Build Slide Decks in Claude Code

Shane Butler: Alright, I’m admitting now.

Hai Guan: I’m gonna share my screen here.

Shane Butler: Hey, everyone.

Sravya Madipalli: Hello, everyone!

Shane Butler: How’s it going?

Attendee: Hi, hello, everyone.

Shane Butler: Hey, yeah. Can you all see the, the screen? Can you see the slides that are being shared right now? You can drop in the chat a yes or no.

Hai Guan: Nice.

Shane Butler: Perfect.

Hai Guan: Nice. Cool.

Sravya Madipalli: Awesome.

Shane Butler: Maybe as we get warmed up. If you want to drop in where they’re coming from, where they’re viewing from. I’m based out of South Lake Tahoe, California. You drop it in chat where you’re from. It’s nice to see, kind of the distribution of regions we get. So we got Toronto. Wisconsin, Amsterdam, nice. Calgary, Illinois, Berlin, more Amsterdam. East Bay, Fremont, SoCal. San Francisco, another Toronto, yeah, we’re from all over. That’s cool. Munich, nice.

Hai Guan: So weird.

Sravya Madipalli: So nice to see people from everywhere, and also a bunch of California folks, too. Nice.

Hai Guan: North Carolina, U.S, welcome, welcome.

Sravya Madipalli: Awesome. We’ll get started in, like, a couple minutes, waiting for everyone to come in.

Shane Butler: Yeah, we got a lot of people rolling right now. Miami! Oh, hey, Attendee. Nice. Texas. We got some next-door people in the… in the house. That’s cool.

Sravya Madipalli: Awesome.

Hai Guan: Yeah That’s so cool. Yeah, we’ll give folks a couple more minutes to row in, and we’ll get… we’ll get started.

Sravya Madipalli: For the new people who are coming in, we are basically asking people where you’re from. We have people from everywhere. California, Munich, Europe, like, so many different places. Illinois, Toronto. So if you’re joining in, please Add where you’re from, yeah. Nice.

Shane Butler: San Jose, yeah.

Hai Guan: San Francisco, San Jose, Seattle, New York. Wonder if we can hit, 50 states in this session.

Sravya Madipalli: That’ll be pretty cool.

Shane Butler: Another free month.

Hai Guan: Lots of Fremont folks. Maybe, your neighbor, Sravya.

Sravya Madipalli: Yeah.

Hai Guan: No, Attendee is. Alright, let’s see… I think we can probably get started.

Shane Butler: Cool.

Hai Guan: Alright, cool, let me… All right, so, great. Welcome, everybody. Thanks for joining in. Today’s lesson is on how to build slide decks in Cloud Code. And hopefully, hopefully this is, sort of, like, something interesting for you. And, by the end, hopefully you’ll be able to at least see the capability of what Cloud Code can do on things beyond just coding. So just a quick intro for all three of us. My name is Hai. I am the head of data at a legal tech company called ENTRA. I… was previously at other consumer tech companies, at Nextdoor, LinkedIn, Pinterest, Meta, so on and so forth.

So, been in the data science analytics space for, almost 2… or almost 20 years, now, and, yeah, excited to, to lead this session and, talk to you guys about this. I will hand it off to Shane, or Shravia, pass it to Shravia.

Sravya Madipalli: Hey everyone, I’m Stravya. I’m a data sense leader with Expedience in Microsoft, eBay, Nextdoor. That’s where I met Shane and Hai, and most recently at Superhuman. Yeah, very excited to share everything we are doing in the analytics and AI space. Cloud Code, and, you know, you also started working to look into Corex and other things, so excited to share with you all our learnings as part of these workshops. Head over to Shane.

Shane Butler: Hey everyone, I’m Shane, yeah, Lead AI Analyst Lab with High and Stravia. I’ve been in data science for about… 10 years, currently a principal data scientist at a same legal tech company that HiWorks at. I specialize primarily in agentic analytics and AI evaluation. But, today, we’re gonna be talking about Psydex, so… Pass that off to Hai.

Hai Guan: Cool, yeah, we’re gonna talk about slide decks, without touching PowerPoint, so hopefully, this would be something interesting, and you’ll walk away with a little bit of mental model for how to do that yourself. Okay. Cool. Let’s see. Oh, and just before we get into the rest of the content, so we’re all from AI Analyst Lab. You can check us out on AIanalystlab.ai. We have a lot of free courses, free workshops of the past that you can watch recordings for. We also have a lot of upcoming Workshops and lessons just like this, so you can sign up for those as well.

We run a couple… boot camps, or I guess weekend workshops, that you can join us as well, that we get more deeper into, Cloud Code, in terms of building AI analyst systems and, and, and setting up Cloud Code and things like that. We’ll get into that a little bit more as we go. Cool. Okay, so here’s, here’s the scenario that we’re gonna, that we’re gonna walk through. So suppose you have the data of what… whatever you’re analyzing. So, you know, we’re all data professionals here, the three of us, we typically have to do a lot of. Kind of, like, presentations and, deck buildings to convey our findings from the data that we have.

And so, in the upcoming, you know, like, rest of the lesson, we’re gonna ground it in kind of, like, the data, analogy, or the data world, sort of. And so, you know, like, we do… or maybe many of you do as well, do a lot of, like, data analysis or some sort of information gathering, and you have a deck that’s due tomorrow. And most of the time, I guess in the old way, for the most part, it’s a, you know, spend a lot of time in either PowerPoint or in Google Slides, and maybe there’s some templates that you already have. That matches your company’s themes and brands and things like that, but, like, you still have to, like, drag and drop and do all of that kind of stuff. Across the board.

Now, with AI tools and, just, you know, generic, general, sort of, like, horizontal, chatbots, or other agents that can help, things like clock co-work and stuff like that, you now have a lot better control and a lot better tooling to help you do this, where, you know, perhaps you can fire off a prompt, and then you can, you can get some, some stuff back, pretty quickly. So, you know, like. things have advanced really, really fast, and there are also, like, other, tool building or presentation building sort of tools out there, like Gamma, so on and so forth.

