7 Mistakes Economists Make With Claude Code & Codex Skills (USC Seminar)

0:00our last uh Econ on the AI seminar of this semester. Um very happy to have uh Aniket Pendharkar here uh today. Who's kind of made a name for himself as the AI economist. Mhm. And um um you originally uh were going to talk about verification, but he now switched to skills, which I think is even more exciting. I mean, we all need to verify, but first we got some skills to So, [laughter] I'm looking forward to your your

0:29presentation. Very good. No, thank you David. Thank you all for having me and for all your hospitality today. It's been a wonderful to be here. So, I'll be talking about skills. So, I think that we're approximately, I would say, a year in now into the AI revolution in economics. And so, for me, it's kind of began as a storm in last winter when everybody during their holiday break seems to have decided that now was the time to try out Claude Code, and many people's workflows began to

1:07change right about that time. And so, I think that at this point, most economists have tried Claude Code or Codex. It may not be a daily part of what they do, but they have tried it. And so, I do still see, however, economists making various classes of common mistakes with these tools. And so, what I want to focus on today is specifically on mistakes that I see

1:35economists make using skills. So, just by way of brief intro about me, for those of you don't know me, I'm an economist by training. I did my PhD from Northwestern, but then I worked as a data scientist and machine learning engineer in fintech for several years. And right around in the middle of 2024, while at my full-time job, I was moonlighting for a startup where the CTO was like, "You should try out this thing

2:03Cursor." And so, what's Cursor? And I immediately got hooked, and I sort of saw that this was going to be the future. And I started using these tools daily right around then. And so, in 2025, I was still wrapping up some research projects. And I was doing a little bit of pair researching with friends, where I would primarily be driving the coding UI, and my friend would be stunned. And then, I sort of realized that, you know, this probably is of value to economists. And so, I called up my friends, and I was like, "You really have to get on this thing Cursor, or then Claude Code." They

2:42said, "Don't bother me. I'm trying to get tenure." And I, you know, messaged them again, and I said, "You really need to get on this thing called Claude Code." "Fine. Holy crap." And so, that's kind of been my experience. It just snowballed from there. And about now, I've worked with maybe 500 economists across one-on-one and group workshops. And so, what I hope to give you today is a little bit of what I've learned from

3:05working with all these people. So, I want to start by defining what a skill is, because that may not be something you're aware of. So, we'll start with an intuitive definition. And maybe in like the early middle of this talk, we'll get to something of a more formal definition. And one thing I'd want to say up front is like, I don't know the culture here is in the seminars to just ask lots of questions. Ask them

3:29right away. Ask them throughout, please. And so, broadly, a skill is something which extends the capabilities of an agentic coding tool. It's a set of instructions, and that's what it does. It tells the tool, "When you encounter this situation, behave in this way." And so, a simple mental model of a skill is is it's a clipboard of useful instructions that you can easily pull into the prompt that

3:55you're giving the Agentic Coding Tool. Very good. Okay. Now, the fastest way, I think, to understand skills is to start by just looking at some concrete examples. And so, that's where I'm going to start. So, this here is a skill that I use a lot. I call it Fathom Research. What does that mean? Fathom is the tool that I use for transcribing many of my meetings. And so, if I'm having a meeting with a co-author, or meeting with a client, I have this transcriber just in the background. You've seen these AI note takers. But then, what I do, just a little bit different from what other people do, is I sort of close

4:33the loop with these AI note takers. I make sure that the notes get into some centralized directory very regularly, and then I have them available to me to query upon whenever I need them. The benefit of that is that if there is like some meeting that I had with a co-author 2 months ago, and I feel like it could be relevant to something that we're discussing today, that's available to me through the Agentic Coding Tool to find that context and help me, uh you know, act on whatever it was that we

5:05were discussing quite a while back. Okay. So, that's one example of a skill that I use quite often. Now, one thing you'll see here is that if I sort of scroll through here, uh this skill's pretty long. It's got like a lot of technical nonsense in here going on. And then, if I go to this one, on the other hand, is very short. So, this one seems to have like all sorts of code and references in here. This one here is just a small set of uh uh like maybe three or four sentences in total. And so, what is it? This is a grilling skill. And so, I use this skill

5:43quite often, as well. What it allows you to do is to enter into a sort of I want to say adversarial, but maybe a relentless Q&A session with the AI. Now, the situations in which I tend to use this skill is when I'm working on something and I've created some document with the AI, but perhaps I feel that I don't have a complete understanding of this document, at least to the level that I'm comfortable proceeding on with without achieving some higher level of understanding than where I am at right

6:19now. This sort of simple set of instructions, I just type in dollar sign grilling, then initiates a Q&A session with the Agentic Coding tool, which I find quite useful. Now, it's possible that some of you have had the idea to do something like this in situations where you feel that you don't understand what's going on exactly with what the AI is giving

6:43you. But, it's possible also that you, if you just had the ability to initiate the session with dollar sign grilling, would do it much more than you otherwise do. And because of that, you would develop a higher level of understanding through use of these tools than you do right now. And so, that's a big part of the

7:05advantage of what skills can get you. Yes, David. I mean, you've kind of jumped into skills as one of the important aspect of Agentic AI, but where do you kind of draw the boundary where like, you know, there's obviously a lot of capabilities that come straight out of the box Yeah. of an Agentic AI tool. Like, where do you feel the need to go from just vibing your needs. Uh, I mean, you could write really me about this, and it will it will do something like this. It won't do it in such a specific way like you described it here. Uh, but it will do

7:38something like that. Yeah. But, then at some point you decide, okay, you know, this needs to be a skill. So, where do you draw the line between, you know, the capabilities come out of the box versus things that you feel like you need to add on top to really customize it for your for your

7:50benefit? Great question. Uh something I'll be getting to in a few slides, but I'll answer it now. The sort of threshold or rule that I use is that if I've noticed this as being something that I've repeated three times, then I make the skill. In general, I find most economists have talked to suffer from making or using insufficient skills, and like my personal experience with the tools is greatly improved by having little sort of nudges like these that are customized to me, tweaked to me over

8:27time. Yes. I guess a related question. I'm so I I did not even know it was called a skiller, but when I was a I I typically prepare, I spend time preparing a document and kind of the structure of what this whole project's going to be about with the rules and decision process rules. Uh you seem to suggest that you want to have many such smaller chunks of

