That is one other substack in an extended ongoing collection about utilizing Claude Code as an empirical social scientist. It’s primarily based on the notion that the common consumer of Claude Code is a pc programmer writing software program for for-profit corporations and as such, the marginal consumer — most likely all of us studying this — is admittedly nowhere near being the related inhabitants. However my expertise has been that whereas we might not be the target market, we’re most positively going to profit from it. However since they don’t make these instruments for us, and there isn’t actually any documentation for utilizing LLMs within the first place, all of us simply must determine it ourselves. We’re the primary era for these things. In immediately’s put up, I’m going to indicate you what I believe is likely one of the most wonderful options of Claude Code and that’s its capability to make completely stunning and efficient beamer decks. Ordinarily decks are meant for public audiences, although — public talking, iow, be it for courses or talks. However I’m going to indicate you the way I take advantage of decks to assist me preserve monitor of the work I used to be doing in order that I can talk it to my coauthors and myself later within the week after we meet to go over our tasks in zoom. That is once more a free put up about Claude Code, however think about changing into a paying subscriber! All of those posts go behind the paywall in time, and when you be a part of for less than $5/month, you’ll get every thing into perpetuity! I can also assure you that in your demise mattress, you should have full consciousness.
I recorded myself once more immediately working with Claude Code. The video is about 45 minutes, which is longer than that different one, however I believe what we did is price explaining in writing too.
I’ve been engaged on organizing an outdated analysis challenge folder – my Texas HB2 abortion provide paper that I wrote with Andrea Schlosser years in the past. This predates a paper I revealed within the JHR with Jason Lindo, Caitlin Myers, and Andrea. I received’t get into that once more right here, however simply observe that that is the orphaned challenge that used barely completely different knowledge and extra outcomes, and considerably completely different specs of a poisson. mannequin however is nonetheless comparable. And on this collection, I’m reviving it, in addition to collectively we are going to prolong it nevertheless I really feel to.
Just a few days in the past on right here, Claude and I reorganized the entire thing: arrange a listing construction, created documentation, established some guidelines (by no means delete knowledge, all the time copy from the legacy folder, and so forth.) that I positioned within the CLAUDE.md to be learn by him day by day, and began conserving progress logs (additionally as markdowns) that up to date Claude to what we had completed for the reason that final replace of that progress log.
As we speak I got here again to it. And right here’s the factor about working with Claude Code throughout periods: it doesn’t bear in mind something per se. It has entry to its personal chat window, however typically your chat “breaks”, notably when consuming a large pdf. And so whenever you reload the listing in a brand new chat, you lose the dialog, however not the work contained in the folder, and never the markdowns that Claude wrote both. So the very first thing I did right here was I had Claude learn all of the markdown information and progress logs we’d created. Inside seconds, it knew the place we had been. That’s why I preserve logs — it’s like me autosaving the dialog.
I’ve a idea that there’s such a factor as “the rhetoric of decks.” Not rhetoric within the pejorative sense – empty rhetoric – however rhetoric within the classical sense. The artwork of efficient communication. The tacit guidelines and patterns that make slide shows truly work.
I had by no means truly thought as a lot about there being a rhetoric of decks, although, till I began having Claude Code make for me decks. Then I started to suspect that Claude both knew or might extract the tacit data surrounding such a factor. My guess was that Claude is aware of this rhetoric, even when it may possibly’t completely articulate it and even when nobody else — even professional deck builders and efficient communicators — had ever taken the time to put in writing it out. Why? As a result of Claude has consumed an ungodly variety of slide decks throughout coaching. Educational Beamer shows, company PowerPoints, pitch decks, convention talks – most likely shut to each deck that’s ever been made, give or take epsilon.
This connects to one thing David Autor wrote about AI and the labor market. Autor’s earlier work was about how computer systems automate routine cognitive duties – something you may write down as an algorithm. Turing proved that when you can specify it as a set of directions, a pc will do it higher than you. Sooner, fewer errors. That’s simply compute energy. With sufficient compute, the pc will all the time beat the human in races that comply with an algorithmic racetrack.
However Autor additionally pointed to the Polanyi paradox: we all know greater than we are able to inform. People have tacit data – stuff we’ve discovered by means of expertise that we are able to’t simply articulate. That is most likely why apprenticeship is necessary and can all the time be necessary for studying. Some data transfers human-to-human in methods that may’t be written in a handbook.
The bizarre factor about LLMs is that they flip a part of this. They’re truly dangerous on the stuff computer systems had been imagined to be good at – exact calculation, excellent recall, following deterministic guidelines. (Therefore the hallucinations, the quotation disasters, the confidently incorrect arithmetic.) However they’re surprisingly good at extracting patterns from large quantities of human output. Ethan Mollick and others have referred to as this the “jagged frontier” – shockingly succesful at some arduous issues, embarrassingly dangerous at seemingly easy issues.