What we’re talking about today here is to leverage clock code to do the end-to-end, such that you can, like, the final output of a deck could actually look Look like something that’s been reviewed by, you know, like, multiple, kind of like, experts, and things are very coherent in terms of the storytelling aspect of it. So, I just pulled these random screens from, from kind of like a, using Claude Cowork to prompt for a presentation deck. So, many of you might have heard of the term AI slop, right? Like, or I guess, maybe, let me, let me, let me ask you this. Drop in chat if you… If you feel like this… these… Slides look great. Okay. Attendee says no. Stravia says not bad.

Attendee says, they look okay-ish, okay? Seen worse, for sure. Okay, wow. Looks clawed. Not great. They don’t feel like real-world decks. I would put together at work. Average. Yeah, so they’re pretty mid, right? Like, it’s… looks pretty generic, it’s got the themes that feels like it comes from an AI, and for the most part, well, I mean, like, you know. to be fair, I didn’t know what the topic was on this one. Apparently, it’s for Octopus. But, you know, I don’t know how interesting you can make octopus, but just looking at the look and feel of the slides here, it doesn’t look, you know, like. super compelling, right?

Even if you expand it to the entire deck, it’s probably gonna be something very similar for the rest of the… of the sequence, if you will. So, So, the reason why, kind of like, you know, if you don’t give it enough context or instructions that AI defaults to something like this is… pretty simple, because it’s trying to just do the fastest way of generating… accomplishing the goal that you set it out to do. I gave it a pretty generic prompt of, hey, build me a deck about a random topic, and then that’s what it gave me.

And so, the power, a lot of times, in being able to leverage tools like Claude, or Claude Code, or ChatGPT is the ability to give it instructions, and so we’re gonna go through what a well-chained, kind of, like, sequence of instructions would bring in terms of the outputs and the, and the difference of it. And if anyone has any questions, along the way, feel free to drop them in chat, and we’ll answer them, or, you know, raise your hands, and we’ll have a Q&A at the end. Cool. Okay, so, what we’re gonna be talking about is, all of you might have seen… might have seen our LinkedIn posts, of the past, or have been at, past Lightning lessons that we’ve run.

We’re gonna be using the… we’re gonna show the… the AI Analyst repo that we’ve open-sourced and created for everyone to freely download and work with it. Basically, what this is, is an agentic system that you can work with in Cloud Code that basically mimics the, the, the ability of a senior product data scientist. And so, you know, like, feel free to, feel free to clone it, feel free to, to take a look at it. It’s got many different agents, skills that effectively is trying to do something that a product data scientist would be doing, and all the best practices that go into it.

So, it understands how to frame better questions, it understands how to, how to define metrics, understands how to Provide guardrails for the metrics, understands how to do, experimentation, and then, obviously, the final piece of it is, like, making recommendations, having… getting the context about the business, and then making connections of them, and building decks, and stuff like that. So we’re gonna take a look at the deck building component of, of this Full pipeline, and we’ll go into some of the, some of what it does, and then… Have it, sort of, show us what it can do at the end.

We’ll, let’s see, oh, and if Xavier or Shane, if you want to drop the link to this guy, so people can clone it, that would be, you know, folks, feel free to follow along as well.

Shane Butler: Yep, dropped it in the chat.

Hai Guan: Okay, cool. Great. So, let’s see, Cool, okay. So, you know, like, we have the… we have the Agentic system freely available. We also run different workshops and bootcamps where you can… you also can get to kind of, like, learn it with us in terms of… both how to set it up, how to run against it, and also how to set up something very similar for yourself. Like, if you want to build something from scratch, we have a bootcamp that teaches folks how to… how to actually do it. We’ve run one in, back in April, and, the, the, sort of, like, the vibe was really, really, really, really cool, and, I think our students got a lot out of it.

We’ll share more about, the specifics later down the, later towards the end of the, of the lesson, But just keep in mind that, we have these things, that we run these things as well, and so if you’re interested, we’d love to have you join us. And perhaps, Shane, Dravia, you guys can drop a link to kind of, like, those things for people to look at as well. But we’ll come back to it. Okay, cool. So, so what I wanted to get into is kind of, like, the few components that, that makes, decks look better than, for example, like, if you just prompt AI out of the box, for, for a deck.

So, the four things that are pretty important to… to nail for AI to have enough information to get a really good-looking and compelling deck are these things, and would love to know if people feel like, you know. more should be… more should be included, or, like, other things that you’ve, that you’ve seen that works. Number one is obviously the audience. Like. telling or understanding who you’re presenting to and having that context available is pretty important. Is it an executive readout? Is it a workshop like this one? Is it a team stand-up? Or, you know, like, whatever the setting is.

really gives the high-level, overarching kind of context setting that’s really important for both the tone, but also, like, the information architecture for what should be included in the deck itself. This is just like humans, right? Like, this is… You know, like, we always think about who the audience is, and then tailor the message based on that. If it’s for, you know, like, in the world of data, if it’s for, like, the data scientists, you can go… probably go pretty technical on the… on the deck itself, and explain methodologies and findings and whatever it is in a very detailed manner, and show your work very extensively.

But if it’s for, for example, like, a CEO, That would be completely different. You’re probably gonna mask away 90% of all those things, and then just show the takeaways, the important evidence, and what you’re asking for. So, and then everything else would be somewhere in between. So… It’s, you know, like, audience is the most important piece, where having that context helps really shape the, the, you know, even how many slides. we should have in the, in the deck itself, versus, you know, some other, some other audience that would be caring about something completely different. So that’s number one. Number two is the context, specifically the decision that the deck is driving.

So, for example, here is, you know, like, hey, if I have a deck that conveys you know, like a, like a topic around should we be shipping a specific feature, then that’s a very different, a very different messaging than if, we’re producing a deck for, are we on track of, like, a, you know, progress updates. And then that would be very different from if you’re asking for resources to further some sort of investment or, or, go… Double down on some sort of a… some sort of initiatives. So, the… Decision is also really important here, where, you know, like, in addition to the audience, is the intent of what you’re trying to do with the deck. Make sense so far? Hopefully.

And then once you have those, then you need a story, right? Like, you know, hey, what is the narrative? So… Right here it says storyboard before slides, so, you know, like, typically there’s a structure for what works, in terms of the flow of how you’re gonna deliver the presentation. So, it might have something like a TLDR, Might have the context setting so that everyone’s on the same page before you get into the detail. Certainly for the case of analysis, and presenting findings out of it, obviously the actual findings that you saw. that are most relevant for the group of people that you’re presenting to, that’s really, really important. That’s the meat of it.

