Claude Code Series (part 10): Producing Highly Effective Decks for My Data Science Class
Including a 1 hour 40 minute video of me doing it!
Today’s post is in my series on using Claude Code for the type of research that applied social scientists find important, but this time it won’t be research. Rather it will be using Claude Code to make your classroom lecture slides. I will be illustrating it “live coding” so that you can see what’s possible. This lecture I’m focused on is day 2 in an undergraduate data science class at Harvard, in the government dept. They are all still very early in their quantitative careers, and so I wanted to focus on five things:
The rhetoric of quantitative social sciences.
Good empirical workflow (e.g., hierarchical folders, writing good code, automation, version control, replicability)
The video is long at 1 hour and 40 minutes, but it’s like watching me paint a scene because I really don’t take the position that the optimal way to do anything in Claude Code is just make some demand on the task and then wait for it to be completed. I would say I am just as intimately involved in the task as I would be without AI. You be the judge for how well I do it.
But the thing I try to emphasize in this video is that Claude Code is not just writing the decks for me using latex because it “knows how to code in latex”. I think that is actually the superficial understanding. It is not merely doing that. Rather, it has as I’ve said before absorbed the tacit knowledge involved in what great speakers who use decks to communicate know but have probably never written down anywhere, and may not even know how to explain it.
Dictating the Production of Your Lecture Slides
The way that I make my lecture slides using Claude Code is through exhaustive dictation. Talk, in other words; share your ideas, your objectives, the desired approach; explain your personal beliefs about pedagogy; elaborate on what your goals are, who your audience is; share the big picture outline as well as the tiniest minutiae, as it all appears.
I call this dictation, not vibe coding. You’re creating the lecture by talking into existence but you’re also modifying it constantly as you read through the appearance of each slide, as well as get a feel for the rhythm of the talk. So you will see me move things around, tweak things, try new things out, scrap old ideas entirely, or comment them out.
What’s a great about dictation approach to constructing decks is that you can also introduce concepts that you only vaguely have in mind. For instance, I throw out a request the Claude Code adopt an approach to what I’m suggest wherein the deck follows a rhythm such that the “marginal benefit to marginal cost ratio” across each slide is the same — a kind of “optimal rhetoric”. As an idea for what I want, that is both pretty accurate and pretty far fetched at the same. But in my mind, I think I see how the perfect deck does often follow that kind of smoothness. I just can’t say I possess the skill to do it, or the patience, but I suspect Claude Code can. And unsurprisingly, he does understand what I mean and can make an attempt at it in how each slide connects to the next.
Original theme and .sty file
One of the things you will see in this video is the way I asked it mid deck production to create a different beamer aesthetic design, and so it did — by literally creating from scratch its own .sty style file!
And so if you watch it, you’ll see the style change where I went from what felt like pretty boring grays and crimson colors to what I felt like was more color, and more interesting styles throughout.
The other thing that I ask for are ambitious drawings using tikz, the powerful graphics package in latex. Sometimes what Claude code can do feels like sorcery. Like the filing cabinet that he creates. Which I described anecdotally but did not explicitly request. I was speechless when I saw it, and especially when we perfected it together, and I’ve been doing this now for two months straight, since Thanksgiving in fact. And I was still flabbergasted when I saw it.
So as long as you are willing to remain immersed in the creative and technical process of producing the deck for your lecture, I think you will find it incredibly useful — a very real transforming of the productivity of your time, at least anyway for the modal academic. I think the modal academic finds deck creation very tedious and you know what my evidence for that is?
The fact that all academic beamer decks look more or less identical to one another, and that they are also usually bad, especially mine. Not everyone is born the Michelangelo of making decks.
A Life Lived in Decks
I think maybe 40% to 60% of my week involves being in a deck in some way, shape or form. Either I’m making one or I’m using one, but whichever it is, decks are a nontrivial portion of my life.
Which means I’m doing something with decks regularly. And they are therefore both important on the side of production and consumption. Getting them right is therefore important and also very time consuming. Put another way — time spent on decks has likely diminishing marginal returns, and rising marginal costs, and therefore we spend an amount of time on that where those two lines cross for each of us, and the quality of what we make is determined by that quality adjusted time use.
