Sravya Madipalli: Hello, everyone! Hello, hello? Good morning. Good afternoon, good evening, wherever you all are from.
Attendee: Good evening.
Hai Guan: Hey, everyone.
Sravya Madipalli: Hi, could you accept people? Are you a co-host with me as well, on this?
Hai Guan: Yes, we claim… let’s see… yeah, I can, I can admit, I can turn off the waiting room.
Sravya Madipalli: Awesome, that’s great, that’s great. Hello, everyone! Welcome to the free workshop today. Today, we are going to chat about, you know, Google Slides Decks. To kick off, could you please share where you all are joining from? That’d be great, like, which part, are you coming from? I’ll start with me.
Hai Guan: Yeah, maybe your role and where you’re from, put it in chat.
Sravya Madipalli: Yeah, that’d be great. Air joining from, and probably later, which role? Yeah.
Hai Guan: Got, Mumbai. Attendee, that’s, It feels very late for you. Thanks for joining.
Sravya Madipalli: Yeah, I see people from different parts of the world, that’s awesome. Cool?
Hai Guan: Bangalore, India…
Sravya Madipalli: Nice.
Hai Guan: Attendee’s from New York.
Sravya Madipalli: Great.
Hai Guan: Attendee’s from Lahore.
Sravya Madipalli: diversity.
Hai Guan: Wow, it’s very international. Love it. Chicago, Minnesota…
Sravya Madipalli: Would love to know which, for some of you, if you didn’t share yet, which part of, like, what’s your role, where you… Where you’re currently working, that’d be great too. Nice. Okay. Cool. I’ll start sharing my screen while you all keep sharing where you’re from. Give me one minute. Da-da-da… yup. Okay. Sure.
Hai Guan: Super diverse, that’s so cool.
Sravya Madipalli: Nice. Cool. Okay, before we get started, why don’t we do this? I’ll do a quick introduction. Hai is gonna do a quick introduction, and then we’ll get started. Hello everyone, I’m Sravya Madipalli. You might have seen my post on LinkedIn. If you don’t, please follow us. I’m a co-founder here of AI Analyst Lab. I have around 14, 15 years of experience now, starting with Microsoft. I started as a data scientist there at Microsoft. I was in Redmond, Seattle for a bunch of time, and then moved to the Bay. I worked at companies like eBay, later Nextdoor, and most recently, Grammarly. And when I was working at Nextdoor, that’s where I met with Hai and Shane.
Shane Butler, who is another co-founder along with us, he’s unable to join us today, he’s currently on a long road trip right now. So, Shane, Hai and I, we started, in doing a podcast a couple of years ago. We, if you didn’t check it out, please do check it out. It’s called Data Neighbor Podcast. We started off with a regular part, and then saw there is a big opportunity in the world of Agent Analytics and AI, and that’s when we decided, let’s do something about it, and we started AI Analyst Lab. We have an open source.
Please do check it out, I’ll share, Hai and I will share a bunch of links about a lot of free material and free depots we put out there, so that you all could benefit from, you know, a bunch of the stuff that we already made, for you. And we also do courses, paid courses on MeUN. So, yeah, that’s about me, and that’s about what we do, but I’ll give it over to Hai for a quick intro.
Hai Guan: Cool. Hey everyone, my name’s Hai. I currently lead data at a legal tech company. I’ve been in the data science and analytics space for roughly 20 years, and have been… have spent most of my time in the consumer technology space, so you might have heard of companies like Meta, Pinterest, LinkedIn, so spent time working at those places, and certainly met Shavi and Shane at Nextdoor. So, yeah, co-founder at AI Analyst Lab, where we provide education paid and free, materials to, help people up-level in AI and Agentic analytics.
Sravya Madipalli: Awesome. Okay, can you all see my screen? A thumbs up! Yup. If you do. Nice. Awesome. So, if you have any questions, please feel free to, you know, message in the chat. We generally try to have, a Q&A, a specific Q&A, by end of the session as well, okay? Cool. So, what is going to be today’s, you know, free workshop going to be around? It is going to be around building Google Slides from data with AI. So, let’s get started. Okay, so in the next hour, we’re going to basically talk through all these four different things, right? I want to basically talk through, first, what makes a deck good. What, the title test, what does each title mean?
Why do we need a specific, you know, framework for how do you want your, a good slide deck to look like? What are the four questions each and every single slide has to survive to get to a really good, insightful slide deck. deck that actually has some insights for the people who attempt, get, you know, they could take away from it. So we need to talk about that first, because, you might have seen this if you’ve worked with AI before, that without, a good grounding of what you want the AI to work with and do.
whatever output that AI receives is not going to be, you know, grounded in, like, the reality, or grounded in the things that you’d like to, that you expect from the AI to give you, right? So you need to be very, like, sure about the type of output you want, type of frameworks you want the AI to work with, and that’s what this particular, You know, this particular piece is going to talk to. And the next one is, we’re going to look through a real analysis. We basically use a repel. We have a free repo, we’ll share the free repo with you, which has a bunch of skills and agents that could literally run an end-to-end analysis for you.