And then also, like, you know, the recommendation, like, so what? You know, having gone through the actual presentation or the slide deck, what are you trying to accomplish at the end of the day? So there’s always, like, there’s different components that make up depending on the audience, again, that make up your sequence of messaging and what gets accomplished at the end of the day, but that’s all can be wrapped into this whole category of something like a narrative arc. It’s like, you know, like, at the end of the day, what is it? And it’s not a one-size-fits-all, it certainly is you know, not a chart dump in chronological order.

I don’t know if A lot of you have seen some, you know, like, let’s call it, like, a lot to be desired sort of presentations, where it’s just one chart after another, with very generic titles, and at the end, you’re like. Okay, then, like, what is, what is going on here? What do you want me to take away? And, you know, left scratching your heads. So, what we’re trying to do here is hopefully steer away from that sort of output, and then we’ll see how AI can do it, itself. And then finally, here’s the, more, more kind of like the, The user-facing aspect of it, so what is the theme of the, of the… of the presentation theme, meaning, like, literally, like, what’s the look and feel.

So, you know, perhaps… and we’ll show a few themes, as well in the demo, but, you know, like, if something is more fun or whatever, you know, you can have a different theme than if it’s something more formal, for example. there’s a reason why, what do you call those, like, McKinsey’s or the consulting companies are very particular about, what they want in their, in their client decks and things like that, precisely because they want to make sure that, number one, it’s being consistent, number two, it reflects, kind of, like, the, the nature of, of… Of the, of the talks that… or the presentations that they’re giving. So, you know, like.

I think if it’s hopefully no surprise that, you know, all these things are actually carried over to the AI Analyst repo that we just talked about, and, it knows how to do, sort of, like, all of these, independently, like, by itself. So, the… we call it, like, the pipeline, so basically it’s like a, you know, like a sequence of steps for the AI to execute. All the skills and agents, effectively tries to orchestrate a, kind of like a sequence of events, or a sequence of execution steps that leads to the final output, and it mimics what we just talked about in terms of the sequence.

So, it’s got the storyboard where, you know, like, given a set of data, given a set of findings, so assume that we’ve done the analysis up front, because this is a building a deck sort of course, or a lesson here. We’re gonna, you know, not look at the actual analysis, or how that’s done, and stuff like that. So, in your mental model, assume that we have a set of completed analysis, and now we need to build a deck for it. And so the pipeline runs like, okay, now you have a bunch of, stuff, so let’s see what makes sense to build a, to build a storyboard. So, what’s the arc? What is the, what is the, climax?

What is the, what is the information that we should be telling people, and, kind of, like, plan it high level like that. And then the second piece is, the narrative. Like, how should I tell that story based on the story arc, based on the storyboard, and based on the actual data and finding that’s already been done? So it understands how to do executive summary, how to do which findings to pick out, what recommendations you can do, based on those, and then connect it back to, sort of, like, the business context or whatever else it has access to, to formulate the actual, you know, like, the actual messaging on the slides themselves.

And then they’ll build the charts, and then, you know, compiling all that together becomes the deck, and then the fifth… the fifth, the fifth step here is where humans come in to, you know, like, hey, maybe play around with it, give it feedback, and, have it iterate. And we’ll see all of that in just a little bit. Does this make sense? Like, generally speaking? In terms of the flow? Like, hopefully this is similar to how you would think about doing it yourself if you were to build it manually, right?

Shane Butler: Yeah, I think one of the other things to think about this is that, You know, you can do this for any use case, it doesn’t have to be analytics, but in terms of, like, analysis, which is obviously our kind of focus area, so much of the work goes into Cleaning the data, framing the questions, doing the analysis, reviewing the analysis, like, just, you know, so much work that it’s really easy at the end of the day to kind of, like. Hand off something that’s not in a very well-curated deck, that doesn’t have a story or a narrative.

To it, and it’s just a shame when that happens, because so many… so much, like, hard work that really could be driving decisions doesn’t end up being actioned on, because the story or the deck, like, doesn’t get, showcased or surfaced, In the way it should be, so it’s really nice. Even if you’re already doing this, to have a kind of, like, automated kind of co-pilot with you, so that every time, like, you complete your work, you’re able to make sure, like, we’re gonna check these boxes at the end to make sure we get that last mile of our work done.

Hai Guan: Yeah, that’s a good point. Yeah, definitely. The last mile, probably 10% of the work, drives 90% of the impact. So, you’ve heard, did this thing land, or, you know, something similar. It’s basically the… that’s the, if you’ve done the hard work, did it actually get people to be influenced by it? And so we’re trying to make sure we can leverage AI to do that as effectively as possible. Alright, cool. So, let’s see, so… Yeah, so what we’re gonna be showing, probably, like, right now, is, we’re not gonna look at any PowerPoints, we’re not gonna do any CSS work, we don’t need to be super precise on our prompts, and certainly we don’t need design background.

All of that is baked in in this, in the instructions and… or I guess the agents and skills in the AI Analyst repo, and we’ll see how that works. So just a little bit of backdrop. In the repo itself, we have a test dataset. It’s an imaginary company, an e-commerce company, think of it like Amazon. It’s, called Nova Mart, and, assume we’ve done this analysis where we saw we were looking at checkout to purchase rates for the company. And, we’re seeing some of these results. Now, the specifics don’t really matter here. Again, like, we’re trying to separate the deck building aspect versus the analysis or the analytics aspect at the beginning.

What I’m gonna show… what I’m gonna demo next is, given these contexts and given these findings, what does it look like when we build a deck with it? Alright, so let’s get into it. Okay, hang on one second, let me find my quad. code here. I am gonna switch screen, and hopefully you guys can see it. Did I share the right thing? Do you guys see a Claude Code setup here?

Shane Butler: Yeah, how’s, like, mobile checkout conversion has collapsed?

