Appendix D — How this book was made
This page is a short account of what this book is, how it was designed, how each chapter was written and reviewed before it reached you, and how AI was used in writing it. A course that asks you to disclose and verify your own AI use should do the same.
D.1 What this is
A four-week, pass/fail elective on AI in medicine for fourth-year medical students at the University of Colorado School of Medicine (IDPT 8079; 49 students in fall 2026), delivered over Zoom. The course director teaches it live in two recorded talks a week, and the book goes with the talks: eleven chapters, each standing alone, read on your own time. The other live hours are for office hours, guest talks, and small groups. It is built with Quarto. The source is in a GitHub repository, seandavi/ai-in-medicine-curriculum, which is private for now and will be made public.
D.2 How it was made
The design was done in the repository, and every step left a file or a GitHub issue behind. In outline, it went in four phases.
Design before drafting (August to early September 2026). The fixed facts came first: four weeks, Zoom only, pass/fail, no installs, free tools only. Then a review of eleven published competency frameworks, from which the course’s own competency list (AIM-1 to AIM-9 and AIM-X) was built. The frameworks students most need are summarised in What the field expects. Then the course director’s earlier lectures were catalogued, one session plan was written per module, and a chapter pattern and six imagined readers were defined. One chapter, critical appraisal, was written first as the model for the rest.
Drafting at scale (September 2026). The other ten chapters were drafted at the same time by ten AI agents working from one brief. Each looked up its citations in PubMed, wrote an answer key, and reviewed its own chapter as the six readers. The author then read the whole book as deployed.
Passes over the whole book (September 2026). What the author found on that reading set the next steps. Each was a pass over every chapter with one narrow purpose, run by AI agents in parallel: plain language; a second review as the six readers; links between chapters and consistent terms; direct statement in place of metaphor; and working through the claims still flagged as unchecked. A script confirmed that the language passes changed wording only, not numbers, citations, or figures.
Fitting the book to the course (October 2026). Once the dates and enrolment were fixed, the course director chose the format: week 1 taught live, guest talks in weeks 2 to 4, Friday small groups, and open labs. The book followed. The syllabus was rewritten. A page on patient information was added, after the institution confirmed which AI accounts are covered for patient care. Added at the same time: an appendix of open labs, starter resources picked from more than 650 links students submitted in earlier years, a section on how ambient scribes work, and a new visual design with clickable self-checks. Research notes compared the book with Stanford’s AI-in-education materials, with published surveys of what medical students want to learn, and with the competency frameworks. The coverage table in the preface comes from that last review.
Then an independent review, run by a fresh AI agent with no stake in the book, found what the author had suspected: the book was long, cautious, and gave students little practice actually using a chatbot. Students the year before had said they were afraid to use these tools and felt lost. So the course was refocused on skill and confidence: one core chapter a week plus a menu, a weekly routine, and most of each week spent using AI to learn something the student cares about. The chapters are being revisited one at a time to match. Then, on October 8, the course director chose to teach more live. Weeks 2 to 4 now have two recorded talks a week in place of the weekly open lab first planned, and each core chapter goes with its talk.
D.3 What the ledger shows
The AI ledger records each piece of AI-assisted work and what a person checked. Read as a whole, it shows four things.
The division of labor held. The author decided what to teach, how the course runs, and what the book may claim about local policy. AI tools drafted, looked things up, and checked drafts against the book’s standards. The author read the result and decided what to publish.
Checking found real errors, and each fix was recorded. While the visual design was being changed, two charts turned out to misstate their data: the charting tool drew bars from the bottom of the axis, so a 4 percent fall looked like a rise. A sentence about privacy law said more than the law does. A research agent described a set of national principles as rules for students, which they are not. A fill-in template disappeared on the dark screen setting until a person noticed. Each is in the ledger with its fix.
Where a check could not be done, the book says so. Some publisher pages block automated tools, some podcast pages give no length, and some sign-up rules could not be confirmed. Those stay flagged in the text or noted in the ledger, not filled in with a guess.
Some things could only come from people. Which tools the university covers for patient care, which scribe the local health system uses, what fourth-year students are worried about this month, and which guests are coming: these came from the author and colleagues, and the book marks them as local facts.
Slides for the live session. The foundations chapter also exists as a slide deck for the live Zoom session. The deck follows the session plan rather than the chapter’s order. It adds what a page cannot do: small programs that run in the browser (two sliders that fit a curve by hand, a model sorter, a slider that shows overfitting, a real tokenizer, a next-word sampler), a slide on asking the chatbot to check its own citations, a further-exploration list (with the TensorFlow Playground), and a short video. The speaker notes are written to be read aloud, so the deck can be narrated for students who miss the session. A script pulls the narration out of the notes. The deck was reviewed by the same six readers as the chapters, in three rounds, until none of them would put it down.
