Appendix F — For instructors
This page is for whoever runs the course next: the author, a colleague, or someone at another school. It holds the parts of the design that students do not need to read: the module catalog and its variants, the session structure, how to run the weekly resource log, how to ask a guest to speak, and how to adapt the course for another cohort. The design files themselves are in curriculum/ in the repository.
F.1 The modules
Modules are independent. Each is one Zoom session, tags the competencies it serves (AIM-1 through AIM-9 and AIM-X, defined in curriculum/competencies.md), and assumes no prior module. That is what makes the course survivable when a third of the class misses week two.
| ID | Chapter | Minutes | Competencies |
|---|---|---|---|
core-foundations |
How these systems work | 75 | AIM-1, AIM-2 |
core-centaur-cyborg |
Centaur or cyborg | 90 | AIM-5, AIM-X, AIM-4 |
core-appraisal |
Critical appraisal | 90 | AIM-2, AIM-9 |
core-bias-equity |
Bias and equity | 90 | AIM-3 |
core-patient-communication |
Talking to patients about AI | 90 | AIM-6, AIM-4 |
core-ethics-regulation |
Ethics, liability, and the device line | 75 | AIM-4, AIM-7, AIM-X |
| ID | Chapter | Minutes | Competencies |
|---|---|---|---|
enrichment-agentic-coding |
Agentic coding | 75 | AIM-5, AIM-2, AIM-4 |
enrichment-prompting-lab |
Prompting lab | 75 | AIM-5, AIM-2 |
enrichment-workforce-economics |
Will AI take your job? | 90 | AIM-3, AIM-7 |
enrichment-environment |
What AI costs the planet | 60 | AIM-3, AIM-X |
enrichment-rag-notebooklm |
Grounding AI in real sources | 60 | AIM-1, AIM-9 |
Eleven modules, eight slots. The surplus is deliberate. Swap by cohort interest.
F.1.1 The default assembly (eight sessions, two a week)
The order students see on the syllabus. Week 1 is orientation and stance: foundations, then centaur or cyborg. Week 2 is judgment: appraisal, then bias and equity. Week 3 is practice: agentic coding, then talking to patients, which is the most important session. Week 4 is consequences and closing: workforce or environment by the cohort’s choice, then ethics and regulation, which closes on AIM-X, what will you not delegate?
Coverage check: AIM-1 through AIM-6 all appear; AIM-4 recurs three times, AIM-5 twice, AIM-2 three times. That meets the coverage rules in curriculum/competencies.md.
F.1.2 Variants
One session a week (four slots). Run centaur or cyborg, appraisal, talking to patients, and ethics and regulation. This keeps the stance, the judgment, the main session, and the close. Make foundations pre-work reading and cover bias inside appraisal.
A hands-on cohort. Swap foundations and ethics for the prompting lab and the retrieval lab. Keep the ethics content as a five-minute opening to every lab.
A cohort interested in policy. Run both workforce and environment in week 4 and make foundations pre-work.
F.1.3 Three ideas that recur
Modules are independent, but three ideas recur across all of them. Name them in week one and call them by name whenever they reappear. They are what turns eight sessions into a course, and they are what students will still have when every tool taught here has been replaced.
- Proxies fail. Almost every AI failure in medicine is a stand-in that looked reasonable. Cost stood in for need in the Obermeyer algorithm. Light skin stood in for normal reflectance in pulse oximetry. Race stood in for unmeasured biology in eGFR. A benchmark score stood in for clinical performance in FDA clearance data. The question to train is always: what is this actually measuring, and what was it substituted for?
- Efficiency gains get eaten by volume. Per-unit improvement is cancelled out by growth in use, and this appears in three separate literatures. Google cut per-prompt energy while total emissions rose. AI made radiologists more productive and imaging demand grew faster. AI scribes cut documentation time per note, and the time went into seeing more patients. Efficiency is not the same as less.