Hai Guan: Yes, okay, cool. Okay. Right screen. I’m always nervous, because I’ve shared wrong screens in the past. That, that didn’t lead… that may or may not have led to… Awkward. Things being shared. Okay, cool. So, let’s see. So, for those of you who are familiar, this is VS Code, and, pairing with Terminal, or pairing with Cloud Code and Terminal. So, for people who are not aware, VS Code is, like a, like a work… like, IDE workflow where you can browse your folders and files on the left, and you can have terminals at the bottom and file explorers on top. Whatever you want to make your You know, almost like a desktop equivalent to be tailored towards your workflow.

And free to download, feel free to do that. You can use any IDEs to run clock code, and, you know, like, certainly it’s, there’s anti-gravity, there’s a cursor, you might have heard of all those things, but effectively, it’s just trying to make your desktop Or, make your workflow as clean as possible. So, for folks who are familiar, this is Cloud Code in Terminal, and I am in the AI Analyst repo, and just, you know, just showing very quickly We’ve got a bunch of agents, skills, so agents are things like, files to… to make Claude code act a certain way.

They have access to different tools, they can accomplish certain goals, and just, you know, as a quick example, Data Explorer means, this file, when you ask Claude Code to act like an agent that can… that does exploring the data, it will be able to follow instructions for how to do that. And then we have a bunch of skills as well. Skills are recipes for how to do certain things, so… Things like, oh, what does metrics mean? How to, whatever, it’s, How to do forecasting, how to explore the data, that kind of stuff.

So, feel free to explore those, we’re not gonna get into… the detail on that, but what I’m gonna do is… so, we already have the analysis done with the Novo Mart, checkout, sort of, like, conversion, conversion, story. So what we’re gonna do is, So I’m just gonna do, hey, can you show me an ASCII diagram of the deck building pipeline? So what I’m asking for is, show me a step-by-step diagram of how you actually do, deck building in this, specifically in this, in this, agentic system here. So, it’s gonna take a minute to think about it. I’m on Claude, opus 4.7. And… let me zoom in a little bit… So, it is pretty much telling me behind the scene, when I use it to build decks.

what exactly it is that, that it’s doing. So, you can see, okay, here’s the deck building pipeline, and let me give you a visual of that. So, assuming, again, like, the analysis phase has been completed. So, it’s like, what’s complete before deck building, but once you’ve done, and, you know, this, This AI analyst is also able to do, you know, framing, design, hypothesis, explore, analyze, root cause, all these things to understand what the data or the piece of analysis that you’re trying to do.

So all those are kind of, like, done… Feeding in, and then this is what it looks like once we have the findings and all the opportunities and, you know, like, all the, all the detail from the analysis. So you can see that, oh, okay, there is a story architect that builds this storyboard. It has a narrative generator, and it reviews, sort of, like, do these things make sense? And then it’s got the loop of, like, oh, if it needs fixes, then let’s go back to the architect and keep doing it until… It passes these, different reviewers and stuff. And, once it’s done, on the upper tile there, then you go into, you know, Chart Maker.

So, making… building the actual charts to, to really get at what the story is. It’s got, like, a critique for visual design, so just make sure that, it follows the right, The right, formulas for how to… how to make those look good. And, just keep doing the, you know, like, hey, do I approve it? Not I, as the human, but, like, does the system approve it? And if it doesn’t, then give the feedback and keep going. And then storytelling, write out the narrative. And then after all those is done, then it goes into kind of, like, putting the deck together, and then so on and so forth. Close the loop, blah blah blah.

So, roughly speaking, there’s, like, a multi-step thing happening here, where it’s doing its own, kind of, like, understanding if, things pass the bar, and then it fixes itself, or, it passes, kind of, like, the… The task from one to the other until the whole sequence is, is accomplished.

Shane Butler: Maybe, just to jump in there, hi. Yeah. Yeah, that… I think this is one of, like, the, really strong differentiators is, like, being able to set these loops up where you work with Claude to set, like, here’s, like, my… you know, my passing criteria. You can have many different passing criteria. to where, like, I know this deck is ready. Because as you… if you’re thinking about it. When you build a deck, like, you review it yourself many times, you might share it with a colleague, your manager, some stakeholders that you trust. And you’re working all of those share-out loops into your first pass, right here.

So your first pass on the deck is… actually ends up being what you would get, like, 4 or 5 passes. But it happens really quickly. And it can… and you can have it, you know, break and ask you, like, oh, should I go through another pass, or it can just loop on autonomously, on its own. So I think that’s, like, one of the most powerful things, I’ve seen so far when I build decks with this, versus, like, you know, using something else like, Gamma, which is, like, an AI slide deck creator, or of course, like, doing it on my own, or doing something with, like, the Beautify slides and Google Slides.

I think those… those criteria can all be, like, custom to… to what your beliefs are, should be… the passing criteria should be. Also had… A question here in the chat for you, hi. would the output quality be different if we use the same workflow in Cowork instead of, Cloud Code and VS Code?

Hai Guan: So, the… or I guess, different, we also have an AI Analyst plugin for Cowork, so it’s almost like a, a companion version for Co-Work itself. The AI Analyst, if you port it directly over to Co-Work, it’s not gonna follow the the exact instructions because of the differences in, you know, like, the files and systems and the system prompts and stuff like that. There’s also the probability nature of the LLMs, so even if you prompt it twice in the in the same, in Cloud Code, it may not actually produce the exact same thing. Now, the data, as long as it’s generated from code, will be referenced the same way.

The actual kind of content and how it, you know, phrases stuff, how it frames things, how it picks what’s the most important thing to support a given storyline could be a little different. Hopefully that’s helpful. Alright, cool. So, I’ll move on. Let’s see. Okay, so now we’ve seen what this thing looks like. Oh, and by the way, I don’t know why I have this chart up here, just to show that some analysis was done on the mobile checkout screen, so something has happened. Okay, so, let’s see. So, I am interested in trying to see what kind of themes Do you have for presentations?

Because, like, I wanna… I wanna know, you know, what it has access to in terms of the look and feel for the final deck itself. So it’s gonna think a little bit, look into some of the skills, look into the directory for what is currently available, and then give me an answer. So, right now, there’s 6 on here, but it’s, you know, like, anybody can add as many as they want. It’s got some of these corporate, minimal, New York Times, economist, analytics, specific, themes, and things like that, so… So, you know, like, it’s got a corpus of pretty interesting stuff, so let’s… Let’s take a look at the New York Times. Can you… make, New York Times themed.