Well that is still true. But, I think what’s changed is that the curves have shifted. Specifically, the marginal cost curve has shifted left, and the marginal benefit curve has shifted right. And when that happens, the amount of time you spend on producing your deck is ambiguous — and you can see that played out in the video as I spend an hour forty minutes and I’m not yet done — but I would contend that the quality of what you make goes up.
I think the output so far has been fantastic. The decks have been fine tuned to me and my goals, tailor made to the experience of these specific students I have in mind. And you can just look for yourself here at my teaching tab at either Gov 2001 (the PhD probability class) or 51 (the undergraduate data science class) to see as I add the lectures in. And once I reach a completed deck, I’ll also add the .tex file for those of who want to see that too. Here for instance is the lecture I did on foundational concepts in probability for the PhD students.
Is Claude Code a Level Shift, a Slope Change, or Both?
Sounds ridiculous to say this, but Claude Code has probably changed my life. I was trying to explain to someone who uses ChatGPT intensively what it’s like. I basically said “you know how you felt like pre-ChatGPT you were doing things at x=100 but then after ChatGPT came out, you felt like you were doing things at x=1,000? This is basically like going from x=1,000 to x=1,000^2.
I know that it sounds like hyperbole to say that, but for me it isn’t hyperbole. It is all entirely caused by the fact that Claude Code is in the folder with access to the command line interface and therefore all these shell commands. And I also wanted to leave you with a few anecdotes too. First, here are two tweets (I got this from Ethan Mollick’s LinkedIn). The one on the left is by someone at OpenAI. The one on the right is by Boris Cherny, who created Claude Code. Look at what they both say when asked about how much of their code is now being produced by AI Agents. Both of them, computer scientists mind you, say that 100% of the coding is now done at OpenAI and Claude by AI agents. Not AI — not the copy-paste method of taking the output from ChatGPT and putting into a script. Not the autocorrect method from Github copilot. No — agents. Heck, Claude Code wrote 90% of its own code itself!
Then look at this tweet that Andrej Karpathy, founder of OpenAI, former director of AI at Tesla, said. Interestingly, me and him both started using Claude Code two months ago. I started I think the middle of November. I vaguely remember using it to do something, a week or so before Thanksgiving, which increased a lot over thanksgiving when I had a project I needed done, and then more after that, and then massively during the Christmas break, then a massive increase in January where I was spending 8-10 hours a day working feverishly prepping classes and finishing projects that had stalled, to now where it’s so deeply integrated into my way of working that I cannot even imagine a world where I could do without it.
The reality is the learning curve in using an AI Agent to help you with your classes, particularly writing decks, is not only flatter than people think it is — it’s easier. The learning curve may frankly even be negatively sloped for all anyone knows. It’s not just easy to learn; it isn’t even clear to me what you are needing to learn. There is no magic “prompt engineering” skill anymore. That day where you needed to have some special ability to talk is long gone, if it ever did exist in the first place — it was honestly always probably exaggerated by a bunch of snake oil salesmen online trying to make a dime out of a nickel as “AI Influencers”.
There is nothing to learn best I can tell. There is only needed a willingness to try and experiment. But that is how it frankly has always been with LLMs practical value to researchers and teachers. Claude Code is no different. It is akin to some kind of magical sandbox where you can basically dream up castles made of gold and rubies using sand, all using only your words. And its ability to do it? You give it the shell, and you give it your folders.
We Are Not All Naturals, or Absolute vs Comparative Advantage
My former colleague, Rebecca Thornton, is a genius at many things. She is hard to keep up with in basically any of them, let alone be better than her at any one of them. She is the classic example of someone who has the absolute advantage in just about every single part of the tasks of being a modern professor at a research university — great teacher, great researcher, and all the tasks involved in each therein. And I’ve listened to her describe the process and vision of her classes before — how much time she spends thinking about the underlying architecture of a course, what the training builds on top of itself, ways of eliciting student engagement, how every element reinforces every other element leading to human capital accumulation for the students and even herself over a 13 week period — the amount of time a semester lasts. It was beyond impressive to observe, and that isn’t even touching on her research acumen. She simply knows how to design a curriculum.