We also have a bigger repo, which contains even more agents, even more skills, and we teach that, we share that as part of our paid courses, but we plan to make that also free at some point. Please look out for updates from us. But I currently used our giant reaper with a bunch of skills and agents to run an end-to-end analysis. And I want to show you what was the analysis that was built, and what was this giant Google Doc that it created, and a bunch of charts and, you know, document… like, insights that it created. But you wouldn’t be sharing that with your VP of product, for example, if you want to share the insights, right? You want to share a Google Slides deck with them.
So, we basically take that analysis, and we feed it into AI and ask it, hey, could you build a Google Slides deck for us based on this analysis that you have? And then… Cloud Code, AI here in this session is going to be Cloud Code, that’s what we’ll work with. So Cloud Code would help you create the Google Slides, and then we, you know, challenge it, we critique it, we ask for rating, reviewing itself, did it do a good job or not? If not, where was the mistake? Can it fix its own mistake and, you know, work with it, and then, we basically get the final output. So that’s the plan that we’ll work with. Sometimes, I’ve been getting a few API errors when I was trying to do this myself.
to do a dry run. Let’s see, I’m hoping, fingers crossed, that I don’t have any issues with Cloud Code today, or any API stuff. If I get into any of those errors, I have a bunch of output that I, you know, because I’m going to do a session, I did this early in the morning, I can share what an output would look like in case of that, okay? Cool. So, before we go into that, I want to have a quick poll here on… How many of you are absolutely new to Cloud Code, have not worked with Cloud Code? Maybe that’s a zero. And how many of you have worked with Cloud Code, but not with AI analytics? Like, anything with analytics related? That’s a 1.
And how many of you have worked with Cloud Code and also did some Agentic analytics work? Could you please type in the chat for me? Awesome. Thank you so much for, replying here. Okay. So, that’s… looks like a varied amount of people. I see a few zeros, there are some ones, and there are some twos. Nice. Because this is very important for me, so that, when I’m going to talk through a bunch of things, I’ll know whom, I’m addressing, and that’s the reason we keep asking this. Awesome. Cool. Thank you so much. Okay, so let’s get started. I hope I could get to finish all of this as well.
If not, maybe I will share with you the final… if I can’t get to the fourth step, of critiquing and fixing it, I’ll probably share with you, later over the Google Slides and stuff. Nice. Okay, let’s go into what makes a deck good. Right? What does a good slide deck look like? This is absolutely going to be tool agnostic, right? The part that you keep, the part that you reject, I’m sure whichever part of… you know, workflow, design workflow, like, engineering workflow you’re from. You could be from data, you could be from product, you could be an engineer yourself.
Anything that you’d like to share, you… you definitely… need to do this over a presentation, and get, your VP, or whoever is your leader that you’re presenting to, get them to agree or acknowledge the type of insight you’re saying, and get them to, you know, work on the recommendations that you have for them to go change and to go do something, right? Okay, I see a question here from Attendee. Would it work only with Google Slides? Just curious if it has been explicitly called out. If you could help us understand if this could be applied for PowerPoint as well. So, Attendee, great question.
So, today’s session is going to be with Google, because… okay, this is something that we teach in-depth, we probably have a session about it. And we’ll have a session later, so… as well. So, Claude Code is your software, is your tool that you work with. And then, all the software that it works with is via an MCP, or a command line interface, CLI, right? These are two ways that you could connect your Cloud Code. into a software, for example, Google Slides is today’s software that I’m connecting with, right? You could also probably connect it with PowerPoint, right?
So, the whole point here is about How… what is the framework you want your analysis to look like when you show… you… you want to show it as a deck, or as a presentation? Which software it uses to to actually present those findings is going to be dependent on if you have an MCP connection from Claude Code to that software, right? So today’s session is going to be with Google Slides, but all the frameworks that we use to come up with a good slide deck, how does a slide deck, like, what things do we want to have as part of, like, a conversation with a VP? Like, how do you want your, you know, things to look like?
All of that is going to be totally interchanged, like, you know, it’s just soft… one software versus the other. All of that’s going to be usable across multiple Slides. It’s purely the MCP connection and the CLI, if you have. That’s what needs to be changed to work with different software. And sometimes you probably need some, you know, helper files to help you, to work with the quirks of different software. But I would say that’s just a one-time effort that you do to get it done.
Most, of the, I would say bang, like, you know, the importance of the content today is going to be around the framework, and how, like, what does the skills and agents and the, and the harness, how does it look like to get you from an analysis to a deck? Okay? Cool. So, let’s go back now. We are basically right now discussing what is the framework of a good slide deck, how do you work with, like, when you have a nice analysis, when you know this is something that you’d like a broader… your broader team to know, or your leadership to know, what is something that you, you know, do and work with, right? So, the slow part here is actually never the thinking, right?
So, if So, let’s say you have a quick… you have a readout that you need to send, and where does your evening, before your readout, let’s say it’s tomorrow, where does your evening go into? It, at least for me, was a lot of exporting the church. Sorry, resizing the boxes, rewording the titles, oh my god, this particular font looks crazy, let me just make sure all fonts look the same. So your entire… instead of spending the time into deciding… spending more time, I’m sure we spend time into deciding what the story is, but… the entire time should ideally go into the storytelling, and choosing what number carries, and actually ensuring the numbers are valid, and is the story clear enough?
