Appendix A — Open labs
An open lab is an hour spent trying something: a tool, a workflow, a way of working that is hard to picture until you have done it. The objectives are loose on purpose. The point is the experience, and what you notice.
The format borrows from NYU Langone’s Prompt-a-thon, where staff picked a project from a set of “project cards” and worked through it with a chatbot. Participants reported more confidence afterwards, whether or not they had been given sample prompts (Small et al., 2024). That was a survey after one event, so it shows the format was liked, not that it taught skills. The project cards are public if you want more ideas.
How they run. Each week of weeks 2 to 4, the class votes on one lab, and we do it together live for an hour on Zoom. The session is recorded. You can also do any lab on your own, at any time, from this page. Every lab is self-contained: it says what you need, it uses free or university tools in a browser, and it uses made-up material only.
How to vote. VERIFY: link to the vote (a Canvas survey or a Microsoft Form), opened each Monday Pick the lab you would most like to do live. The others stay here for you to try on your own.
Every lab uses invented patients, public documents, or your own made-up material. Do not swap in anything from a rotation. Your university ChatGPT Edu and Copilot accounts are covered for patient care, and these labs are not patient care.
What counts as patient information, and why taking out the name is not enough: the patient information page.
A.1 The labs
| Lab | You will have… | Time | Status |
|---|---|---|---|
| Hear it, draft it, check it | watched a scribe mishear a visit and a model invent part of the note | 60 min | Ready |
| Same question, three tools | seen how much the wording and the tool change the answer | 45 min | Ready, in the prompting lab |
| A study guide that cites its sources | built a notebook from papers you chose and caught what it adds | 45 min | Ready, in the grounding chapter |
| Build a calculator, then break it | built a working clinical tool in an hour and found what is wrong with it | 60 min | Ready, in the agentic coding chapter |
| Draw me a doctor | counted who an image model pictures when you ask for a doctor, a nurse, a surgeon | 30 min | Idea |
| The patient who talks back | practised a hard conversation with a chatbot playing the patient | 45 min | Idea |
| Deep research, checked | checked every citation in a chatbot’s “deep research” report | 45 min | Idea |
| Teach a machine in ten minutes | trained an image classifier in your browser and watched it fail on new pictures | 30 min | Idea |
| Write your own board question | written exam-style questions with a chatbot and checked the answer and every wrong option | 45 min | Idea |
| Build the tool you wish existed | described a small tool you would use in clinic and had an AI build a working version | 60 min | Idea |
| Ask questions of a dataset | analysed a public health dataset by asking questions in plain English | 45 min | Idea |
| A handout your patient can read | made a patient handout plainer, translated it, and checked what changed | 45 min | Idea |
“Ready” labs are written out in full. “Idea” labs are a short description for now; the ones the class votes for get written out first.
Have an idea for a lab? Post it in the course channel or send it to the course director. Good ideas get added here, and you may get to run it.
A.2 Hear it, draft it, check it: a scribe from the inside
You will play a clinic visit, turn it into a transcript, have a chatbot draft the note, and then check the note the way you would at 6:40 p.m. with fourteen notes to sign. The point is to see the two kinds of error from the ethics chapter: words misheard when speech becomes text, and sentences never said when a model writes the note.
Time. About 60 minutes. You need your university Microsoft 365 login (for Word for the web) and your university ChatGPT Edu or Copilot login. The visit is invented. Nothing installs.
The visit below is fiction, written for this lab. Do not replace it with a visit you saw on a rotation, and do not record a real patient. Your university accounts are covered for patient care, and this is coursework.
What counts as patient information, and why taking out the name is not enough: the patient information page.
A.2.1 The visit
Three speakers: the clinician (C), the patient, Mrs. Ellen Park, 79 (P), and her daughter (D). Read it at a normal speed, with the pauses and corrections as written. It takes about three minutes.
C: Good morning, Mrs. Park. What brings you in today?
P: The dizziness. It started Tuesday. No, wait, it was Sunday. Sunday after church.
C: Any chest pain with it?
P: No chest pain. No. And I'm not short of breath.
C: Are you still taking the hydralazine, twenty-five milligrams three times a day?
P: Yes. And I took some of my husband's hydroxyzine for the itching last week.
C: Okay. Anything else changed with your medicines?
P: I stopped the atorvastatin about a month ago. My legs ached.
D: She's been falling, though. Twice this month.
P: I didn't fall. I sat down hard. That's different.
C: Are you checking your sugars at home?
P: Usually around one fifty. Sometimes two fifty in the morning.
C: Let's hold the hydralazine if you're dizzy when you stand,
check your blood pressure lying and standing today,
and I'll see you back in two weeks.
D: We can't come back in two weeks. Her grandson's wedding is in Fort Collins.
C: Then three weeks, and call us if the dizziness gets worse.A.2.2 Part A: record and transcribe (10 minutes)
Two or three of you read the parts. Alone, read all of them and change your voice a little for each speaker.
- Open a blank document in Word for the web with your university login.
- Choose Dictate, then Transcribe, then Start recording. Read the visit. Stop, and let Word finish the transcript.
- Copy the transcript into a new document.