Presentation for the mobile checkout conversion… let’s see… It’s called story. The… Audiences, executives at the company. So, I just gave it this prompt, so again, like, this… the mobile checkout conversion story refers to an analysis that’s already been done, so something like this up top here. So what it’s doing is, it’s gonna go… refer back to the actual analysis, which lives somewhere in… under working. somewhere here. Latest… Yeah, it’s, outputs. Yeah, it’s got, like, a… like, a full analysis, done… somewhere here. I’ve done a lot of stuff here, so… not gonna pollute the… not gonna… not gonna make it a little too messy, but that’s the idea.

So what it’s doing is, looking back at the analysis, and then it’s going to build the end-to-end reusable and New York Times-themed CSS, putting together the supporting charts, making the deck, and then rendering it in PDF. And… let’s see… Existing checkout artifact, so it’s found the analysis, and then, And then it’s gonna do the actual rebuilding of the different charts and stuff like that. And what it’s doing here is it’s invoking some of the skills. that, that has instructions for what a good chart looks like, how to build those charts, what the title should look like, so on and so forth.

So there’s a lot of best practices incorporated in some of these skills that’s already explicitly been baked into, to, to these instructions for Cloud Code to follow, so we don’t have to, you know, like… so we don’t get that AI slob that we… That we saw at the beginning. Okay, so it’s looking at the story a little bit, and then, it’s gonna… read the real analysis, I don’t know what… What, why it was not real before, but… And then it’s giving me a plan for what it’s doing, so the to-do list, if you will. Now it’s like, okay, I’m gonna build the CSS, the New York Times themes. We didn’t ask for any… we didn’t give it any instructions for any of that, and it’s gonna generate the charts.

And then it will write out the deck and, render the chart in PDF. But we’ll ask it for an HTML that I can share with you all as well. Hopefully, this… Go square… WIC, I’m hoping. While it builds, any questions here? I know there’s a bunch of stuff that I kinda can’t keep up with.

Shane Butler: Oh, there’s a question, I can kind of answer this one, too, around, it says, like, with this workflow, does Claude develop predefined design templates for each theme, or create fresh design layout based on the content? So, like, the themes we’ve, predefined here, if you just, kind of had Claude code. Create its own. theme each time on its own, I’ve found they’re… they can be pretty bad. So we have a few different themes. that we leverage. It’s pretty good at, building themes with you if you have something you’re inspired by, like, say, like, hey, give me the qualities of, like, an article from, say, like, The Economist or the New York Times.

I want to generate a theme that’s kind of based on that. format. And then within that theme, it’ll create a few different types of slides, like, you know, title slides, or big number slides, or chart slides, or bullet slides. But it still will reason through and create new types of slides on its own, so it’s not necessarily limited to that, although you could make it limited to that if you wanted to.

Hai Guan: Cool. So, it’s run into some errors and stuff, and that’s all good. It’s, self-healing, so agents are autonomous in fixing its own errors and figuring out what’s wrong, all that kind of stuff. So, you know, like, for those who are new to Claude code, like, it, like… pretty much until it hits a point where it can’t solve certain things, it won’t really bother you, so it’s very compelling.

Sravya Madipalli: Also, There’s one question by, I think, Attendee about, the course details on the Saturday workshop that we’ll have. If, there’s a question about do we need… it says that they could bring their own data set, and how do they work with it, so…

Shane Butler: On the Saturday workshop? Yeah, so for the Saturday workshop, and also for the bootcamp. and also our 5-week course, we do have a… A practice dataset we use. That’s based on a, fictional e-commerce company. So, you know, you can think of it as, like, SKUs and orders and usage kind of stuff. Basically, like, a fake Amazon. So we’ll have that data set we can use, and then, yeah, if you want to connect your own data, or if you want to work with a, public dataset, that’s an option as well. We can chat with you in terms of, like, how to connect to that. But… I think your question here was, like, if you don’t have your own dataset, can you still do the workshop?

And the answer is yes, we’ll all start with that practice data set.

Hai Guan: Yeah, totally. Alright, so it’s still building. Hopefully it should be done soon. What we can take a look at, for example, is it generated some of the charts in the New York Times theme. Oh, now let’s start to… okay, it’s almost done with the deck, but we can take a look, for example. Well, too tiny, so… this is one of the charts that it builds with the New York Times theme. You know, if you’ve followed New York Times, this is roughly similar to what you would have seen in the, some of their articles when a chart is, embedded. It’s got the… different, I don’t know, the signature sort of red and gray kind of, kind of look and feel.

And with the other skills in terms of how to tell a story and stuff like that, it actually highlights the, like, the takeaway at the title instead of, you know, whatever the description of the chart is, for example. So, okay, cool. So it looks like it’s done, so it’s… so it did a bunch of stuff, but basically it outputs a PDF file, outputs a MARP file. used the New York Times CSS, so on and so forth. Now, I wanted to just, can you render this? I can show you the PDF as well, but it’s probably easier to do the HTML. Can you render this in HTML? And then I’ll share a screen for… of, what that looks like. This one should be pretty quick, alright.

Shane Butler: The HTML rendering is pretty cool, because you can also have it create speaker notes. For you. I don’t know if this one’s automatically creating that, but the speaker notes is cool because you can have it, you can set up different voices For yourself, based on your audience, or who you are, and, have your, kind of, speaker notes, you know. Reflect whatever kind of voice guide you have set up.

Sravya Madipalli: Also, there’s a question from Attendee around what is the advantage of creating deck in Cloud Code versus Cloud UI? This is too technical to create a deck. So, I just wanted to say that I’ve been a… I’ve not been a terminal person all my… carrier, I was trying to stay away from it, just use directly VS Code and all, but as you get used to Claude Code, it’s a very straightforward, approach. It’s almost like you’re chatting with it. Yes, it gives a lot more information, probably, like, the bash commands, but we don’t need to, like. understand them or, what they mean and stuff. It’s almost like you’re, you know, Cloud UI. Probably, it’s just… Getting to understand to work with it.

The one thing I can say, how is it beneficial than Cloud UI? Because I have access to UI, Cloud UI, Co-Work, and also Cloud Code. Cloud Code is a lot more flexible, and it remembers things so much more easily. You can work with context, you can work with memory, there’s so many more tools you could work with. It’s just the starting friction. I would say probably 2-3 days of friction, and that’s it. And after that, you would understand how much more powerful this is.