8:51instructions. Uh you can turn on and off. Yeah. I I I think that that is not a skill. It's probably not a skill. That might just be a document which describes everything that you are intending to do in the project. I typically put it at agents.md Okay. Yeah, yeah, very good. So, the So, that's not a skiller. No, the the agents.md is highly related to a skill uh in that they both are mechanisms by which to manipulate what is in the context of the AI. With the agents.md, that enters the context of the AI always whenever you're in this project. With these skills, it's more elective. So, you wouldn't necessarily

9:36want to enter a grilling session every time, but maybe you would always want the AI to know these details of your project irrespective of what you're doing in the folder. So, that's a distinction. Good. Okay, this is from Pedro Santana, who's an econometrician at Emory. And so, he's put together a skill, which you can just look here, describes, you know, his standards of how to do a staggered difference in differences or event study according to the way that he likes doing things in the programming languages that he

10:16prefers with the packages he prefers. And so, relative to the other ones, this one is quite a bit more complex and has, you know, it's like it's huge. I haven't read this thing, maybe some of you have, but it describes how he likes to do difference in differences. So, just getting a sense of the types of things that could potentially be encoded

10:39in a skill. Yes. So, so my question is all these kind of skills uh they are located like in detail or something? They'll be located on your computer. So, and I'll talk about like where they come from, how you make them, making ones that other people have made versus making your own very shortly.

11:03Yeah. Yes. a follow-up question. But they're about like, do you use AI to actually make like Are the authors actually typing all of this or is it like, "Hey, I want to do this. Write me a skill to do it." Also something we'll be, yeah, covering throughout. Yeah, but it's it's good. I'm prepared, you know, so I got got

11:18your questions coming ahead. Um this is a skill to control your Chrome browser. So, one thing you all may not be aware of is that AI tools, Codex and Cloud Code, can pretty easily be used to manipulate your Chrome browser, So, manipulate your computer. If you want to like edit a cartoon video of your job market paper, that is something you can do now just by having

11:43Codex edit the video for you. And this is a skill that I use sometimes to write like myself. And so it applies my writing voice to different types of writing. I'm going to discuss and I'm happy to discuss the degree to which like I found this useful. Um I would guess the sort of summary is that I find it most useful for emails

12:06and like that's the main thing. Sometimes for a single slide and I don't really have success using this for longer um longer pieces of writing. And I can uh discuss with some of you today how I approach AI and whether to use AI in longer pieces of writing, but um that's something that we can get to

12:27during the during talk. And this is uh from Chris Blattman who's a development economist at the University of Chicago. So he describes um back in February, you know, early in the revolution, "Second best project management skill he has created. Every week for each of about these 10 research projects, a skill pulls and integrates the transcripts of all meetings, all project-related emails, all discussions in the project WhatsApp groups, and then it generates a project overview covering everything he does. So he has a memory of every major discussion discussion decision and discussion. And so like for context, if you're sort of thinking about Chris's production function relative to your own, he's probably

13:10managing like a lot more projects. He's doing so at a higher level relative to being in the code, actually coding all the time. But he nevertheless has found immense value in these tools and making his own skills and he was a pretty early adopter. And so, one of the, let's say, intuitions or framings I want to make is that if you're not using these skills or you think that it's something that's maybe only appropriate for those who are really writing code in the weeds, that's uh I think really not going to be the

13:43case. Yes. You You probably get into this later, but that deep diff-in-diff example pages and pages of these reminders, it seems like it's faster just going on Stata and just doing it. Or I think that what that skill is doing is like establishing how he wants a diff-in-diff to be done when an agent is encountering some situation in which he's told with this data, "Write me the code for a diff-in-diff." So, it I mean, conditional on that being I I guess the the degree to which it's costly is the degree to which he individually had to write out that

14:24skill. Like he he didn't have to do that. Yeah, so I started thinking I mean something that would be very useful could be diff-in-diff stata XYZ, and then AI does it does everything. Ah, so like Or write down the model and it's in a sloppy way. So, like maybe like make like more of an interface that instead of having all these lines of everything that like put this argument

14:54in. already, it's fine, but if if you're just writing it for yourself you don't need those steps. I think that's correct. Yeah. Or you can make a skill for that unified interface how you prefer it. There's one more. Yeah, your face, yeah. It's useful one. I was like I tried to install but the skills that I found online like couple months ago, and I feel like I just kind of forgot that I have them. Like I would remember like the skill of writing and I would remember like sitting there with specific skills, but most of them I feel like I got no use just because when I'm asking uh

15:29you know, I'm writing a problem that I don't know oh, which of the 100 skills that I have would best fit this problem? And I sometimes see that you know, he he will call the skill uh some at some and don't really understand when he is I will call the skill or not. Uh without me prompting it, but how do you go about this? Like how do you get the best out of these skills without forgetting the ones that you actually

15:51have? Yeah. So, you can make skills that are specific to a project or ones that like are across all projects. And as I'm going to show, um if you have skills that are just hanging by the wayside and you forget about them, they don't exert a huge cost to you in terms of like your uh like a diminishing of your experience with these tools. So, if you have 100, ends up you only use 10 of them, that's

16:20fine. You can just use the 10 of them. Those 90 do actually exert some cost to you. I'll explain how that is, but it's not that big of a cost. Yeah, I'm I'm more worried that I'm under using my this good skill. Oh, okay. Um yeah. Uh I will show you in probably I don't know, 15 minutes or so how I

16:44would address that exactly. Yeah. But there's there's like a way to set up basically say, "Okay, from now on uh into the future whenever I want a different diff, I want it to reduce the key ID skill, right?" Correct. So, I I I I'll I'll sort of I'll explain exactly how it occurs in terms of like what you are saying or typing to that message coming through, probably in a few

17:11slides. Yeah. Okay. Cool. Uh great Very good. So, I think that this what I wanted to establish here initially is that skills seem potentially quite useful. And what is it that economists do wrong when using skills? So, the first mistake I see people make is that they do not make their own skills. Now, a pretty common question I'll get when I present skills in some workshop I'm giving is where can I find the skills that are useful for me or how can I get that DID skill from you? And making skills is actually remarkably easy at a first pass. You know, you you can borrow skills and there's probably

18:01some skills I'll mention here that are useful to you and I'll show you how you can go and get them. But economists workflows are very idiosyncratic and so a place to start might be to look or be inspired by some skills that somebody is making when they sort of explain how it fits their workflow and then maybe it's sort of, you know, it's not quite the