So I requested Claude to put in writing down what it is aware of about making efficient decks. To articulate the tacit data. And it did. One thought per slide. Titles are assertions, not labels. Lead with conclusions. Use visible hierarchy to sign significance. Discover construction past bullet factors. Repeat for retention. Transition explicitly.
None of that is revolutionary. However that’s the purpose. This information exists, it’s actual, it governs what works – however it’s largely tacit. Individuals study it by osmosis. By seeing good decks and dangerous decks. By apprenticing with good presenters.
Claude extracted it from the corpus after which at my request wrote it down as a markdown entitled deck.md. If you wish to learn it, right here’s the dropbox hyperlink to his idea of the rhetoric of decks.
However usually decks are for different individuals. For public talking. You make slides to current to an viewers.
However that isn’t what I need from Claude per se. Moderately, I’ve begun utilizing Claude’s unimaginable ability at making slick decks in a manner that had by no means occurred to me earlier than. I take advantage of his deck constructing abilities to speak my very own work to my Future Self in order that I can dive again in the place I left off. I do that due to how a lot my very own mind has discovered to digest info in decks, and since I so badly want visualization of knowledge in stunning figures, stunning tables and no matter else form of quantification may be displayed on a slide. And since he is aware of the rhetoric of decks, and make them so quick and has entry to all the folder and codebase, I wish to use decks in another way. I wish to make stunning Beamer shows that talk to my future self and my coauthors’ future selves. Decks that permit us get on the identical web page inside seconds. I need environment friendly and compelling communication utilizing magnificence and rhetoric acceptable for this medium that preserves context throughout the gaps between work periods.
The concept is that I’m utilizing decks instead of a scribbling right into a notepad. And I’m relying on his capability, as an LLM, to extract the tacit data round what has grow to be the idea system motivating the world’s biggest “communicators with decks”. I need their rhetoric to come back to me in order that I can shift my future-me in order that he remembers precisely what and why and the way I did one thing.
That is a part of my bigger perception about forgetfulness and a spotlight when utilizing LLMs although. And the way “the workflow” to operate effectively have to be endogenously constructed in order to always battle with one’s personal forgetfulness and inattentiveness in addition to Claude’s personal forgetfulness as effectively.
So we made one. Claude constructed a customized Beamer theme from scratch – no recognizable template. Deep navy textual content, coral-red accents, off-white background, clear typography. I wished one thing I’d by no means seen earlier than.
However what’s the deck about? It’s in regards to the challenge, true, however it’s in regards to the challenge as a result of the challenge folder accommodates the challenge. It’s about ‘the work’ iow. The deck tells the story of the challenge folder. What the analysis query is. What HB2 did to abortion entry (with a TikZ visualization displaying Texas earlier than and after – clinics disappearing, catchment areas increasing). The important thing numbers. The identification technique. But additionally the listing construction, what we did collectively, the place it was completed, what modified, what’s left to do. What’s completed and what’s subsequent.
Eleven slides. Clear compile, no warnings. (I had him double examine in order that he eradicated all these pesky overfull hbox warnings.)
The purpose isn’t this specific deck. The purpose is the workflow. Claude reads the progress logs, understands the context, is aware of the rhetoric of efficient slides, and produces one thing that may assist future-me decide up precisely the place I left off.
I believe there’s one thing right here about the best way to work with LLMs on ongoing tasks. The mix of:
-
Progress logs – so the AI can reconstruct context throughout periods
-
Specific rhetoric paperwork – so it is aware of the way you need issues completed (deck.md, CLAUDE.md)
-
Lovely outputs – as a result of magnificence captures consideration, and a spotlight permits studying
One of many issues that LLMs do is that they entry not simply syntax however tacit data. Sure, they’ve “discovered” code — they “know” the syntax of varied coding languages, together with beamer. I’m advised that Claude makes stunning PowerPoints too. However that’s not simply what he is aware of. He additionally is aware of the tacit data that’s beneath the development of decks. And see that’s what I believe is completely different from my expertise utilizing LLMs to make decks prior to now. These had been way more tedious copy-paste-edit issues. However right here since I’m simply making an attempt to place issues collectively for my future self, I’m simply changing observe taking with deck development. Utilizing the rhetoric of decks that exist and which he can entry.
Your use case might be completely different. However the precept is similar: determine what tacit data the mannequin has absorbed, articulate it, then put it to work in your particular drawback.