Shane Butler: I think the other thing with, you know, setting up with Quad Code is that you can automatically trigger creation of decks based on other things that you have going on. So if you have, you know, something going on in Slack or an email, It’s like, for instance, like, the deck here… is its own pipeline, but that’s actually just, like, one small piece of our much broader, more longer pipeline of our, like, analysis pipeline. So for us, it’s really nice, because we can not only just… it’s not like, I’m going to create a deck right now, I’m going to go into Cloud Code and create a deck. It’s, oh, I got a question from a stakeholder in Slack.

I’m gonna go in Cloud Code and start, like, framing this, reframing this question, doing all the analysis for this question, and at the very end of that, it’s gonna automatically kick off This, deck creation for me. So it’s also just, like. It’s just, like, being able to automate things, based on, like, other triggers is one of the, kind of, like. key differentiators of doing something in, like, cloud code, or codecs, or whatever.

Hai Guan: Yeah, totally. Alright, cool. So, let’s take a look at the thing here. So this is the deck it came up with. So the… It’s got kind of, like, the, bit of a… look and feel for the New York Times things, although it should have been a little wider, I would say. But anyway. So the deck is pretty much kind of like going through the tailor for an executive audience. It has done the minimal, almost like the minimalist, way of, showing the information that is relevant for the audience itself.

Notice there’s not a lot of clutter here, for example, and what were the findings, so on and so forth, supporting ticket, all of these were done by clock code, like, I didn’t touch, I didn’t give it anything in terms of, In terms of the, the, kind of, like, the display requirements and things like that. And so, you know, like… And then it goes to what we’re asking for, for an executive audience. Typically, it’s like, okay, we’re looking to ask for a certain certain things, and not just, you know, tell a story and leave without, without any sort of, next steps. So hopefully this is, helpful. And, you know, it’s got the appendix.

But this is a pretty short deck that it created, because the audience itself is, is, is pretty tight. So… That’s kind of like the… that’s kind of like the idea here. And, we can obviously, you know, like. if we switch back to the… switch back to cloud code, we can obviously, if you remember, just, 10 minutes ago or so, we have the ability to swap themes, right? Like, we have the, economist theme, for example, or the corporate theme, and so it would just be a matter of, like, you know, hey, I don’t like this theme, let me change… let me swap that. one prompt to clock code in the same instance, in the same session, and then it will… it’ll do the rest.

And then you go check, and you go iterate, and you give it feedback, and just keep doing that. So that’s kind of like the idea here. Alright, let me go back to here, and I know we’re coming up on time, so… Okay, cool. So, we did the demo here. Okay, before we get into Q&A, so a couple ways to, you know, like, we have pretty limited time in, in the lesson, but a couple, couple ways to continue to take the next step, if you’re interested. This Saturday, so 2 days from now, we have the Intro to Cloud Code Analytics workshop, which is 3 hours, for Pacific Time people in the morning. hands-on workshop.

We’ll help you install Cloud Code, clone the repo, run real analysis, and we’ll basically get you set up. We’re doing it almost for free here. We’re doing kind of, like, a $25 with this code, because we do want people to show up, and not just, you know, like, sign up for free and not, not take it seriously. And then we have, in another week, we’ll have a two-day bootcamp, ClockCo Analytics, so where we help where we work with you to actually build the system, so you can actually build something like an AI analyst yourself that’s tailored towards your own use cases, your own context, and your own expertise. So we’ll show you and walk you through how to do that.

We’ve done a cohort of this, and we’ve got really great reviews off of it, so we’ll send over the link in the chat as well. But, feel free to kind of, like, ping us for any of these. If you have questions. We have a promo code here for 20% off of that bootcamp that happens in, I guess, a week and a half now. Expires Saturday Pacific time. End of day Saturday, Pacific time, so, definitely take advantage of it if you’re, if you’re interested. Alright, we’ve got 10 minutes left, roughly, so I think we can probably open it up for questions.

Shane Butler: Gonna read through some of the ones on here. I think Attendee just had a question. Attendee, do you want to ask your question, actually?

Hai Guan: Oh, that’s a really good question, yeah.

Attendee: So I was curious of how does, this workflow differ from using Claw design? Particularly because, like, right now, I have some free credits that I can try things out, it’s about 20 prompts worth, so I really want to get value out of that. And then also, I want to keep the costs low with whatever workflow I come up with. So I’m realizing Claw design can’t be, like, that one place that I stick to.

Hai Guan: Yeah, I tried Cloud Design pretty extensively myself. They’re certainly a new product, and I’m guessing you’re talking about the slide deck builder component of Cloud Design, right?

Attendee: Yeah, exactly, but we are hoping to, like, build out our design system and start using that to rebrand old materials and create new ones, and I’m starting to realize, like, that’s gonna get expensive. So, I want to come up with, like, the most efficient workflow I can, but also take the… Leverage the value of this new tool.

Hai Guan: Yeah. Totally. Yeah, Claude Design, for those who haven’t used it, it’s, the new, the new system that Claude rolled out, probably, like, two and a half weeks ago now, or three weeks ago. Effectively, you can go to claud.ai slash design, something like that, where you can build interactive apps or do anything related to design, and building slide deck is one of those. I built 2 decks with it, and I exhausted all my… all my credits, and… I think over time, it’s gonna get really good, definitely. That’s… that’s no… that’s no question. Right now, it’s pretty painful to do. So, instruction following, understanding your intent, understanding your context. that’s been… pretty rough, I would say.

Now, over time, that would be different. The, you know, like, what Shane got into earlier is that in code, in Cloud Code, specifically in Terminal. it’s just a lot more flexible, and and, you know, like, I think the cost is a lot more predictable in a way. And so, you know, like. pretty sure design… cloud design is gonna keep getting better, and it’s gonna keep getting, like, being able to follow instructions and all that kind of stuff, improvements pretty well over time. Right now, it’s pretty hard. to do. And, you know, again, like, you lose the ability to sort of, like, chain different workflows together.

you have a very single purpose kind of like a workflow when you look at something like a cloud design, for example, versus, you know, Cloud Code and Terminal. You can chain it with whatever you’re doing all together, and just perhaps, like, oh, now build me a deck. After some of the other stuff that you’ve done. I don’t know if you guys have anything to add.