18:23way you do things. It's very easy to make your own. Let me show you how. There's a skill for making skills. This is how you make skills. And so in Codex if you just type in dollar sign skill creator, there's going to be a skill and then you describe to that skill what skill you want and it'll help you make a

18:44skill. In Claude code, there's a skill creator plugin and so you instead type in backslash skill creator and then you invoke its skill creator skill and it'll help you make the skills. I'm going to get all these slides for you but yeah, take all the pictures you want. Um Yeah, so the the bottom line is that you don't need to know anything about skills to make skills. At this point like hopefully you know, and decide this guy is full of I don't want to listen to him anymore, that's fine, you can leave. And like, you should be able to work with skills without really having anything more of a than a vague

19:18understanding of what they are. And so, that's one point that I wanted to make. So, out of curiosity, like a raise of hands, who here has made a skill before? Okay, so it's like 50% about, maybe 60%. Good. All right. So, try making a skill. And let's say, suppose you were working with code extra cloud code in a way that people typically do it interactively, just like in one long thread, and you suspect that some of what you did is

19:50something you'd like to go do again. You know, you can ask, you can actually ask the agent, is this currently an appropriate place or reason to be making a skill? And I'll explain, you know, situations that confuse people where they think they ought to make a skill and it's actually, let's say, not a

20:07skill-shaped problem. And then you can ask, just if so, make one for me or suggest the one to make. And so, I think this is like a perfectly reasonable prompt to to have in in some of your coding agent sessions when you're starting out. So, this first, if you haven't made a skill, I would, you know, push you to to try it out and

20:27see how it works. Now, the second mistake I see people making, okay, so we guess we talked about the uh uh uh intensive margin, now the extensive margin, uh making too few skills. So, I have hundreds of skills. And some of these are very project-specific, and then some are user-specific, meaning they're things that I want to be available across all

20:53my projects. So, the heuristic that I suggest is that when you notice yourself doing broadly the same thing on three different occasions, time to consider making a skill. Now, maybe that's too aggressive a heuristic, but given that I tend to see people, you know, with like a set of three skills that are made, generally I would recommend you try following that heuristic for, let's say,

21:16a few days, and then see what happens. [snorts] Okay, here's an example of a skill which I find useful, which I have made using this heuristic. So, yt-dlp skill. So, I often come across interesting long YouTube interviews, okay? Like I maybe hey, AI space, maybe on econ, maybe on poker. I I like poker stuff. And so, I

21:40rarely have time to watch these, okay? So, the solution to this is a tool called yt-dlp. What is yt-dlp? It's a free, open-source command-line tool for downloading video, audio, and subtitles from YouTube, and also other sites. Incidentally, if you guys are like thinking about doing research in economics or the media, something like that, great data source. You can, you know, scrape YouTube at scale using this

22:06tool. So, for this workflow, uh what I'll often do, or I've noticed myself doing, is that I would have this like Dwarkesh Patel podcast on something related to AI, I'd want to understand a bit of it, but I don't have time to watch it. So, I'd ask like coding I'd ask Codex to have like with yt-dlp download the captions, and then summarize what is

22:30important to me. And so, it's very easy to install. You have tool brew, you just install it like that, or you can just ask your coding agent to install it on your computer for you. And so, what I made was a yt-dlp skill. And like before, you know, I'd find myself typing in something like this, and maybe I would notice that it would enter in the arguments wrong, so I'd have to manually tell it not to do something that it would tend to do, and

22:57to create a summary for me. And then afterwards, I just type in yt-dlp and paste in the URL. So, this is like a small thing, but it sort of turns into a little bit of muscle memory that you have. You know you've made this for yourself before, and it becomes a more pleasant tool to use. So, this isn't a a hugely significant skill like I don't know some of the other ones, and I'll show you some ones that I consider absolutely critical, but as an example of that thought process I had. I noticed this thing being something I'd want to do three times. Okay, just make a quick

23:30skill. It's very quick and easy to make a skill. Hey, it's Aniket. If you're enjoying this talk, I wanted to take a minute to explain some of the ways that we can work together. First, I can conduct agentic AI trainings to make your researchers or employees more productive. My trainings are not generic. I'll work closely with you to adapt the training to participants' particular needs and incoming levels of familiarity with agentic AI. So, for example, even though let's say researchers at a hedge fund and a group of applied microeconomists may make use of the same tools, I'd tailor a workshop quite differently for each group. The

24:09examples and exercises would be built around the work the group actually does. Second, I'm also able to help your organization directly with AI implementation work. I have several years of experience as a machine learning engineer, and I was previously director of AI and ML at a fintech startup. You can see all these options

24:27at ai-economist.io. To book a call with me, go to ai-economist.io/call or email me at aniket@ai-economist.io. Okay. Now, we've talked about making too few skills. Now, another mistake that I see people making is making unnecessary skills. And so, this is the question of understanding like what exact problems require skills versus not. So, skills in general are for tasks which repeat. So, scraping YouTube transcripts and like getting a summary out of it, that should be a skill. Cleaning the data set for a paper should not be a skill. This is a set of code that you are making one time for a particular paper. It's not

25:11something I would make into a skill. However, reviewing code for cleaning data sets to create a report on all substantial decisions made in creating a sample, that could very well be a skill. You may have many different research projects on which the coding agent is making its own idiosyncratic scripts to scrape data sets, merge them together, clean them. And maybe you want to have a standardized way by which another agent evaluates the code written by the coding agent and tells you like what are the decisions that you should look out for, if you should look at the code, where should you look for these little details

25:47that matter to you. So, that's a little bit of the distinction here on what should be made a skill. So, let me give you like a more concrete example. Like suppose you're studying the cereal industry, okay? If you want to collect cereal prices across markets, you just have Claude code or Codex write that code. So, it'd write some code, scrape.py, collect the data. There's no

26:11skill required here. Uh if you're uh extracting manufacturer information from tens of thousands of PDFs that you have, I still wouldn't make a skill here. So, what I would do here is I would have Claude code or Codex write code to extract the text, okay? And then I would have the script call a model, a large language model with this extracted text, and ask it to just from this extracted text parse out the structured data that you need. So, this still here is not really a skill-shaped problem because the call that you're doing here with your open router model is pretty straightforward. It's like just give me

