Takeaways from a Workshop on Claude Code and Canvas | Krish Ramkumar, Pauli Murray ‘28

The workshop

Inoue had a clear picture of the flipped classroom he wanted. A student would log onto Canvas, open the announcement for that day, and find everything for the session gathered in one place: the pre-class video to watch, the graded quiz to complete beforehand, the worked solutions to the previous assignment, any notes relevant to the day, and the textbook sections to work through for extra practice. Assembling that by hand is *tedious*—it would mean writing a formatted announcement before every single class meeting, each one built through the same slow sequence of dragging and dropping course files, linking the right video, and setting open and due dates one at a time. For an instructor teaching several sections, Inoue’s Canvas dream could never come true.

The workshop showed how that Canvas dream can now be built in about twelve minutes. Inoue opened Claude Code, a leading agentic coding tool, in his terminal and handed it three things: his academic calendar, the web address of an open-access textbook, and a playlist of the lecture videos he had already recorded, each named for its topic. Claude Code read the textbook, worked out how it was organized, matched each video to the relevant sections, and drafted the daily modules along with short pre-class quizzes whose questions it pulled straight from the readings, leaving Inoue to review and approve each one. The finished shell came together in a single sitting, the announcements queued to publish on schedule. Inoue remarked that when he had run the same process on his real Differential Equations course earlier in the year, it “made no mistakes that I could tell.”

Inoue is quick to add that you do not need anything as ambitious as a flipped classroom to feel the benefits of connecting Canvas to Claude Code. The deeper convenience is being able to speak to Canvas in plain language instead of having to learn and maintain the platform by hand. A professor who cancels a single class for an illness, and then has to push back every reading, quiz, and due date that follows, could describe the change in a sentence and let the tool re-date the whole term in one pass. Moreover, this kind of plain-language access to course tools has arrived at a convenient time. A 2024 Department of Justice rule now requires the digital content of covered institutions to meet a recognized accessibility standard, WCAG 2.1 Level AA, and although the rule formally binds public bodies, it is widely expected to reach private universities like Yale through Section 504 and Title III.¹ Bringing years of untagged PDFs, uncaptioned videos, and inaccessible pages into compliance is exactly the kind of large, repetitive work that coding agents can grind through far faster than a person can.

Reproducing any of this, though, is less simple than twelve minutes of screen time suggests, and the gap is mostly technical. For running his real Differential Equations course, Inoue relies on a scoped API key, letting Claude Code reach into Canvas and build while deliberately restricting it to modifying the course’s structure, its modules, pages, and dates, but with no access at all to gradebooks or student enrollment. Even with those restrictions in place, Inoue did not want to test the workflow on a course full of real students protected under FERPA, so for the demonstration, he built everything inside a personal Google account on an empty Canvas instance, with no student data anywhere near it. Yale does not currently provide Claude Code to its faculty, and the Canvas@Yale Support Team has disabled the kind of scoped Canvas access that Inoue’s method depends on. In other words, Yale faculty do not have the access that would let the tool send announcements to students automatically during the academic year as Inoue does; they cannot point the tool at their real course site using their Yale account. You can watch a full walkthrough on the My Robot Teacher YouTube channel; though the walkthrough is marketed as a quick and easy twelve-minute miracle tutorial for a non-technical audience, do not feel ashamed if following along does not feel so seamless.

The interview

In conversation, Inoue’s conviction about how a teacher comes around to AI became clear: you find the one task they resent most, and you take it off their hands. Inoue learned this conversion method on himself. He had wanted to flip his classroom for years, convinced by a body of research most instructors now accept, that students in active classrooms outperform those sitting through lectures,² even though they tend to feel they have learned less.³ He never had the time, because keeping a flipped course running is that same relentless Canvas upkeep, and he had not, as he put it, gone to graduate school in mathematics “to become an expert at Canvas shell construction.” Over one winter break, he realized he could hand that job to Claude Code, and with the hours it gave back, *he finally built the classroom he had wanted all along.*

His own story is the whole argument in miniature: solve a problem a teacher already has, and curiosity about the tool follows on its own. There is an honest complication in holding him up as the example, which is that Inoue is more technical than he admits. He has written in LaTeX, the plain-text typesetting language mathematicians use, for most of his career, and that long habit of working in structured text is much of why an agentic tool felt natural to him so quickly. A colleague approaching this cold has no such head start, and closing that gap by yourself is exhausting: staying current with a technology that reshapes itself every few months, on top of the expertise you were actually trained in, asks a lot of anyone. The least exhausting way through it tends to be sitting down with someone who already knows the terrain, rather than working alone through a pile of tutorials.