Shane Butler: Not for me.

Sravya Madipalli: I think there are a couple of questions around, it looks like… I don’t know if you’ve covered in the previous one, it looks like AI agents can operate here without being invoked by slash commands in the terminal. Have you built .md files in the background to be able to respond to you? That’s one question. I think another question, this was by Attendee, and there’s another question around something similar, is about how did you get, Claude Code to get the analysis for you done? Like, did you… like, you know, share… how was the analysis fed to plot code? In which format was analysis or data, stored? So I think both are kind of similar, so just want to do it.

Hai Guan: Yeah. Shane, do you want to answer the slash command one? And I’ll take the second one.

Shane Butler: Yeah, the question is, like, have we built agents and slash commands in the repo to do stuff? Yeah, we have, like, so this repo is, like, There’s, like, dozens of… Agents. Skills slash commands, and so a lot of those will… trigger automatically based off of, like, phrases that we’ve put in there, where you’ll… it’ll look if there’s, like, you know, semantically similar phrases, like, I want to create a deck, I want to create a presentation. It’ll automatically know to kick off that skill, which will then, invoke these… these agents. and this entire workflow.

You can also trigger them Directly, like, sometimes… honestly, a lot of times when I use, any sort of Gentex system with cloud code right now, I’m so paranoid that it’s gonna go off the rails that I’m, like, I first… prime it to say, hey, tell me… and I kind of did this, right? I was like, tell me about the, pipeline we have in this repo that, That can build out a deck for us, and then, okay, go leverage that pipeline to do it. Because sometimes it will… you know, escape the harness as well. But yeah, you can take a look at the whole, kind of, Open source repo to see what’s in there, and then we’re gonna obviously… In our boot camp, in a week and a half.

We go into how you build all of that from scratch for your use case, what our kind of methodology was. when we built this for ourselves. And then we also have a 5-week course where we spend the last week of that 5-week course purely on, presentation, and we kind of talk about not just the, the agents and skills building, but the kind of, like, frameworks and stuff. Behind them. I don’t know if that answered your question.

Hai Guan: And then the second part of the question was, how does Claude know where to pick up the analysis? So, the analysis was pre-done. Again, the AI analyst itself is a full end-to-end agentic system that goes from the very beginning of, you know, hey, given a question, how should we reframe it, how do we… what do we look for, and what is actually useful, that kind of stuff, all the way down to, hey, here’s the finished deck. And so, it has all the different wirings and, again, very similar to what Shane has said about, automatically triggering, invoking.

different commands and skills, same thing, like, if it’s like, oh, like, hey, I’ve done this analysis, and then I’m referring to that analysis, then it knows, kind of like, oh, okay, like, here’s what I’m looking at, and let me see what that looks like, and then, you know, it passes over and, like, continues with the pipeline. Hopefully that makes sense.

Shane Butler: Another question for the Saturday workshop, are there any prereqs, such as software tools that need to be installed beforehand to get started? So this is that Intro to Cloud Code Analytics. It’s a 3-hour workshop we’re doing this Saturday, May 16th. This is, like, like, I think we said this in a couple things in the chat, like, we, like, we find that, like. Anyone can really… leverage cloud code, for any sort of use case that they have, like, just no matter what their technical ability is, but there is some kind of initial friction. And getting things, set up, so that’s kind of what this whole Saturday thing is about.

It’s, like, getting folks set up hands-on with live instructors, and kind of, like, unblocking people so they can run through, like, the analytics repo, or start thinking about what they want to do on their own. So, that being said, there’s no prereqs. Our expectation is that you could have never used any of this before. You’ll need a computer, obviously. You’ll need a… you’ll need Wi-Fi to be able to access Zoom, and also to hit the Anthropic. API, and then you’ll need a Claude code Pro subscription, which is, $20 a month currently.

Hai Guan: best if you have a Mac, we are… we don’t know PC much, but we’ve had students who work… worked themselves out before as well.

Shane Butler: Yeah, we’ll spend some time on how you can translate the code to, from Mac to… Whatever, PC or Linux or whatever. Another question here… Or, Attendee, do you want to read your question, actually?

Attendee: Yeah, my question was just sort of around the repo that you have, and how I can take pieces of it and incorporate it with the repo that I’m already working with at my company, but I saw Attendee’s response and can probably just ask Claude how to do it for me.

Shane Butler: Yeah, I think it’s pretty cool how, how, portable Things are, so, Yeah, at my work, we all work on, like, separate kind of repos locally, and then kind of have, like, a master one. And we’re not kind of pushing to those… as we typically would with, like, in-git, because it just gets… things get so messy and different. So what I usually will do is, like, I’ll have Cloud Code be like, hey, go look at that repo, look at mine, tell me… everything that mine does better than that, and it does better than mine. Tell me all the gaps to fill, and now create a plan to basically, like, bring mine up to parity, or to add any… to fill any of those gaps.

Then you can review the plan yourself, and… Have it executed.

Attendee: Okay, awesome. Yeah, that’s super helpful, thank you.

Shane Butler: No problem, good question.

Hai Guan: I saw Attendee has a question about the token… or not token, the… how… if the $20 plan is, Is good enough. Yeah, so, short answer… oh, Shab, you already answered. Yes, should be… should be… should be more than enough.

Shane Butler: Press that ever workshop, yeah.

Hai Guan: Yeah, especially Claude, I don’t know if you guys have seen, Anthropic just announced that pricing war with, Codex, so they actually upped the limit for both the 5-hour window and also the weekly limit by 50% or something like that, or double, or something like that, so… it actually is quite abundant, if, if you kind of think about it that way, versus before, which is easily running out. So far, I think… I think now is probably the best time to kind of learn, because you can burn not… You know, like, not wastefully, but, like, you can burn compute in a way that you can really get into the learning without shelling out for the max plan, or, or things like that.

Shane Butler: Let’s see, we got another one here. Pipeline slide from before, the numbers weren’t vertically aligned. Is that limitations in MARP? Or wondering how good the slides can get.