We want our entire attention to go into that, like, ideally in an ideal world. But that’s not what the reality is, at least before AI. You spend so much time in all of this formatting, and, you know, repasting, and, you know, all of that stuff. But guess what? With AI, today, we’ll show that with Cloud Code. All of this stuff that’s related to formatting, resizing boxes, and anything that could be handed over to AI just makes so much of your workflow of making a good deck so much more easier. And a quick inside tip, hi Shane and I, we use Claude Code for everything that we do, from free lightning lessons, that we do, like this one right now, to our course tags.
It’s literally, especially if you have the workflow, if you know what you want to present, it’s literally a 5-minute thing, or even lesser, for us to come up with a deck that actually follows our guidelines of how the colors of AI Analyst Lab, which is our company right now, and what is the presentation, how we want the presentation to look like, the storytelling. We have all of this inbuilt in our scales and agents, and it’s literally one button away for us To click this, to click through this, and, you know, go through, yeah. Attendee, so there’s a question from you around, so I suppose it is the same workflow with Gamma, attaching the MCP, etc. So, good question. Hai can chat with this more.
I personally used Gamma, not as much, but Hai and Shane used to use Gamma a lot before. And I remember, was it 2 months ago, Hai, or 3 months ago? We decided, like. Claude Code’s doing everything for us, and we don’t need to pay Garma anymore, and we just quit Gamma.
Hai Guan: Yeah, the honest opinion from me is, you… or at least for us, we really don’t need gamma anymore, because clock code is just much more flexible. We can instruct it to be so much more tailored and bespoke than a gamma, which is a lot more limited. So we don’t even pay for something that is less powerful, in our opinion. And yeah, like, I think that’s true of a lot of softwares out there, generally, and that’s what the whole SaaSpocalypse narrative has come in. I think it’s overstretched by a mile, but then there are certain pockets where the software are not You know, like, they just can’t beat horizontal large language models.
Sravya Madipalli: Yep. Thank you, Hai. We can share more details later, as well. Okay, so, this is something that we call it a title test, right? So if you read the left column, let’s say this is how your decks look like. Revenue overview, monthly trends, channel breakdown, AOE analysis, next steps. This is basically something that I would say, a genuine, not a gen… a general, like, a slide deck could look like. This is literally something that I’ve seen in my career, all my data centers and my team has done. But, if you read through the titles, does it tell you the story, right? in my left hand, in whatever is on the left side, I don’t think this tells me the story. What about the right?
Literally the same slide, with the same chart and the same information, but look through the titles, and if I read through the titles, like, if I read only the titles, top to bottom. That literally tells me the story. Revenue grew 78% in H2, but December growth nearly stopped. TikTok drove all of it. AOV fell in all 6 channels. Aoe’s average order value, it fell in all 6 channels. Protect TikTok, fix basket size. If you go through this, this talks through what exactly is the problem you’re dealing with, where did the problem start. And what exactly was the cause behind the problem, and what do we do about it?
The entire story and takeaway is literally in the titles, and we call this the title test. So, this is, like, one of the frameworks that we use to ensure that you have a good slide deck that actually speaks the entire story, from what’s the problem, what’s the insight, and now that we know the insight, what do we do about it? The recommendation, the most important part. And these are the four questions that you want every slide needs to answer, right? So, basically, the first question is the so what. Does the slide state a finding, or does it just show the data?
A chart with a topic title fails this one, because you want your slide to talk through and give you the actual finding, not just displaying the data, right? Now that this part is done. What about the second one? That is stakes. Would the audience actually understand why it matters to them? Did you convey that as part of this slide? That’s a very important one. So, starting with the so what, and then you do the stakes. Who is at stake? Why does… why? Why should people take the this particular insight seriously, right? And you need that as part of your deck. And the third one is basically the evidence.
Attendee: What is…
Sravya Madipalli: the evidence. What is the evidence that you have? Two or three, you know, supporting points? you don’t want 15 points, right? You don’t want to throw a bunch of things at a person because they get lost in all that information and go overwhelmed. So you need very, simple, not simple in the sense, or very, like, easy to understand and not get overwhelmed two or three supporting points, and that’s the evidence. You need that as part of your slide deck as well, slide. And then. You need the ask. Does it end in a decision request, or is it just, what do I do about it? Like a shrug, right?
So, is it, like, continue monitoring, there’s no ask behind it, or is there an ask, or is there a decision that you want to make, now that you know it? So, all these four things, you want your slide to answer, at least most of the slides, if not all of it, to answer the so what. And what’s at stake, and why does it matter to the audience? If they’re looking at the slide, why does it matter? Why are they looking at this slide? And then the evidence. What are the data, what’s the data that we want to point to? which talks about the insight, and then they ask, what is something that, they’ll do now that they know this insight? So these are the four questions we want every slide to answer, right?
So, you, we want you all to write these down, because in about 20 minutes, we are going to run through them over a real deck, but these are some things that, if you’re working with a deck, you want to, you want yourself to go through these questions and answer them for every slide that you have. Okay, so before we start with the deck, looks like I’ll run late today, let’s see. But before we start with the slide deck, I want to share with you an analysis that I have worked with our Cloud Core repo, and generated analysis in a Google Doc, and I’ll share with you how we got to that. Give me a moment, I have so many of these open. Okay, oh, I’m sharing back the same thing. Cool.