26:55a list of prices back. Now, let's say you have, you know, 120 IO papers that are potentially relevant to your project, and you want to have like a first pass thing for you to skim on how these papers might be relevant to whatever it is that you're doing. This is actually a skill-shaped problem. The way that I would approach this is I would make a skill which consists of my assessment criteria for the way in which um I like want the AI to assess whether a paper is relevant to my project. Then, I ask it using a concept common in agentic coding called sub-agents. It's a method

27:41of parallelization to basically apply that skill in parallel across each paper, and maybe give me back some kind of written output on a per paper basis. So, in this situation, all kind of in the same space of general thing we're doing, studying the cereal industry, here I would want to have a skill. So, that that makes sense? Yes. I guess that we're doing one or two like when Qualidor Kodak say writes the code, do you have skills about

28:12writing the code itself? Yeah. Uh uh I'll get into that, too. Yeah. I had the same question. Man, I follow you these body of dying. Yeah. That kind of skill. Yeah. Yeah. Yeah. So, I mean, like you My my answer is going to be like the compound engineering set of skills. Yeah. Yeah. Very good. Sir. Um

28:34Yes. So, with each of these, one, two, and three, you just you have to come up with a set of prompts that would give you the the right like the collective collective each of our sister cross markets, and you have to come up with some prompts, right? Yeah. Yeah. So, you're saying like this is all going

28:53to come from? That's not the easiest thing in the world. Yeah, I think that like So, a uh a skill that could exist is that um it could be like a skill that help you understand if a skill is useful here. All right. Like So, I mean You you you're saying it's not the

29:17easiest thing in the world. For me for me I I I've done it and I've seen these apps as a RAs, you know, and it's it's fine, but if it's like a multi-step thing, okay, collect data, um open up map, open up this data, analyze the data this way, and then put the result in Yes, that that model that exists. So, that's like a kind Yeah, the realistic Yeah, realistic stream of things you

29:44want to do. So, that that would be kind of skill-shaped because there's a particular way in which you're describing doing things, and then these are your standard operating procedures. Yeah. And then that's where the skill. Like here, this is um like just a piece of code, and then maybe as like these fellows were asking about like if there's a particular way in which you always want Python files to be written,

30:10Yes. then like a skill could be useful for being instantiated every time you try to like have the agent coding tool write code at all. And I think But this thing in the middle is not actually going to be a skill. Whereas over here, this thing in the middle is a skill. Do you want to

30:36Cool.

30:41And so, a partial answer to like situations where you're unsure is again, you just ask Claude Code or Codex, okay? I'm interested in making a skill for doing X. Do you think it makes sense for me to do so? And um the reason I say it's a partial answer is that in like more complex situations, it may not get you all of the way. And I'm going to be like discussing a little bit more about that later in the talk. Like the degree to which you can just depend on Claude Code or Codex to help you without like any external reference sort of make these architectural decisions versus it

31:26being a thing that you have to get a sense for through practice. Okay. Now, another mistake that I see Google making is not using their session transcripts. So, all of your past sessions, your coding agent sessions, are in and of themselves a good source for new skills. So, coding agents, Claude Code and Codex, save transcripts of your local

31:55sessions on your computer by default. And so, this is very personal, but also very useful information. The default locations are inside your home folder. And so, there's different locations for Claude Code and Codex. If you like go in that directory and look, you'll see time-stamped files, which are just the transcript of like all the files being read, all the messages back and forth. Um with Claude Code, I think they now get retired on an every 30-day basis. So, like Claude Code will just delete the old ones. I think the Codex

32:27ones stay around for longer. So, this is this would be the same for whether you use Claude through the command line or through the previous chat of who asked who the previous then I'm not 100% sure. I think for the desktop app and the command line interface, yes, for sure. VS Code might be different.

32:53Yeah. Yeah, you can have code called this thing you did that you were talking about that right? By by what? Like your heuristics so about the conversation like the little Oh, yeah, yeah. Yeah, yeah, yeah. Yeah, exactly, exactly. So that that that is the advantage is that you can have Claude Code and CodeX like do a search through your previous sessions either in a project or across projects and you know, assess if there are repeated workflows here that would make good skills. And so so it's a very handy thing to do. It's also like a handy way sometimes to like remember

33:38when you were thinking about something. If you know to if you know these locations exist even if you don't know the exact locations, you can simply tell Claude Code, "Hey look, like 2 weeks ago we were talking about this, but I forget what the exact data set was, where it was located. Can you look through my coding agent sessions and try to find it?" And it should be able to find it

34:00like that. Yes. Is this just for like Claude Code or like Claude chat? So because at least like when I I think I've tried doing like, "Hey, write me a transcript of this." I don't know if it's chat or Claude Code, but like one of them just would not give me the transcript. Yeah, I'm it's definitely available with Claude Code. I'm not 100% sure with

34:19Claude chat. Like I just recommend you don't use the chat UI. Like like if you use the desktop application, there's like a chat bubble and then a code one, just always use code. Yeah. Even if you use it like chat, just don't use Yeah, you can just still use it. Yeah, cuz I I mean like for one it's certain that it will accumulate like this and like also I just see a lot of people who are in the habit of being in the

34:42chat interface. I don't have like a complete description right now of the things that are like bad in it but part of it is that it's like a more bounded interface. Like it's like you get sandboxed and so better to just use the code one. Yeah. Cool. [snorts] All right. So a skill which I have developed which I use quite a lot is kind of as I described these cloud code sessions get um

35:14deleted after 30 days. Um and so I've made a skill myself to kind of get all of these sessions into a single database on a regular basis and then be able to search across them in like an optimized way. Um not something that you need to do but I just wanted to bring it up as like a kind of meta skill that I find extremely useful and use

35:41quite often. Okay. So from here I think this is like a good point to describe the technical definition of a skill where these actually live on a computer and so on and so forth. And so a skill is a folder which has certain properties. Here's what it looks like. So this is this difference in differences event

36:06study skill. Okay. And one of the first conditions for this skill is that it has to have the front of the skill like be in a particular way. It really needs to have these three dashes or it won't show up as a skill to cloud coder code X. Needs to have a name field and a description field. These two are

36:29important. Then, it's going to have a body, which describes the contents of the skill, and I've shown you a little bit about what that looks like earlier. And then, everything else is optional. So, these things are required, the rest is optional. But, there could optionally be, let's say, scripts, references, images, data, templates, all which are captured into this skill folder. So, maybe that contains, you know, an

37:00interface that the skill can reference. Maybe it contains references of like a checklist to go through when doing an event study. It could contain all sorts of things. Now, one of the fundamental features of skills is a concept of progressive disclosure. Let me explain what this means. It is kind of core to like what