- The published number is the optimistic one. Before-and-after pilots on self-selected early adopters beat randomized trials. Vendor benchmarks on curated hard cases beat real-world deployment. The gap between the most prominently reported statistic and the trial result is the most reliable teaching material in this field.
F.2 Running a session
Zoom attention decays faster than in a room, so every session runs in blocks with a change of activity at each boundary.
| Minutes | Block |
|---|---|
| 0 to 10 | Introduce the session and restate the three standing rules: no PHI, verify before it counts, you sign the note. |
| 10 to 20 | Demo or short didactic. Never longer. |
| 20 to 60 | Breakouts of four or five. Not larger; participation drops sharply above that. |
| 60 to 80 | Report-back and comparison across groups. |
| 80 to 90 | Close on the standing question: what does this change about how you’ll practice? |
Running notes, from the design constraints:
- Pre-work is account creation, not reading. Any session that needs an account says so a week ahead. Creating accounts live costs ten minutes and leaves students with unsolved account problems.
- Everything uses synthetic data. No real patient information enters any tool in this course, ever. Consumer tiers come with no business associate agreement.
- Ungraded means students attend only if a session is worth their time. No exams, no graded work. Every session has to justify itself on the day, which is why each one produces something: a worked case, a judgment, a scored rubric.
- Cost is an equity problem. Free tiers only. If a session would work better on a paid tier, it gets redesigned, not assigned.
F.3 Running the resource log
The student page has the instructions and the entry template. This section is what the instructor does.
F.3.1 What changed from earlier years
Kept as-is: the weekly schedule, three to five resources a week, three to five sentences per resource, complete freedom of topic.
Changed:
- Each entry answers three fixed prompts (what it says, what it changed for you, what it left open). Same length as before. The prompts ask for reflection rather than summary.
- Optional encouragement to read across areas. Students are encouraged, not required, to cover two or three of the listed theme areas across the block.
- Week 4 is synthesis, not reading. Students pick one theme from their weeks 1 to 3 entries and write a short piece about it, then present it in the last small group.
- A short AI-use note each week. One or two sentences on whether and how they used a chatbot and what it got wrong. This is deliberately low-effort. Its purpose is to make the tools discussable, not to control their use.
Grading load is unchanged or slightly lower, since week 4 replaces three to five entries with one short piece.
F.3.2 Learning objectives
By the end of the block, students will be able to:
- Find and select resources on AI in medicine from a range of source types: tutorials and courses, primary literature, reviews, policy and regulatory documents, and essays on ethics, equity, environment, or economics.
- State the main point of a resource accurately and concisely, distinguishing what it shows from what it argues.
- Say how a resource changed, sharpened, or challenged their own prior understanding.
- Identify an open question or unverified claim that a resource leaves behind.
- Synthesize their own reading across a block into a short piece on one theme, drawing connections and disagreements among sources.
- Use AI chat tools reflectively as a reading aid, noticing where they help and where they mislead.
F.3.3 Timeline
| Week | Deliverable | Small-group tie-in |
|---|---|---|
| 1 | 3 to 5 resources, three prompts each, AI-use note | Each student names one resource and one surprise |
| 2 | 3 to 5 resources, three prompts each, AI-use note | What connected to last week; what contradicted it |
| 3 | 3 to 5 resources, three prompts each, AI-use note | Which theme are you choosing for week 4, and why |
| 4 | One-page synthesis on a theme and a five-minute presentation | Presentations |
F.3.4 What to prepare
- A short menu of theme areas (the list on the student page) with two or three starter resources in each, clearly marked optional. Students who already know what they want will ignore it; students who cannot decide will use it.
- The entry template posted as a file.
- One exemplar week written by you: three entries showing the three prompts done well, including one on something non-technical (an ethics essay, a piece on energy use) so students see that breadth is welcome.