So… Okay, can you all see my screen? This is the Google Doc that we have, that we created based on, a run end-to-end analysis pipeline skill that we have. We basically have data called Nova Mart. We created this synthetic data to share as part of our workshops and share as part of our paid courses as well, where we work with multiple thousands, like. You know, thousands of data, tens of thousands of rows of information from, like, a fictional company called Novart, where it has 2024 data, and I basically had it do an end-to-end analysis to understand the performance review of the revenue of You know, 2024 for this company. And this is what it came up with.
It basically wrote when it created, like, you know, what, prepared by who, distribution, it gave, like, the ticket, this was a ticket, like, it just gave all this information in the subtitle, and this is the executive summary. It talks about revenue for the second half of 2024, it was total so-and-so, these are the scope and objectives, what was the data source and methodology, this is the top-line performance. It’s basically a dump of really good information.
I’m not saying this is, like, bad information, but if you look through it, you’ll understand it’s a dump of a lot of information that was shared with, probably this is something that your data person, a data scientist, data analyst, whoever will come up with. They’ll look through the entire thing, they’ll have a bunch of tables to share, they’ll have a bunch of charts to share, and a bunch of insights about these charts, right? And some notes about the methodology, of the quality of the data, what are some limitations, and look at this, the appendix. So this is what the first, run of the data would look like, but you would not want to share this with your VP, right?
You want to share a simplified version of this. with your VP, that contains the literal… the format that we just talked about. What does a good Google Slide, or any slide deck, should look like. So, we’ll take this. analysis, this Google Doc, with all this giant information, charts, and, you know, a bunch of text, and make it into a Google Slides Deck. That’ll be the demo today, but before that, I want to share with you something. this… how did we even come up with this analysis in the first place?
I know this is not the point of this workshop, but I wanted to quickly share with you what all, what all agents and skills were used to come up with this analysis at the first place, because I see that some of you, have not used Claude Code, for analytics work, right? So, let me share… where am I? Okay. So you’re seeing Claude Code, this is our repo. So Claude Code, we work with Cloud Code Terminal, but the software that I’m using, the actual, you know, IDE, that I’m using is Anti-Gravity. It’s a free, thing that you could download, anti-gravity, it’s from Google.
So I used that, and I’m sharing my screen of anti-gravity today, and I could go to anti-gravity terminal, and I could work… I installed Claude Code, and I could work with Claude Code from anti-gravity, right? So, I… this… is basically the ASCII diagram. I asked it, go through the… if I do the run analysis pipeline, if I go… if I do the run analysis pipeline, can you share with me the entire, set of skills and agents and your workflow that you do to get to a place where you come up with a Google Doc with the entire insights and learnings? I literally did not do anything.
All the charts, all the tables, and all the insights was entirely generated by Claude Code, and it could generate that because we created a harness, we created a set of skills and agents for it to… for it to work with, and on the data, the Nvomar data, the synthetic data that we created, and it used all the skills and agents and came up with the Google Doc. So this is the list of all the skills and agents that it uses… it used to create the Google Doc, right? So it basically… you know, understands the data, and it comes up with, this is the phase one, right?
It uses the question framing, what is the right questions that it needs to ask, because all you’ll give when you use the skill and agent from our repo, you’ll be like, hey, we have a question from our VP to understand how is the… what’s the performance of the business right now for 2020. 24. Can you create… can you run an end-to-end analysis to understand this? You literally give one sentence, and Claude Code runs through all the skills and agents to create that Google Doc for you. So, it uses a skill called question framing, analysis design. So, first, it creates the questions. What are the questions that we need to ask to the data to even create this You know, to do the entire analysis.
So it fixes those questions first with using the skill question framing, and then it, it basically goes through analysis design spec. Then it looks at the metrics, what metrics that it needs to use, what is the design of the analysis, and then it looks at what are the gaps, are there any experiments that are running in the same timeline that needs a brief of that? It also checks a data quality check, it does data profiling, it does an ESRM check if it’s an AB data set, and it looks through all these multiple, skills and agents in Tier 0, like, one after the other, right?
And after that, it basically verifies the information that all of these skills and agents have, you know, created and came up with. And then, it does the verification, and it looks at, did… did it go through, like, you know, this particular framework of business? Question is specific and question-oriented. And once it verifies it, it starts building the analysis.
It has, it goes through the skills of triangulation, it looks at Guard rail skill, ensuring that, hey, it’s not just looking at the main metric, but you’re also looking at guardrails, and it does some semantic validation, ensuring that what are the questions that are being asked, and what is the answer that’s being given, and how does it, like, are we looking at the right things or not? And then, we look through descriptive analytics, looking at multiple segments, different funnels, different drivers, all of that is part of our descriptive analytics skill. And then we have a root cause investigator scale.
We basically drills down, it drills down even more to find out, hey, the revenue is going down, the average order value is going down, what is the reason behind it? Let’s look, into… all the segment-level data that the Tier 2 created, what is the exact root cause? And then it validates the prior tire output, and then it comes up with opportunity sizing. Now that we see this error with this root cause, how much is the opportunity size if we go fix it?