37:27these agentic coding tools are. And so, something which you may be aware of is this concept of a context window. So, when you're using Claude Code or Codex, you can't just keep cramming in your questions infinitely forever, so there's no end. Eventually, you hit a limit. This limit varies depending on the agentic coding tool and model. And then, once that limit is reached, an event will occur called compaction, which summarizes everything that's been done to that point, and then continues on with the summary going forward. So, a little bit concretely, Okay, so that that's how it works. And so, when you start a coding agent session at a baseline,

38:15some of this context window, even if you haven't sent a single message, is already going to be filled. What is it going to be filled with? Well, one of the things that's going to be inside there is the system prompt. And so, you may have heard of this with the Claude, for example, is like, you are a helpful assistant, you know, you do not try to kill humans, things like

38:36that. That's all in the system prompt. And there's also access to tools, which are the actions that the coding agent has possible. And so, these may include a tool to read files, to write files, to do browse the internet. These all take up space in like say the first 30,000 tokens that are here. Now, what skills are effectively is references to tools. That's basically what a skill actually is. This is a

39:06reference to a tool. And so, you see here that also, before we started sending any messages in this context window, there are eight skill descriptions that have filled up a little slit, just 800 tokens of the 258,000 that get filled up. And what is being filled up specifically is everything here. So, what gets filled up is the titles of the skills and the descriptions of the skills, which I'm just for let's say convenience saying let's say if they're about 100 tokens each. So, another 800 tokens gets filled

39:45up. Okay. Now, let's say the user sends a message, can you update this figure workflow? Okay. And so, in this instance, I haven't explicitly specified like the particular skill I'm referencing. So, you'll see if you look through here, the most relevant skill here looks like the speaker update skill. Nevertheless, like these agents are intelligent and they're able to use that intelligence to sort of even if I didn't say it's bigger and I didn't say update, but I was like semantically in the direction of a figure update, it's very likely that the

40:27following thing would occur. Or you can explicitly invoke it. In either case, the skill would be that if read by the age is decided that given the message of the user has sent, the skill makes sense to read and so balance body of the skill is read. This skill.md file and maybe that takes up another

40:475,000 tokens. And it perhaps decides that reading some of the references is useful like in the scripts file, but maybe not like the checklist or something. And so the other stuff that gets read is optional. This definitely gets read. And so that's what's occurring with skills, but importantly as opposed to the agent.md or the claw.md, which is more like the system file, it's just always there. With skills, only the name and the description are being read every time and then everything else is read into context progressively. And so that's

41:29sort of the core of what skills are. It's a way of managing this context window which is precious to us while giving us the ability to selectively do lots of different things as we need them. Yes. What you just said like I understand that there are two costs in like creating new skills that might not be

41:53happening at the right. What is like the token concept? Great. Yeah. So what if if you had 100 say skills, then this would take up 10,000 tokens, which is something, but it's not the worst thing in the world. Okay, I agree with you. The other more subtle one though is that the most agents are intelligent that can like then understand that you are to use one of your skills from like just

42:20language. Like it's you have many skills like they can use one that you Yep, that's correct. So like it's possible that I have let's see I have whatever DID advanced study and like DID analyze or something. Okay, you could imagine like I I've forgotten which skills I've been making and so I have these redundant skills in there. And so sometimes like the wrong skill gets

42:55triggered. So you basically have to be attentive to this if you're in some big set of skills. There is I think somewhere I have to remember to set skills in the front matter such that they can only be invoked explicitly. Like I think in both Codex and Cloud Code there's some way to modify the name and the description so

43:17they can only be invoked in that way. And that solved that problem and some people believe that like the best way to use these agents is to only have explicit invocation. And personally like I like having it because like I I sometimes I'm working like from my phone just like go on a walk, say something and I want that to be able to

43:35you know go through freely. Yeah.

43:44Okay. Now another mistake this is I would say a lesser mistake but still a mistake is to not clean up your skills. So skills can go stale. How does that happen? Well, your project might change. So if it's a project specific skill, some files or folders previously referenced name no longer be there. And so you'll reference a skill, and you'll probably notice in the transcript, "Oh, I looked for this file, but it's not there." You

44:12may wonder, "Where is this coming from?" It's very likely, or it could be possibly a stale instruction in some skill that you and they wrote a while ago. Your computer system might change. You upgrade packages, and an old skill is referencing some other package. External systems might change. You're trying to hit some API, which itself has

44:33changed. And the most important, and the most common, is that the model might change. Or the model will absolutely change as as we've all seen. It's a good example of this is with Stata. A year ago, coding agents really struggled with Stata. And so, I made a video, a very popular among economists on YouTube, on how to use Quarto with

44:57Stata. And these days, that video is basically irrelevant because explicit instructions, kind of like all I remember describing in that video, needing to explain what Stata is and how to use it, are unnecessary. Or are they? Stata skills may still be useful for non-frontier models. So, I have a grant with which I'm doing some research on how to use local large language models for working with restricted data, with data with which you can't send queries to frontier models. So, for this, you or an institution you're working with must have procured GPUs, put on local large language models. Do they say that some, you know, potentially accelerated

45:46workflows are being bottlenecked? So, the question is, can local large language models be helpful for this? And so, I did just like a back-of-the-envelope thing as part of this to ask a bunch of like syntactical questions to different types of models. And this is a popular coding agent, Python 3.8 27B. I can fit uh like a $4,000 or so worth

46:14of GPU. Um and it probably just based off this performed poorly likely to be usable without any skill, which is the way I tested it. But an interesting thing to look at for me would be to see if I added been a skill and like gave it some context on what state it is, how to use it, does it work

46:36better? And so if in some of your work you're using smaller models for whatever reason, that's something to be aware of. So the main suggestion I have with respect to this mistake is to revisit skills when models change. So I didn't quite understand why was your YouTube video made obsolete? Because the models themselves have progressed to an extent that you don't need to give the model explicit instructions as much as at all or as much of that described in that video. Uh here's what state it is, it's a tool that economists use. Here's like the whole Here's where the PDF is located on the system where you can find its

47:19instructions. It very likely sort of knows that already from the model training process itself. So it can you can just go and start right away then. Yeah, without all that preamble. You should test the hypothesis. So you and maybe false and if it's false for whatever reason, just make a Yes. It's a bit of a digression but I'm curious since I'm on Stata, I mean how how far do you think we are where you know even telling agent what kind of statistical language they should be using is relevant for the quality of the results. My My sense was that it was recently I'm not sure if I remember this