D.4 How chapters are reviewed
Every chapter is read as each of six invented readers, twice: once by the agent that drafted it, before its pull request, and once in the review pass, after all eleven were merged. Three of the readers are students; three are the people who would adopt, adapt, or clear the course. The full descriptions, with each reader’s review questions, are in the Review personas appendix. The review files, one per chapter per pass, are in reviews/ in the repository, and the open items from the second pass are GitHub issues labelled pass.
| Reader | Role | What they catch |
|---|---|---|
| Elena | MS4, avoids AI, free tiers only | Chapters that assume you already use these tools, or read as enthusiasm |
| Marcus | MS4, uses AI daily, matching radiology | Chapters that are all warnings, or slower than the tool itself |
| Priya | MS4, wants to use it safely on the wards | Practical guidance buried under mechanism |
| Jordan | Instructional designer | Objectives with no matching activity; dishonest time estimates; images without alt text or with no purpose |
| Dr. Okafor | Educator at another school, not an AI expert | Chapters that cannot be taught without the author in the room |
| Dr. Lindqvist | Clinical ethicist | Activities an IRB would question; harm without a named population; false certainty; unlicensed or misattributable images |
The order matters: Elena reads first, because if the avoider stops reading, nothing after matters, and Dr. Lindqvist reads last. When readers disagree, students win on content, faculty win on form, and the ethicist wins on harm.
The pages added in October 2026 (the patient information page, the open labs and starter resources appendices, and the new section on ambient scribes) have not yet been through the six-reader review. The ledger marks them as awaiting the author’s review.
On 12 October 2026, the first day of the course, seven AI agents reviewed the whole book, the course pages, and the transcript of that day’s session. The author made decisions from that review the same day. The fixes are being made in batches, each before the week of the course it affects.
Two further checks run on every chapter:
- Citations resolve or the build fails. Every reference is written as a PubMed ID or DOI. Before each render, a tool (quartobot) looks each one up and writes the bibliography. A key that does not resolve stops the render. This does not prove a paper says what the chapter claims; it proves the paper exists, which is the failure mode this course spends a chapter on. The agents that drafted the chapters also read each paper’s PubMed abstract before using its key, and their ledger rows say which numbers were checked against an abstract rather than a full paper.
- Unverified claims stay flagged. Research notes carry
[VERIFY],[UNVERIFIED], and[CONFLICTING]tags. They are copied into chapters unchanged and are not removed until a person has checked the claim. The flags still open across the book are listed in one GitHub issue, grouped by how much a mistake would matter.
Answer keys are kept in _answers/ and are never rendered. They are notes for instructors; nothing grades students with them. They were drafted by the same agents as the chapters, and the ledger says where an expected result is a prediction rather than a recorded run.
D.5 On the use of AI
This book was drafted with an AI coding assistant (Claude Code, running Claude models), working in the repository alongside the author. The division of labor has been consistent: the author decides, the model drafts, the author edits and commits. That covers most of the prose in this book, the research notes, the module plans, the answer keys, the review files, and the review personas above. It does not cover the decisions those files record, which are the author’s: the competencies, the constraints, the four-week structure, the choice of reviewers, the rule that every chapter opens with a clinical scenario, and the order and contract in the chapter plan.
What changed as the book grew was scale, not the division of labor. One chapter was drafted in conversation. The other ten were drafted by ten agents working at once, in parallel, each on its own branch, from the same brief. The passes that followed ran the same way, one agent per chapter. The author’s part moved from editing each paragraph to writing the brief, reading the result, naming what was wrong, and deciding what to merge.
The model has been most useful for four things. Editing, which is bringing material written over several years into one consistent style. Drafting, which is producing first versions of explanations, activities, and answer keys from a module plan. Looking things up, which is retrieving the PubMed record for every citation, reading the abstract, and fetching the primary page behind a regulatory claim or a podcast episode. And holding the chapters to a standard, which is reading each one as the six reviewers above and reporting where it fails.
A language model produces fluent text that can feel authoritative while being wrong, and the appraisal chapter exists because that is true of citations in particular. So every reference in this book was resolved against PubMed by tool, every chapter was read by the author in the deployed book before the passes that finished it, and the author is responsible for what the book says. Where the review was less thorough than that, the ledger says so. Where a claim has not been checked by a person, the chapter says so with a [VERIFY] flag.
The AI ledger records, commit by commit, what the model did and what a person checked.