It does all of this, and then finally does another verification and does storytelling using our story architect, and then finally does a narrative coherence review, or sometimes it says something at the top, but then it contradicts itself at the bottom. So, that’s the reason we have all of this stuff. So, these are… I mean, I don’t want to go through the whole thing, but these are a bunch of things that it goes through when you call a run analysis pipeline, an end-to-end analysis pipeline. Hai shared with you an open source repo. The AI Analyst open source repo also has most of this, other than some of the deck creators and stuff.
So, I would totally suggest for you to go through that repo, even if you don’t have cloud code, you don’t have it on your system. go through the repo, just read through. It’s all, English that you could read through and understand. Understand the skills, understand the agents, and that’ll basically help you know what is this framework and how does it build a real analysis end-to-end. Okay? So, this is the entire process, that it goes through to create the document that I shared with you. Okay? Let me stop sharing my Cloud code, and… Where am I? Yep. Okay, so this is the analysis that it, finally, you know, went through. Ein, it created this deck. It created this, you know, document.
But if you look through the document, we don’t have recommendations here. We only have a summary, right? And there’s no insight that’s popping up right away. But. we don’t want this type of story in our deck. We want our deck to the VP of Product that we’re going to demo now. to contain the insight, to contain the recommendations, to have all of the information that makes it easy to grok, right? Though this is a very rich document that contains all this information, we want our Google Slides Deck to actually have The insight. of the, like, the insight of why the performance is… is the performance lower to start with? And if it is lower, what is the reason behind it, right? So… yep.
Any questions so far? Hai? Anything we want to get through before I go to the next 10 minutes? I’ll probably go through the slide deck.
Hai Guan: I think you can keep going. Awesome. Yeah, I think, we’re just answering them as they go.
Sravya Madipalli: Awesome. Okay. So, basically, let me… da-da-da…
Hai Guan: Attendee has a question. Was the output of that workflow you showed, this document, Actually…
Sravya Madipalli: Yes, yes. So, the document that I shared with you is an output of the workflow. Some of… I didn’t go through the entire workflow, because I don’t need a deck, because I wanted to share the deck with you all, so I made it stop at a Google Doc. There’s a later end of the workflow that also creates the deck. Yeah. So, I made sure that I had some flaws in my pipeline, related to, just give me a document that contains a dump of information of the data using run analysis pipeline, and I could use that document to create a deck which contains the insight. This is… this was my prompt to Claude Code, and that’s what… I can share the prompt with you guys later as well.
Like, what was the prompt that I worked with Cloud Code to, and what was the fine, like, that got me the document? Because I wanted to document a certain way, because if you are part of data or worked with data orgs, you know that the first, the first output from the team is going to be a bunch of information with lots of slides and a lot of information, right? So I wanted to at least have some version of that, where it contains, like, a lot of information, and how does Cloud Code work through all of that to come up with the final, you know, output, like, the final insight, and actually grab that and show that as part of the deck. Okay? So, let me go back to Claude Code right now. Where am I?
Okay. Cool. So, since… I only have 10 minutes before we go to any questions. I want to share something that I kinda created for her. Okay. Oops. Where am I? Give me a minute. Hmm… Sorry, I’m working with a different monitor today, and it’s been a tough time working from this monitor. Okay, Cool. So, let me go back… Ein? Share with you guys, this. Where’s the Zoom? Okay, found it. Okay, cool. So, you’re basically seeing a demo that I was dry running myself before this meeting, because I was afraid of actually running through some cloud code errors in the meeting. So, this is something that I gave it, right? So, here is the finished analysis.
This is the Google Doc I gave, I was sharing with you all. So, I just told it, the audience is the VP of Growth who owns the channel budget. The meeting is 20 minutes. Do not make slides yet. First, give me a deck brief. the… I’m… so this is how we work with, right? Because we don’t want Cloud Code to do an end-to-end, you know, give me a slide deck and you have so many problems with the deck. You actually want to see… before it creates a slide deck, like, you know, what is it that it’s going to put in that slide deck, and do you have problems with it? Do you agree with it or not?
And once you agree with it, that’s when you go let it create a slide deck, because you’ll waste your tokens on something, on an output that you would probably hate. So, that’s the reason I tell it to not create the slides right now, but give me an arc in the, you know, four sentences. Like, what’s the situation? What’s the complication? Give me an answer and an action, right? And I’m also asking it to create a slide-by-slide plan. For each slide, what is the takeaway? Title, written as a full sentence, right? And what is the finding, and what’s a single chart that provides me that insight, right?
And, the title only read… lists just the titles in order, because If you remember, we had this framework where if you read all your titles, it should tell the story. So I give it all the details here, and then… It gives me the arc. It gives me the… so, it took me quite some time to get here. That’s the reason I’m sharing with you the output directly. I will also share with you in your giveaways on how do you work with Cloud.ai directly to, you know, for it to work with a document that you share with it, and, like, create an, like, an output like this one that I’m sharing right now.