47:58correctly, but I think there was a piece of research that showed that it actually doesn't matter when you tell it to to use data or MATLAB or Python. Mhm. It will basically arrive at the same set of results nowadays. That's all good. Do you feel like it's still relevant to prescript I mean, of course, if you want to read the code and you're more familiar with one language than another,

48:18then I'd get a reason. So, like the short answer is I don't know. And then my long answer is that something that I feel is probably a missing piece in the economics profession is evals in general of like different models in different situations on how they perform on economics research relevant tests. That's what you would want concretely to

48:46be able to answer that question. Otherwise, right now it's just kind of hearsay. And like from what I did in preliminary literature review, like these evals don't exist. That's what I'm interested in doing with this grant for like local models, but it's probably important to also have other frontier models. Yeah, it might have been something else

49:10or not. Cool. Okay. So, the main suggestion I have with this mistake is to revisit skills when models change. And so, what you'll usually see when a new model is released that the lab will put out guidance on how you should work with this model relative to previous models. And so, this for example is what OpenAI put out when its GPT-6 Astra came out. And so, if that becomes the main model through which you're using the skills, what you can do is just tell your coding agent to take a look at it, take a look through your skills, and see given the guidance being given in this document on how you

49:54should now think about skill descriptions, agents.md, and etc. With this new model, what should we modify? Yes. So, you you recommend doing this cleaning up like only when new models are released or just regularly then? I would say like definitely do it in an instance where you notice it breaking in some way that's bothering Like I don't do it that often. It's just like this

50:23particular type of cleanup. I I guess what I'll say is that people do evals on how much more efficiently they're able to work with their tokens. Like before and after doing this kind of thing. And sometimes the differences that I'll see, which may not be in a context that not perfectly to mine, but the differences that I'll see from people I trust with the evals is like 30 or 40%, which is like enough that, you know, I I I want to do that

50:51kind of thing. And so like there just could be instructions that you had in the past which will push, you know, the tool in a wrong direction in some way that's not necessary anymore. It naturally goes in the right direction because of how the training process has been improved.

51:11Okay. Now, a mistake number six I think will be especially relevant to all the development economists, really everybody. It's not making skills for a project management. So, I have two most valuable skills in my opinion. There's the one that matters in my email off. There's the one that I use for researching across all my call transfers. These are the two skills that I can least live without. So, a lot of senior faculty are no longer in the weeds with code, but everybody is managing many projects and struggling to

51:46keep up. So, getting call transcripts and your email to your engine is critical. And one thing I'll say to you, is like this is basically a judgment call with respect to how you feel about the privacy considerations of giving these tools access to this kind of information about you. I can talk a little bit about what like the claimed and actual privacy risks are. And part of my interest in doing this local large language model research is that there's a big segment of the population who's not getting these productivity benefits cuz they have legitimate, you know, privacy concern. And it's possible that a reasonably priced local large language

52:29model could get them the same, you know, use with possible risks. So, we can even if the Zoom is protected by password, is it possible you can get access? So, what you would get is the transcript. So, you need a mechanism by which to get the transcript of the Zoom. Okay, you cannot just go in and do anything and download the transcript

52:50yourself. It can also do but I'll I'll show you the mechanism I recommend. With your slides. Turns out to be quite difficult on the US. Yes, all my meetings are Zoom. The the yeah. They have to allow it. I record it. [snorts] So, let me describe kind of the value that I feel see of email plus AI the benefits. It's going to be context for your research and your non-research

53:17projects. Speaking frankly, I have some reimbursement I'll have to fill out for this and it's all going to be done by AI. Especially when combined with computer use, I'll just like figure out what form I have to put in all the receipts, stack them all together. I'm no way I'm doing that manually. Um drafting emails, I find it quite helpful. I don't like send the emails immediately with AI. I'm going to show

53:42you how to protect against that. Very common concern, okay, what if the AI sends emails asking me? Like what I found is that it's literally sufficient in your email skill to just have something like draft first, ask for my approval before sending. You know, make it like send strong. Never ever ever send without asking for my approval first or like something like that. And

54:07that that basically gets you there. Okay. So, just mechanically, if you want to connect your email, the Cloud Code and Codex email plugins only permit connecting one email inbox. I personally have three inboxes which I use for personal and some different work projects. So, I use a tool called gogcli to coordinate across multiple Google accounts. It looks like this. The tool is made by Pete Steinberger. Does anybody know who

54:37that is? Yeah, he's a Yeah, he's the guy Yeah, he's the guy who created Open Claw, this like crazy AI assistant that got popular, you know, a little about like about a year ago. And so, if you're setting this thing up for the first time, it's pretty involved, right? There's like a lot of steps here. And so, one thing I'd recommend is that if this intimidates you, one, pay a poor grad student. Two, you know, you can just ask your coding agent to walk you through it and maybe even try having your coding agent control your browser and fill out these forms for you that you have to go

55:12through. And so, that's, you know, my advice if you decide to go this route and want to set up your inboxes. Now, let me talk about the value of having call transcripts plus AI. So, obviously, similarly, it's going to be context for your research and not research projects, but it's much richer context than you will get with just Titan. And so, it's analogous to some of the advantages if any of you use voice to text tools. So, I use voice to text pretty heavily for my work, but it's selectively. Like sometimes when I'm writing, I just need to write every like single letter by hand. But sometimes as part of my work,

55:52I'll like be you know, sometimes you have like a feeling that there's like something inside you and if you just talk for 5 minutes, you'll be able to get it out. And so when I have that feeling, I notice that feeling, be okay, time to take a walk, get a coffee, and I'll just talk. I'll be like going around talking to myself like a crazy person for 10 minutes on my phone and then I have it set up so that the phone is connected to my computer. All that gets as input to my computer and I'm able to extract out this thing that

56:21was in my head. Yes. Yeah, that the problem for PhD students is that we don't look old, right? Like we often don't look old. We don't know what our advice is all about is. So, I have a solution for you. You have I mean I I take I think about my what I had. Yeah. Like I I I take notes on Notability on my iPad

56:40Yeah. and then go back to to my my laptop bro. And uh both both read my notes from Notability Yeah. and create some tasks. Okay. It's great. No, that's It's been so great, right? Yeah. Like is there anything more efficient? Yeah, one one site. Yeah. I just want to reinforce something you said which you know, some people might not take as serious as you uh maybe intended it, which is you know, that this idea that you just like stream of consciousness write prompts where you just talk for 5 or 10 minutes. It seems like doesn't doesn't that seem like it would really create really weird instructions? But even some