Takeaways

1. Teachers before students
 

For a few years, the great hope for AI in education was student-facing: a tireless personal tutor for every learner, the technology that would finally individualize instruction at scale. That hype has mostly cooled down. The most prominent of these tutors, Khanmigo, built by Khan Academy on top of OpenAI’s models, has fallen well short of what its makers expected. Khan Academy’s founder, Sal Khan, now describes Khanmigo’s reception as “a non-event,” his chief learning officer reports that she does not see the transformation she had anticipated, and Khan himself has begun saying that the “biggest lever is really investing in the human systems.”⁴ The revolution aimed straight at students has stalled, and teachers remain only lightly engaged: faculty report generally warm feelings about AI while using it in their own teaching surprisingly little.⁵ The order many institutions follow can look backward: they write AI policy for students and roll out tutoring bots before their instructors have felt the technology do anything profoundly useful for them. The order should instead run the other way—*once* a teacher has felt the payoff, they will explore what AI can do on their own time and make more informed decisions about how their course should involve it for students.

2. How it spreads

Once AI solves a real problem for one teacher, word spreads. It moves sideways, one office at a time: someone fluent with the tool sits down with a colleague, solves a single real problem for them, and that colleague carries the news to the next person over chatter in the coffee lounge. The research supports this: studies of how computers first entered schools two decades ago found that the social capital between teachers rivaled any directive from the top in driving adoption,⁶ and more recent work on AI specifically finds faculty wincing at mandates, with most educators in one interview study fearing that a top-down requirement would erode their autonomy.⁷ Inoue makes the same case in his own terms. Writing about the disorderly state of AI policy across the California State system, he argues against settling the question from above and in favor of letting many approaches run side by side to see what works, what he calls letting “a thousand flowers bloom.”⁸ The practical reason is plain: a model released next year will outdate any policy written this year, so the thing that reliably changes how a teacher works is the colleague who sat down and showed them something useful.

3. Plain text is the durable bet

The final lesson is the most concrete, and Inoue’s own history makes the case for it. His early commitment to LaTeX, made long before any of this, is much of why the new tools paid off for him so fast: a course that already lives in clean, structured plain text is something an agent can read, rearrange, and rebuild, where a folder of scanned PDFs and Word files sits close to inert. It is not too late to make the same bet in a smaller key. Keeping a course in markdown and a sensible file structure is what lets an instructor revise it cheaply, hand pieces of it to a tool, and keep it accessible as the standards tighten. Plain text has quietly become the most durable and versatile material a teacher can build with, and those who organize their courses that way now will find that whatever tool comes next fits their work more easily than it fits anyone else’s.

And if your own course still lives in a folder of inert PDFs and Word files, that is not the dead end it looks like: turning those documents into clean markdown or LaTeX is exactly the sort of thing a Poorvu AI liaison can help you set up, and it doubles as a head start on meeting the accessibility standards—we would be happy to help!(mailto:askpoorvucenter@yale.edu)

About The Author

Krish Ramkumar is a rising junior studying Mathematics with the intensive certificate in Education Studies, focusing on questions of pedagogy in the AI landscape. Alongside his work as a Student AI Liaison to the Math department, he is spending this summer in the Netherlands researching AI-enhanced textbooks at Utrecht University through the Tetelman Fellowship. On campus, he serves as co-president of Yale Splash, sings with Yale’s Redhot and Blue, and leads hiking trips for incoming students as a FOOT leader. He is also an avid jazz listener. Reach out! krish.ramkumar@yale.edu.