So, it basically reads the full document, and it talks about there is a real complication, and it is not the one the document leads with. Look at that! So, it identified that the document, actually, that I shared with you does not talk about the actual insight right away. It… the document actually leads with just giving a performance overview, what is quarter to quarter, how are we doing, just like a plain In reading of data, but that’s not what we want. We want a good presentation to actually talk about the insight right away, right? So, that’s the arc that it finalized. The H2 2024 was strongest half, and the biggest month. Ever. And all 6 acquisition channels grew in absolute terms.
But this is the complication. Almost all that growth was brought with traffic. And it talks about what did TikTok give, and what was the marketing spend. And it gives the answer for the complication, that the AOV decline, AOV is average order value, and the decline is real, quantified, and probably not a TikTok mix effect. And this is literally… the entire arc of the slide deck that was not clear in our document, right? Our document was full of charts, random… not random, it was just reading the charts, and it didn’t have the insight right away, and Claude Code understood it. As soon as we gave it the doc. And it also gives the action.
Hold channel budget at current levels, close 3 specific data gaps, and decide at budget log. So… this is the entire arc that it created. And it also created a slide-by-slide plan. It tells that we need 6 slides. Look at the entire data. If we wanted the entire data in the document, in slides, we would probably get, what, 20? 20 slides or more, because it has so much information in those tables and all. But it created the plan that it created is that, no, we don’t need as many slides, right? We basically just need 6 slides, and what does each slide contain? Each slide came up with the exact title. Hedge 2 delivered recorded revenue, but the growth was bought, you know?
Hold FY25 channel budget until we can see spend. So, I would, If these are the titles for each slide, I basically would want the titles to be shorter than this, right? But this is its first pass. You could definitely work with Claude Code to make sure that, hey, your first pass looks great, but I don’t want the titles to be as long. And you could work with it to make your titles, you know, slightly shorter, because this, I think, it’s… Might be longer for a slight title. Anyway, so that’s the title that it gave, and this was a chart, and it also gave this support. And look at this, it also has speaker notes for you, what you should talk to when you present this slide.
And it comes up with the same thing for all 6 slides. So this is the first pass. And it also has a title-only read. Remember, I was telling you, if you read all the titles that you tell the story. So, it also gave me all the titles. And if you read all these titles, it actually tells the entire story of the data, what the data’s trying to say. Like I said, I think the titles are longer, and we don’t… we could work with Lod to, you know, to let that not happen, okay? Cool. So, after I got this, what I did was I actually told it to build the slide deck using a Python file that I created. I made this so that it’s faster for me to create a slide deck right away.
So I took all the code that it runs, and I put it in a Python file, so that Claude code doesn’t have to think on the fly, and it could use a Python file to come up with it. dick. So, let me share the deck with you that Claudecode created. - okay. So, it… This is basically the deck that Claudecolor created for me. No more, it’s your 2024 readout. If you look at the colors, these are the colors of Analyst Lab. Black and orange is what we use with grays and whites, and that’s the reason it used this. So, it basically came up with the chart, it has the title, and it created the, you know, the TikTok growth revenue, and the simple chart. And it came up with, like, all of these.
these little, you know, these little takeaway boxes, everything, all of this was created by AI. I don’t know if you see a difference between this particular slide deck and the slide decks you all might have created using Claude Code. The slide decks that we created without our repo. they are generally very busy. They have too much information, they are, they’re not simple enough, the… the charts that we come up with are not, you know, they have grid lines everywhere and all of that.
The reason why our slides look like this, even the first pass, I mean, I don’t like this first pass, I will probably work with, this deck to make it even better, but even… the first part that you look at, the reason why it looks like this is because of all the business rules and the detail that we added in our skills and agents. We… we have skills and agents for every single thing that you’re seeing here. For example, the chart. How does the chart what type of formatting we want the chart to look like. We used coal, strong… I forget the last name. There’s a data storytelling with data, book by Cole.
We use principles from them, and also principles that we think are the right, you know, ways to look at, like, data, what… how should the charts look like, how should the access look like. We basically used all that information and created skills and agents, and that… all that information is used to create this particular slide deck. So, let’s say we still have issues with the slide deck. I’ll try to go fast, and at least have 10 minutes for questions at the end. what did I do later? As soon as that is done, I basically work… oh, give me a minute… I’m sharing the wrong deck… Where am I? Yep. I have so many of these open. Okay.
Okay, so it not just created the deck for me, but it also created the data for me. I’ll share with you all the data as well, after this. But then, you could use the skill called Deck Critique that we have, or if you don’t have the skill created. there is also a prompt that you could use that basically contains all the information that we put in our skill called Deck Critique. You could literally type all of that in English, too, right? To make it reusable, you could put all that natural language into a skill and just call it using slash command. So, this is what I did.
I called Deck Critique, and I gave the deck that it created just now, and I used this… I put all of that in this JSON file, this deck spec, and I asked, like, hey, the audience is VP of Growth who owns, and it’s a 20-minute meeting. So, it basically used that critique, and it reviewed its own output. Like I said, I told you I have some problems with that deck that I just showed you, right? And then it created this entire deck, and it said, what’s the issue with this deck? It gave… it graded it overall grade C, and it only gave it 6.8 out of 12 for the 9 slides. And, it created, basically, what are the issues, right? It gave, basically, the three defects the verifier can’t catch.