57:20of the leading people at uh Anthropic and Open AI when they talk about their own prompting these days, it is some If you read their prompts, it is just stream of consciousness. Just all that feel like this. I want to be beautiful. You know, maybe you should do a little of this and you know, I really want to be impressed. It's like a lot of it is just like emotions. So, you know, I think it's easy for us to be like all this is kind of BS if we just talk, but you know, at least that that seems to be worth Yeah, so I think Andrej Karpathy, if you guys know who

57:51that is, one of the founders of OpenAI, now I think head of pre-training at Anthropic, he had a tweet where he basically described what you know, Dave was describing that he just like talks at length. Now, the thing that has to be said is that I do believe that people who are skilled verbally get better results out of these things. And so, um like eloquence or precision of your speech probably like increase in value when used in a complimentary way with these tools. So, um that, you know, like a toddler jabbering for 5 minutes

58:31is not going to get useful results. Yeah. for that like getting the skills that like translate your stream of consciousness into a structure. Yeah, so I think so so sometimes like if uh in fact, this is a still that I ought to make because I often find it like not using my stream of consciousness in the way that I want it to. And so, but I guess like there isn't enough commonality between the different ways in which I want to use it. So, I should probably go through my session transcripts and have it identify the stream of consciousness and see the ways in which it erred and how to use it, how

59:14I had to correct it, and then suggest the appropriate skill to make. Yeah. I should do that. Cool. Okay, another like really great feature of these call transcripts with AI is the ability to act quickly. So I'm working on a project right now with a less technical co-author. So we have a call and you know right after the call I'm able to tell Codex to act on it. You know I approve. Right now we're doing pretty simple stuff but it like it's more just simple things that are made clear through clear thinking and so once you have you know like an hour of jabbering at the end of which in the

59:58last 10 minutes like some clarity arose Codex is able to act on that like infer that from the transcript and act very quickly on that 10 minutes of clarity and then I even have Codex printed in the WhatsApp as like where you usually message but I just tell it to send a call to the other message and that's

1:00:16been our way of doing things. Um if you guys ever like enter into private private sector do any kind of consulting as well that ends up being very helpful to be able to respond very quickly to clients. Um Also Fathom and Granola are the two tools that are available that like are the ones that I would recommend for

1:00:42getting these transcripts. So Fathom is what I typically use for video calls if I have a Zoom or a Google Meet I'll invite in the Fathom. And the important piece that you want to make sure you whatever tool you have has is an API by which you can easily extract all the transcripts. Like you could always log in and like cumbersomely get the transcript out but it's annoying and like really with these things the greater degree to which you can remove any frictions that can be removed the better use you're going to

1:01:15be able to get out of them. The other tool is Granola and so uh, Granola can be used outside of the context of a video call, can also be used in the video calls, but importantly, like, it is something that you can pull out if you're having a meeting with somebody at a coffee shop. It's a very easy and convenient form

1:01:37factor in which to get these recordings. The difference between Granola and Fathom is that with Granola, um, the audio is not kept or the video after the recording is deleted, so you just have transcripts and notes. The Fathom does keep the audio and video, so if, you know, you guys are sharing a screen or something and that's relevant context that you want preserved, or for whatever reason you want the audio also preserved, like this is a talk you're giving and then you want to be able to share it on whatever you YouTube, your website, that you would want Fathom for that. And so the thing about Granola is that you have

1:02:15to pay not a terrible amount, like $10 a month, to have this API access. And so I use the free plan for Fathom and then I pay for Granola. That's the way I do things. Yes. I use that your free version means like you you you'll either have to ask for permission who you are. Correct. So, like, uh, You would have to do it, uh, without Yeah, I mean, I guess, check the laws in your state. I don't know how it works in

1:02:42California. [laughter]

1:02:58[laughter] Yeah. Just bank from that and I like it would give you a much better video log than just giving the the just putting the transcript into a into a enclosure. Uh, so this, like, the main use of this is to get the transcript and that you give me one slide to then be able to get

1:03:20it into Cloud Coda Coda. If you get the transcript from Zoom, then you can just put the transcript there. That will work, too. Yeah. Will you go out with it automatically? Like this automate. So, like what I what I'm not I think it will automate. Yeah. What what I'm what I'm not sure is like is it like truly automatic or it happens instantly? Does Zoom have the API or do you have to pay to get access to the API? I think that like given how much value it adds if you like consistently have it in there, if it's the case that this indeed has some advantage over Zoom, it's probably worth

1:03:55the $10 a month for how how useful it would be. Can you use WhisperFlow for So, I use It has a note taker as well. Yeah, so WhisperFlow is the speech-to-text tool, which if you haven't used before and you have an EDU account, you get 3 months for free. Uh I had a year for free um because of my YouTube channel, they just

1:04:19gave it to me. Um I don't plan to renew it because I think there are free tools that are like doing as well this speech-to-text kind of stuff. I'll see if I ends up being the case that I really will miss WhisperFlow, I might go back to it. But, as you're saying, dude, yes, the WhisperFlow, if you have that subscription, now seems to have a note

1:04:39taker as well. What are the free ones? Is it There's one um I don't know if this is free, but it might come with a another subscription I'm going to talk about in a few slides to an organization called Every. It's called a monologue. And so, it comes free with that. And I think there are other like genuinely free ones. But, I don't know them off the top

1:04:59of my head. 10 minutes. Okay. So, you want to make these transcripts available to your agent. And so, by some mechanism, you capture these transcripts, you connect them to, you know, an integration or API, and then you have some still which is able to hit that API and bring in the relevant call and its context into your project. Or, if you're being a pro, you have some automated mechanism by which the transcripts are pulled into a central location, and so they're always available to you. And that's more what I

1:05:39would recommend. Okay. Uh uh mistake number seven, which probably will be the last one I have time for, is not composing skills. There's only one more you'll miss. So, I'll send out the slides. So, skills can call other skills. Back in the old days, when we used to write code, there were these things

1:05:59called functions. When you were doing the same thing many times, you'd stick it in a function, and then functions would call other functions. So, you can and should do the same thing with skills. So, as an example with making these slides, recently, like these slides and all my slides, I've been making them with HTML because they're much more flexible. You saw some of the nice little animations I