I am not Rebecca Thornton a am I close. She has the absolute advantage in, best I can tell, every single task that accumulates linearly and nonlinearly in the creation of university class. And that absolutely includes her slides. Which are themselves based on a rhetoric of her own that she just knows and knows well.
I frankly need Claude Code to just shorten the gap between me and the professors. I most likely cannot surpass them, but I can extend my own production possibility frontier such that the experience for students probably goes up dramatically. My skill as a professor is stymied by my own struggle with inattention, obsession with the wrong things, and misallocation of time and attention across the myriad tasks involved in producing a curriculum and class which includes but is not limited to the creation of decks.
I likely will never have the absolute advantage when put in a competition with the world’s best rhetorician, but I am not trying to either. I am trying to be my best self and Claude Code’s help in writing decks is just one of the ways I am trying to do that.
Optimal Production of Decks
I woke up early this morning thinking of the things I did in the video above last night, and specifically how I was trying to articulate this idea that there is an optimal deck, and that it is where the marginal benefit of the time spent making the deck (which is likely quality diminishing in effort and time use) equals the marginal cost. This is a deck I am working on which I’m going to share now. Here is the pdf (unfinished at the moment), here is the .tex and here is the .sty file which is the same as the one that Claude Code invented on the fly for me in the video above when I asked him to switch because I didn’t like the aesthetics of what we were using at that moment.
But, my contention has been that both curves shift with Claude Code specifically. Not LLMs generally — an AI Agent on par with Claude Code, specifically. One that lives in the directory. The effect that Claude Code has on the marginal benefit curve is that for every unit of time spent making decks, the quality of those decks rise and hence why the marginal benefit curve shifts up. My example is things like creating aesthetically pleasing themes (which you’ll see me do in the video), accessing expert level skill at producing high quality graphics with TikZ (a powerful yet utterly bewildering LaTeX package for most people who value their time at basically any non-zero value), and the ability to try out experimental ideas.
But the marginal cost of producing decks also likely declines. Here I represent that shift with a large decline, but also a flattening out, of the marginal cost curve. It’s unnecessary to state just what the new shape is, but my personal hunch is that it is like this. The point where we start to really hit the strain on the underlying cost function with AI is probably much farther to the right of even the hours in a day and as such my hunch is that the new marginal cost curve is flatter which means that at higher levels of time spent working, the savings gap between the old marginal cost and the marginal cost grows and may even grow substantially at high levels of use.
So that’s an interesting idea when you consider it. And this wraps it into a simple supply and demand framework that I personally think is so obvious to anyone using Claude Code to make decks that it’s not only noncontroversial to claim it, it’s borderline obvious. It’s obviously true to anyone who is using Claude Code to make their lecture decks that the time-adjusted quality of your lecture slides goes up.
Being a Professor is a Collection of Folders and Files
Claude Code in my opinion is for anyone whose job involves being in a directory of folders with files. That is who this is for. It is not just for writing R and Stata code. It may not even be primarily for that. It’s hard to explain, but it’s like you’re able to manifest your dreams through sheer wishes. That is probably a more accurate description of what I have been able to learn to do with Claude Code than if I was to even try to say more.
So, check out the video. It’s long, you can skip around. Share it. It’ll be free and not gated for another 2-3 days I think. Everything eventually goes behind the paywall, so I encourage you to watch it immediately if you’re not a paying subscriber. But I think seeing me make a deck for my actual class is worth it and that alone probably will cause some of you to become immediately a subscriber. And remember, I am paying the $200-250/month or whatever it is but that’s because I am a power user. I am using this so intensively all the time, as I have so many dormant projects on the verge of being orphaned that I am finally getting back up and running. Of course, supply creates its own demand so I’m also doing new things. But that’s okay. I need to and this has helped me to.
Thank you again for your support! Consider becoming a paying subscriber! It’s only $5/month or $50/year. You can also become a founding member which is $250 I think and that’s basically just you saying you want to do that as I don’t have any add-ons. But regardless of what you give, I can give you this.