It talks about the slide 7’s headline is false. AOV fell every month, it rose in… then it gives me the reason why the headline is false. So it caught the errors that you don’t have to go through and, you know, catch. I’m sure you’ll still find something that you don’t, agree with, or that the deck critic didn’t catch, but the deck critique that we built also goes and catches the errors that it… that the claud generated for you with its first pass. So, it talks about what are the slides. And it also gives us, based on the framework that we asked, remember the four things, the so what, the stakes, the evidence, and ask that I talked about at the start?
It looks at, does it contain all that information or not? And it rates Each slide, based on how good you know, each slide is. It rates it based on if… on 3, if it’s a 3, 1, or 0, and it rates each slide into that. And it comes up with an anti-parent summary, and it talks about what is… what are critical things, and what are some, you know, high-impact things, like, if you fix them, it’s going to, like, totally help you. Okay? So, I was working with it to come up with, basically, a plan on Sharing that, the final deck as well, it didn’t complete yet for me, it’s still running, and I had an issue with Cloud Code at that moment.
I’ll share all of that in an email with you all, because we are running late, I don’t want to spend time on that right now. Okay, so let me stop sharing and go back fast to the deck. Okay. Oh, man, am I… Okay. Cool. So… this is something that it also created for me, guys. It also created each, like, a Excel Google Sheet, which contains the data and the graphs. That it used. So that it makes it… it helps us understand what exact data, when to do it, and if you want it to be changed, we can change it. Okay, let’s go back to the slide deck, that I was sharing. Okay, so you all… I all shared this with you all, we turned it into a deck, then we graded it, and then we’ll fix it.
I’ll share with you how the fix looked like, but then this is what the critique gave me back. It gave me a scorecard of Basically, what was wrong, why it was wrong, most importantly, and then it talked about the… it talks about if the story or the insight is buried or not. Okay? And these are all the things we use, the skills and agents we use, to come up with the deck, to critique the deck, and also to come up with the, you know, the different charts and stuff at the end.
So deck critique, deck rescue, the slide transform, there’s a presentation doctor skill that we use to ensure that it graded all your slides in ABC, like, does it fail it, does it, you know, pass it, and what is the number of marks it gave? For each slide. Okay. So, before we go to the questions, I just want to, you know, quickly share with you what I… there is a giveaway, there’s a bunch of things that I’ll share with you. What does the Starter Slides template will look like, and Google Slides prompt pack that you could use, and what are the four questions on each page that I was talking about. I’ll share all of this with you so that you could use it, you know, like, literally.
Starting in your next slide deck that you build. Okay, quickly, I want to share that we have two courses that we share, along with a bunch of free stuff. You should definitely check out our free stuff, the free depot, and all the free workshops that we are having almost every Wednesday. You should definitely join us. And if you want to learn more about what we are sharing, we have two courses where we teach all of this. This… we have something called AI Analytics for Everyone, where we literally teach, I would… kind of call it, like, master’s… micro-masters in product data science of Silicon Valley. We share all the frameworks, the entire product workflows, how do we build metrics.
How to ask… how to ask right questions? How do we run experiments? The causal analysis, like the diff and diff, and multiple causal analysis. And how all of that that we do as part of data science workflows in tech, in Silicon Valley, and how do you do all of that with AI, with Cloud Code. And we give you a repo that you could work with, and that’s… this is the course that’s literally starting next week. So, if you have any questions about this, please ping us. We have, a coupon for you, like, a discount code for you, called Think20 as well. And, if you want to learn how to build a system, to build the repo, like how… the repo that I shared with you.
the Reaper with a bunch of agents and skills, and understand what went into each… the repo, and if you want to build that and take a repo like that to your work, and create something like that for your work, you should definitely join us in our, like, entire end-to-end course on Agent Analytics, where you build an AI analyst that you can trust. Not just building it, but also making it as deterministic as you can, making sure that the output that you have is validated. The context is built into it, and you create systems for your work that can be shared with your stakeholders and stuff. So, yeah. This is what we have. I’ll stop sharing and take any questions.
I could stay slightly over as well, if there are more questions.
Hai Guan: There was a… Let’s see… It’s a wizard.
Sravya Madipalli: Any questions, you could also, you know, raise your hands and…
Hai Guan: Yeah, just pop in.
Sravya Madipalli: as well.
Hai Guan: And then you can… you can just ask.
Attendee: If over…
Hai Guan: Yes, Attendee, go for it.
Attendee: Hey, thank you for… thank you for walking through this, this is very helpful. I just have a quick question, which I posted in the chat as well. I saw that you actually had a whole set of agents and skills to generate the Google Doc, and then from the Google Doc, the deck. Is there a reason why you will have to follow the two-step process? Can’t you just… have the prompts which you… a combination of prompts is the skills and agents to directly generate an executive. Readout, like what you did for the deck, versus the two-step process.
Sravya Madipalli: So, Attendee, that’s literally what, what we do when I created a deck for this lightning lesson as well, right? We do a single prompt to use all the existing, you know, information and create a deck for us. But for showing you the actual workflow, let’s say you’re looking at the data for the first time, right? You want to know all the different steps, every detail of the data, right? You want to understand the data yourself. you want to be the person that hand-holds every information, because it’s your first time running it, right? So in that case, how do you go through? We can actually get even more granular, right?