F.3.5 Grading
Complete or incomplete, with brief feedback in weeks 1 and 2 while habits form.
| Criterion | Meets | Falls short |
|---|---|---|
| Resources chosen | Specific, locatable, varied across the block | Only general news, or only one type |
| Prompt 1 (what it says) | Accurate main point in one or two sentences | Restated title or abstract |
| Prompt 2 (what it changed) | Names a specific belief, question, or assumption that shifted or sharpened | “It was interesting” |
| Prompt 3 (what it left open) | A real question, or something to verify | “More research is needed” |
| Week 4 synthesis | Draws on at least three of their own earlier entries | Disconnected from weeks 1 to 3 |
F.3.6 Running the small groups
Keep them small. Ask each student one question: “What is the most interesting disagreement you have found so far?” By week 3 most students will have one. Those who do not should be pointed toward a second theme area.
F.3.7 The reading companion, at three levels of effort
The student page carries four weekly chatbot prompts, one per week, with less help each week. Three ways to deliver them:
Level 1, posted prompts. Post each week’s prompt as text; students paste it into any chatbot. No infrastructure. You cannot see transcripts. This is what the student page assumes.
Level 2, a shared assistant. A Claude Project, custom GPT, or similar with the weekly system prompt and the entry template attached. Share the link; swap the prompt weekly. The simplest option that gives every student the same experience.
Level 3, a course chatbot. A small app you control with per-week prompts, transcript logging with consent, and a “flag this as wrong” button. Only worth it if you expect to run the course for several years or want the transcripts. The planned chat tutor described in the preface is this level.
Whichever level: in weeks 2 and 3 the interesting AI-use notes are the ones where the bot’s summary of something and the student’s own reading diverged. Ask about those in small group.
F.4 Asking a guest to speak
The design caps guests deliberately: one per session, two across the whole course. Every session pairs an idea with an activity, and a guest-heavy course turns back into lectures without anyone deciding to.
The most useful guest is someone who can talk about a tool actually deployed in the local health system in the last quarter. That is what a fourth-year audience wants and what a survey lecture cannot give. Good matches by module: ambient documentation as deployed (centaur or cyborg); disclosure conversations and a patient-facing tipsheet (talking to patients); how a tool gets approved and bought locally (appraisal); the device boundary and local policy (ethics); a proxy failure in a model that was really deployed (bias and equity); what AI is doing to one specialty’s job market (workforce).
The invitation should say: the date and time; a twenty-minute slot inside a ninety-minute session, never a fifty-minute lecture; the competency the session serves; Zoom only; about forty fourth-year students; ungraded; and the actual ask, bring one thing you deployed and one thing that went wrong. Sending the invitations as one batch with a dean’s co-signature gets more acceptances. Sending them one at a time does not.
F.5 Adapting the course for another cohort
Everything specific to the original offering (people, dates, local accounts) is kept out of the chapters on purpose, so the book can be reassembled elsewhere. To run it for your own cohort:
- Check the design constraints in
curriculum/design-constraints.mdagainst yours. The chapters assume about forty fourth-year students, four weeks, Zoom, ungraded, and free tiers. If you have fewer weeks, use the one-session-a-week variant above. - Choose an assembly. The default or one of the variants. Update the syllabus tables and the due-date placeholders.
- Confirm tool access. The chapters assume one general-purpose chatbot without a signup step. Several
[VERIFY]flags in the chapters are access details to confirm the week before: an institutional NotebookLM login, a student plan’s status, an episode still being online. The open flags are listed in one GitHub issue on the repository. - Set your own institution’s policy text where a chapter says to check it, for example the attestation language for ambient documentation in centaur or cyborg, and your state’s disclosure statutes in talking to patients.
- Choose your guests, if any, using the section above.
- Write the exemplar week of the resource log yourself. Students copy what they see.
Chapters are reviewed against six invented readers before they change; the Review personas appendix has them, and the How this book was made appendix says how the review runs. If you change a chapter, review it as the six.