you the other way. It will ask what else fits, what doesn't fit, and what would change your mind. It will not give you one answer; it will give you a ranked list with the reasoning visible.
The test is simple. After using the tool, can you defend the bottom of your list as well as the top? If you can, the tool is broadening your thinking. If you cannot, the tool has become a crutch, and the next patient will be the one where the rare diagnosis lives at the bottom of your list and never gets mentioned.
This is also the job where AI-patient practice earns its keep. Running real cases against an actor that will not let you get away with a thin history is the closest thing to supervised practice that fits around a shift.
Job 4: Prescribing
Demand: decision support, never decision-making.
This is the job where you must be the most conservative. A prescribing tool should check your dose, your route, your interaction, your allergy, and your renal function, and then stop. It should not write the prescription. It should not suggest the drug. It should not second-guess your senior. If the tool ever produces a script that lands in the drug chart without your name on it, the tool is mislabelled.
The reason is not paranoia. It is the boring truth that every prescribing decision is a medico-legal decision, and the person who signs is the person who is accountable. A tool that flatters you with confident suggestions will, eventually, be the tool that helped you miss the one interaction that mattered.
Use prescribing AI the way you use a calculator. Verify the input, check the output, write the answer yourself.
Job 5: Revision
Demand: retrieval, not re-reading.
This is the one job you can fully control, and the one where the literature is most settled. Re-reading feels productive and is not. Spaced retrieval feels unpleasant and is. Whatever tool you choose for revision, it must force you to produce the answer before showing it to you. If it lets you passively flip cards, passively watch videos, or passively highlight text, it is not a revision tool. It is a procrastination tool with a subscription.
The format matters less than the forcing function. A flashcard with a single, atomic prompt you have to write out is better than a deck of two thousand cards you downloaded. A short case you have to work through, with feedback that cites your own transcript rather than a generic answer key, is better than a textbook chapter you skim.
This is why deliberate practice with cited feedback sits at the top of the stack. It is the only job where the loop is entirely under your control and entirely measurable.
Building the stack: the order
Start with revision in week one of residency. You will be tired, you will be on nights, and you will not have the cognitive bandwidth to learn a new documentation tool until you have slept. Revision is portable, offline, and impossible to do wrong in a way that hurts a patient.
Add documentation and evidence in week two, when you know which template your trust uses and which guidelines your consultants actually quote. Add differentials and prescribing support only after you have a senior who can audit your outputs and tell you when the tool is helping and when it is hiding your thinking.
What "good" looks like for each tool
| Job | The bar |
|---|---|
| Documentation | Dictation in, structured note out, no invented findings |
| Evidence | Every claim linked to a citable source |
| Differentials | Ranked list with reasoning shown |
| Prescribing | Checks your draft, never writes it |
| Revision | Forces retrieval before reveal |
If a tool cannot clear the bar above for the job it claims to do, do not buy it, do not download it, and do not let it sit in your stack where you will reach for it under pressure.
The honest summary
You do not need five subscriptions. You need five clear demands, and you need to know which job each tool is actually for. The stack that survives a first year of residency is the one where every layer pulls its weight, every output is auditable, and the hardest thinking — the differentials, the prescriptions, the signed note — is still done by a person whose name is on the bottom.