1:06:24had. And so, what's unlikely to hit the mark when making slides is to do something like take this paper and make a deck out of it. I think you guys have seen this. This is total trash what comes out. Can't be done. And so, instead, like I think a good way to do things, and we can talk about the exact process by which to get there, is to get the slides out just in bullet points, not in design, not in LaTeX, not like the full Beamer output. This can you like get what you roughly want in slides organized in bullet points in a

1:06:58text file. And so, a way in which I'll talk or work sometimes is, if I have, Let's say set of slides that I want to make and like in this space, I think something's going to fit and here's roughly the things that I want inside. I'll just talk for 1 or 2 minutes and then I'll ask using some like skill I've made here on having the large language model help me think about how to divide that into slides in

1:07:25this space. Then once the slide text is set, I'll think about how to visualize it. Should we call on it? Should we have a figure here? Should we insert some images? And so I'm creating these sort of different skills which are composed which kind of hand off to each other. And so, you know, question here

1:07:45As a question. Yes, sir. And a even the most advanced models cannot do that paper and make the slides. They they can. It just isn't good. Like I mean, it depends on what can means here. Like they they do make some of them. You know, like what it says it's crap. I just if you if you present that,

1:08:04everybody knows that's embarrassing. I can't skip. And I I don't think it's like everybody knows and it's embarrassing to you just because you're using AI is because it's like actually not good. Like there's things that don't follow. There's like irrelevant composition on the slide. Right? Yeah. It's still not there then this this the

1:08:29technology. It's a good question of like what whether Yeah, whether we reached that point or not. Like in some tasks, I don't know. So, like when you're in this situation where there are these like skills that could fit together, kind of like I was describing before, like we add these different serial examples and the question was is this a skill shape problem or exactly what kind of problem is it with all this serial research related stuff? Similarly here, like LLMs can help you sort of figure out the correct shape and composition of skills, but they're not perfect. So, for example, that grilling skill that I mentioned before, I find it quite

1:09:12useful. That skill prominently was not written by a um through like the use of the skill creator skill. It's written by this developer, Matt Pocock, who has like very opinionated tastes on how these things should be made. He's also a very good communicator. And so, he's been able to make good skills or probably leveraging some aspect of something similar to why LLM's are just bad at making slides. There's like some aspects of communication that don't come across. And so, some types of things can be figured out just by asking that AI, and others really can't. And so, it's analogous to why I think in like software development and also just

1:09:53an industry, humans are still seen as and genuinely are needed for like bigger picture architectural and product decisions, but also in, you know, research itself. Like, probably, I'd be curious of your opinions, but I don't know what the quality is of the ideas necessarily that you get out of these tools. Like, are they actually good? I

1:10:13don't think that they're that great. And so, for some skills, like deciding whether it's the case that you have created the correct composition, it's possible to like explicitly evaluate. And so, for example, if you have a literature review skill, where all the skill is being demanded to do is something like find me all like the labor papers studying minimum wage in, I

1:10:40don't know, Idaho, something like that. You could like create explicit tests, cuz you could find all those papers and then see how well a skill does at getting the LLM to recover things for that test. Though, there's other kinds of literature review that you may want to do which the skills are which for which even a literature review eval may be very difficult. And then there are other kinds of more taste-based things like a web design skill or anything with an aesthetic or taste component for which evaluating the skills like explicitly numerically would be quite difficult. And so overall, I think like humans are still in this aspect of like AI and usage

1:11:25quite integral. And I'm going to give one example if I have time in these last 4 minutes of a way in which like skills can sort of be composed. And so this is an organization called Every. They're uh media organization which I think actually writes about AI for let's say a sophisticated general audience at a level which is quite useful for economists as they're mostly focused on sort of general knowledge work. And so if you have like a spare 100 bucks a year, it's probably a good subscription to like check out. They might have a

1:12:02free trial. So I subscribe to them too. And they offer free skills. You don't have to pay for them. And I've been using their skills for over a year. I find them very helpful. And so they've created this plugin for Cloud Code and Code X. For a plugin you can just think of a set of skills made for distribution which is called compiled engineer. And in this they have a set of skills which I would really recommend trying out which have this sort of comp- composition nature really well got out. So for example, when you want to do something and you need to think through what your exact

1:12:37requirements are with what you should do here, they have a brainstorm skill which I use all the time. Like every day probably a few times. Anytime I'm starting to brainstorm something, I I like the way their skill works for this. And then in their way of thinking about things and composing things, after you've brainstormed decided like what it is you want to do, you need to think about the implementation order of what it is you want to do. And so these two in particular, like if you're a grad student working on a fresh research project and like you're thinking about doing something in spit the space, the

1:13:15brainstorm skill here can be quite useful for like thinking through what to do. It might collection, various steps, or whatever. Well, the plan skill is going to be very useful for thinking about the order in which to do things. And so similarly, they have like a skill for implementing work. And so this is what I was kind of alluding to when you were asking like, "Okay, the Python file, where does it come from?" That's like a skill that could be used as a standard way of actually writing your

1:13:43code. They have a skill which I like using at all times in situations where you're considering two possibilities and you need like a second view on how to think through these two possibilities, they have a POV or point of view skill. When you have a document to review though, a document review skill and I

1:14:02have a code review skill. Now, the code review skill, just uh as like something you should be aware of, will often work in ways that are not that useful for economists. It'll start doing like security reviews or like looking for SQL injections, things that don't matter for economists. But these, you know, pieces, the brainstorm and plan, I do really recommend. POV I don't really recommend. Work if I had a little more time I would describe some nuances of that. But the general thing that I would sort of like suggest as you get more advanced or interested in thinking about your workflows is something like these could be quite

1:14:43useful for like having a mechanism by which to have a critical eye at your own work. And so, a mechanism which um like several people have that I've talked to or respect or like gotten use from is having something like the document review skill, but something to criticize their own papers and perhaps having like different personas which are parallelized checking their papers. And so, famous economist has the people that you all respect are embedded very deeply into the weights of frontier models. So, this means that if you want to have like a Gary Becker skill and is like looking for the things that Becker would look

1:15:25for or a Lars Peter Hansen skill or whoever it is like uh skill, you can pretty much just ask the agentic coding model to create that and it will be able to do so um without even necessarily you need to go into any external resources just from its like own weights. And so, the last piece which I didn't get to is not version controlling the skills. So, use Git, but that'll have to be for another

1:15:52time. Thank you. [applause]