If Transcribe is not available in your account, use your phone’s voice-memo app if it makes a transcript. If neither works, skip to Part C and use the script above as a perfect transcript. You will lose the transcription errors, but not the drafting ones.
A.2.3 Part B: what was misheard (5 minutes)
Read the transcript against the script. Mark every place where it differs: a word misheard, a “no” dropped, a speaker label wrong, a correction lost. Drug names that sound alike are the first place to look.
A.2.4 Part C: draft the note (5 minutes)
Paste the transcript into ChatGPT Edu or Copilot, signed in with your university account, with this prompt:
You are an ambient documentation tool in an outpatient clinic.
Write a SOAP progress note from this visit transcript.
Use standard clinical abbreviations.A.2.5 Part D: check it as if you were signing it (25 minutes)
Go through the note sentence by sentence against the script. Put every problem in one of five boxes:
| Box | What it means | Which step it came from |
|---|---|---|
| Misheard | The transcript was wrong and the note repeated it | Transcribe |
| Never said | A statement that is not in the visit at all | Summarize |
| Left out | Something said that the note dropped | Summarize |
| Wrong person | Something the daughter said, written as the patient’s | Either |
| True but not yours | Correct, but not what you would sign, such as a plan you did not make | Summarize |
Then look hard at four lines: the start of the dizziness, the falls, the two “-zine” drugs, and the atorvastatin. Those are where this visit was built to go wrong. If your note gets all four right, compare it with a neighbour’s. The same prompt can give different notes.
Optional, 10 more minutes. Run Part C again, but paste the script instead of your transcript. Any error left now cannot be a transcription error. That shows you what the model adds on its own.
A.2.6 Part E: one rule (5 minutes)
Write one sentence: the rule you will use to read a scribe’s draft before you sign it. In the live lab, we collect these and compare.
In Mr. Alvarez’s case, the note said he “continues apixaban” after he said he had stopped it. Which step most likely produced that sentence?
- Record
- Transcribe
- Summarize
It is a fluent sentence that says what is usual for a patient with atrial fibrillation. A transcription error tends to look like a wrong word. A sentence that sounds exactly like a normal note, and is false, is the summarizing step’s kind of error.
A.3 Same question, three tools
Ask one clinical question three ways, in three tools, and put the answers side by side. You will see how much the wording of a question, and the tool you ask, changes what comes back. Written out in full as the prompting lab.
A.4 A study guide that cites its sources
Load a few published papers you chose into NotebookLM, ask it questions, and check each answer against the paper it cites. You will see what grounding in your own sources fixes, and what it does not. Written out in full in the grounding chapter.
A.5 Build a calculator, then break it
Describe a clinical score to an AI app builder and have it build a working calculator in the browser. Then try to break it with made-up patients at the edges. You will see how fast a tool appears, and how hard it is to know whether it is right. Written out in full in the agentic coding chapter.
A.6 Draw me a doctor
Idea. Ask an image generator for “a doctor”, “a nurse”, “a surgeon”, “a patient with a rash”. Make twenty of each and count what you see: age, sex, skin tone, setting. Compare with real workforce numbers. Connects to bias and equity.
A.7 The patient who talks back
Idea. Give a chatbot a made-up patient and a hard conversation: a new cancer diagnosis, a medication error, a patient who wants antibiotics for a cold. Have it play the patient while you play the doctor, then ask it to play an observer and critique you. Connects to talking to patients. NYU Grossman’s Prompt-a-Thon library has virtual-patient prompts to start from; copy them into your university ChatGPT Edu or Copilot.
A.8 Deep research, checked
Idea. Run a “deep research” query on a narrow clinical question. Then check every citation in the report: does the paper exist, does it say that, is it the best evidence? Connects to critical appraisal.
A.9 Teach a machine in ten minutes
Idea. Train an image classifier in your browser with a free tool and your webcam: two classes, a few dozen pictures each. Then change the lighting, the background, or the person, and watch it fail. Connects to how these systems work.
A.10 A handout your patient can read
Idea. Take a public patient handout, ask a chatbot to rewrite it at a sixth- grade reading level and then translate it into Spanish. Translate it back and compare. What changed, and what would a patient now misunderstand? Connects to talking to patients.
A.11 Write your own board question
Idea. Ask a chatbot to write five exam-style multiple-choice questions on a topic you are studying. Then check each one against a real source: is the keyed answer right, is each wrong option clearly wrong, and would the question teach anything? NYU Grossman’s Prompt-a-Thon library has question-writing prompts to start from. Connects to critical appraisal.
A.12 Build the tool you wish existed
Idea. Think of a small tool you wish you had on a rotation: a page of the formulas you use every day, discharge instructions matched to a patient’s reading level, a sign-up page for a clinic project. Describe it in plain English to an AI app builder and see how far you get in an hour. Then ask what you would need to check before anyone used it. The open-ended version of build a calculator, then break it; connects to the agentic coding chapter.
A.13 Ask questions of a dataset
Idea. Open a public health dataset in Google Colab, a free notebook that runs in the browser, and ask its built-in AI assistant to answer questions in plain English: how many, how has it changed, is this group different? It writes and runs the code. Then check one answer by hand, and ask whether the question was a fair one for the data. Connects to how these systems work.