Where we try to do that as part of our paid course, where we actually, literally spend every week on each of those steps in the workflow, because if you don’t get your questions right to start with, then the entire effort that that you worked with, all the tokens that you spend in coming up with the deck, all of that’s going to be a waste, right? So, ensuring that you have right questions to start with, and then ensuring that it’s picking the right methods, based on the context of that question, right?
Because you might have built an entire harness, like, the entire set of skills and agents for A few types of situations, but let’s say you have an entirely new question that comes with an entirely different context than… that you didn’t build in your system yet, then you want to know how Cloud Code reacts to it, to that particular situation. That’s the reason you stop at every step, and see that you agree with those steps, and then.
Attendee: work.
Sravya Madipalli: you agree with each of those steps, that’s when, I would say, in the next iteration, you’ll probably do less of that checking in every step. You’ll probably combine a few steps together, and then in the third, fourth, or probably fifth and sixth iteration, you’ll just do an end-to-end thing. One prompt in and a deck out put out, right? But… you do that when you’re very sure about every single output that each of those phases have given you, right? That’s literally what I was trying to share with you.
Attendee: That makes sense. No, thank you. Appreciate it.
Sravya Madipalli: Yeah. I think, Attendee, I see a question. Do you do a PII check on the deck? Do you mean that if the deck contains any of, like, security information, that it doesn’t give an email out, probably, of a user, right? Is that what you mean, Attendee?
Hai Guan: He’s here. Oh, yeah.
Sravya Madipalli: I think he just replied with a yes. Yeah, so, yeah, so we basically… so this is all the rules and, you know, the type of, like, logic you build in your skills and system already. So, this is something, because we are working with the synthetic data right now, that’s not something that we, that we dealt with in this, but… it’s literally one… one important Claude rule that you could put. Like, when I was working with sensitive data at my workplace, we didn’t even have that, information open to Claude code. So we had multiple classes, of, you know, like, permissions that you need to access to even get PII data, right?
So, if… in your work, I don’t know, like, if… wherever you’re working, where I was working at, we didn’t even open that class of information, the financial class, of information to Claude Code to most of our data scientists. We opened it just for a few people that we know are very sure about what they’re working with. Especially, like, writing into database, we didn’t give that permission to anyone if they’re working through cloud code on that. So, these are something that could be held, way before even Cloud Code workflows comes through.
But for example, let’s say those are not there, you could definitely and easily build that, lever into your cloud.md file as a must-pass rule, that you never show this information, you never show this Like, we do have rules like that in Cloud.md. You could check them out in our free repo as well. Look through Cloud.md, and look through what’s actually part of it, and it contains, like, you are an AI analyst, you know? You never, show, like, things that don’t triangulate with each other and all of that. That could definitely be a rule that you add. Okay. Any… any other questions about today’s workshop? Any questions about the future of free workshops?
Or any questions about the paid courses that we… that are coming up? Would… would love to answer any questions.
Hai Guan: Yeah, let me share what’s coming up. We have, as Sravya said, we have Lightning Lessons coming up weekly, and so, yeah, we have a series that you can subscribe to as well. So I’ll share it in the chat in just a little bit. And then if anyone wants to join us next week on a 5-week journey to… to, to up-level on, analytical judgment and, and all of that, you know, certainly do join us for the AI Analytics for Everyone course. That starts next Monday.
Sravya Madipalli: Yeah. Awesome, Attendee. Great that you’re subscribed to the Lightning Lesson series. If you’re… if… so, you could also subscribe to the entire Lightning Lesson, free Lightning Lesson series. You will get all the free resources and, you know, the… that we share and the recordings as well. So, make sure you join that too.
Hai Guan: Trying to find a link.
Sravya Madipalli: Awesome. So, if you also have any questions related to the courses, the paid courses that are coming up, please reach out to any of us, me, Hai, or Shane, and you could join our Slack channel. Hai, did we get a chance to share? If not, I have a link I could share to join the Slack channel.
Hai Guan: Not yet.
Sravya Madipalli: Awesome. I can share the link, give me a minute. So… join our Slack channel. Oh, looks like. Hai and I share at the same time. So, join the Slack channel, we’ll keep sharing, everything that we’re doing in this space, a combination of free, giveaways and also paid stuff. Please share with, if you find this lesson valuable, and if you think others in your team at your company would benefit from it, or your friends, please share with them. We just want to make, all the analytics for as Agentic as possible, so… Yeah.
Hai Guan: Alright, I just shared the list for folks who want to subscribe to it. That’s, all the upcoming Lightning Lessons, free, every Wednesday. We have a few that’s coming up that would be really interesting. I think next week’s, we’re going to talk about context. Management, and then the week after, we’re actually gonna… in partnership with, Kimi, you might have heard it from Moonshot AI around, just introducing what open source models are, and what that means in the whole AI space, so that should be really, really interesting. And then so on and so forth. We have a ton of Going from introductory, like Bot Co-Work 101, to, to more advanced stuff. So, yeah, stay tuned.
Hope to, hope to see you all soon.
Sravya Madipalli: Awesome. Okay, thank you so much for joining, everyone. See ya.