Hai Guan: Pipeline slide from before, numbers were not vertically aligned. Oh, I see. Yeah, so the fifth step, if you remember, the iterate piece. So, you can actually give Claude the same feedback, like, you know, hey, I see that you aren’t doing this, or, like, the thing is out of whack, or, like, text boxes were spilling over the page, and things like that. It actually knows how to fix itself, or it knows how to review it, and then fix itself, so it is… By no means a limitation on any of the… any of this stuff. As long as you sort of work with it like a person would, it will figure out kind of, like, the stuff that you’re referring to, and that it will do the fixes for it.

Shane Butler: Yeah, and then, and then you just want to make sure that, whenever you do those fixes, you want to codify them. So, because they’ll default to, like, one-off fixing it. For that deck you’re working on? after you get to a place you want it to be, usually, like, kind of my next step is, like, working with Cloud, like, okay, what was the difference between what you were doing And this, this final step first that, you know, draft before that. How do we update, the existing, kind of, templates Or, skills or agents, or whatever, so that the next time, those are all aligned from the first place. And, like, every time you build a deck, there’ll probably be some little tweak like that that you have.

But it’s really cool because the next time you use it, it just gets better every, every single time.

Hai Guan: Yep. I know we’re out of time, but I can stay for a couple more minutes if people have more questions.

Shane Butler: Attendee on, like, Are folks using… What laptop are people using for cloud code? Is it recommended to get separate equipment? So I mean, like, Cloud Code itself, like, everything that’s running is running on Anthropic’s servers. It’s not running on your… machine. So… like… Your computer doesn’t matter that much in that case, but, like, if you start… if it’s starting to, like. create a bunch of output, on your machine. If it’s starting to, like, call other, like, you know, for us, like, we do a lot of, like.

Analysis works is calling different stats packages and stuff, or if you’re starting to run a lot of things in parallel, that’s when you want to kind of, like, think about graduating to a different machine. I don’t know, I… me, Hi, and Sharavi all run Macs. But I would say, like. you know, I don’t think you need to go shell out and buy a new computer right away, like, see what you can get away with right now, and then, yeah, it might be worth buying something.

Attendee: Very cool, thank you. I… after layoff, I’m at home, I’ve got my old Dell, I’m like, oh boy, can’t blow the dust off of this thing. Yeah.

Shane Butler: Blow the dust off. Try to remember the password.

Attendee: Yeah, yeah, I’m rolling in it, but, you know, even for security purposes, right, like, I’m just getting to the point where I want to start toying around with automating some of my personal stuff, and I’ll, you know, start with some you know, job search or business ideas, but now I’m going, alright, I know I need to do, like, make sure I’m doubled down on all my password stuff, but does it make sense to have a separate machine anyway? From what I’m reading, it sounds like it? And the Mac Mini is popular, but I don’t know, it’s a little… it’s a little iffy out there on the advice.

Hai Guan: Yeah, really depends on what you’re doing with it. I think, you know, the Mag Mini Crew is… Magmini is really powerful, by the way. I’ve been playing around with it for the last week and a half for, you know, really high bang for the buck. The whole, you know, separate computer thing, I think, pertains more to, for example, like, if you run something like an open claw or the Hermes agent, where, you know, like, it’s doing things directly itself.

Attendee: without you.

Hai Guan: you being in the loop, which could be pretty dangerous, if you think about it. So, that’s probably the context that you hear it from.

Attendee: Yeah.

Hai Guan: you know, for the biggest bang for the buck, it’s a Mac Mini is great regardless.

Attendee: Okay, cool. I’ll take a look into it, thank you. Yeah, I wasn’t looking at open club, but we’re just gonna slow our roll and just take it step by step.

Hai Guan: Yeah, open claw, in the grand scheme of things, is, like, level 5, versus, let’s say, like, a clot code is probably, like, a level 3 in terms of complexity.

Attendee: Awesome. Thank you.

Hai Guan: Okay, is there anything else?

Shane Butler: Any other questions? So you asked about the time for Intro to Cloud Code Analytics on Saturday. That’s 7 a.m. to 10 a.m. Pacific time. So, Saturday morning, West Coast.

Attendee: One last thing on the workshop. So, I have, an engagement that would cut me off about halfway through. Do you have these workshops frequently, or if I did it, and I was there for the first hour, hour and a half, and then is the rest recorded where I could just kind of catch up on it?

Shane Butler: Yeah, the rest will be recorded, and I think… That first hour, hour and a half is probably a good one to be there for, because that’s when we’re going to kind of, like. Demo the capability a bit, and then, like, set people up. And then you can follow along on your own leisure. We haven’t ran… so we have a bunch of… we have a couple two-day boot camps that we run monthly. We haven’t ran this Intro to Cloud Code Analytics one before. This is the first time we’re running it. But if, yeah, if it goes well, I’m sure we’ll run it again. But either way, you’ll get the recording.

Attendee: Okay.

Sravya Madipalli: We had… we generally want people to come in person and participate live. Most of the people did, but there were few people who did async for the bootcamp as well, and they… they still, like, you know, liked the work and were able to get a lot of value. Yeah.

Attendee: Great. Thank you. Yeah, I’ll see if I can move my other engagement a little bit, but… No.

Shane Butler: Sweet.

Hai Guan: Okay, if nothing else, thank you, everybody, for joining us.

Shane Butler: Yeah, we’ll see… we’ll see y’all in the Saturday workshop for those who are going to make it, and then hopefully see some of you in the, boot camp in a week and a half. If you got any questions. Feel free to reach out to us. on LinkedIn, or email us, or go to AIAnalystlab.ai if you have any questions. Happy to… Or get a hop on a call with you as well, if you have any questions around the boot camps or workshops or anything. Then we have our Slack community, too.

Hai Guan: Yeah. Oh, I forgot. Completely forgot about the Slack community.

Sravya Madipalli: Yeah, I shared it. I shared it. We can probably share it in the email as well, right?

Shane Butler: Yeah, well, I’ll send it… we’ll send an email out after this with all the kind of details to that stuff. But the Slack community is a nice way, because somebody gets to DM each other so much easier.

Hai Guan: Yeah. Cool. Alright. See you guys around.

Sravya Madipalli: Thank you, everyone.

Shane Butler: Thanks for staying. See ya.

Attendee: Thanks, guys.