AI Data Visualization in Claude Code

0:00 Uh so thanks for being here. I'm Mich. I'm one of the folks along with Hy and Stravia who ran AI analyst lab. Uh we teach courses on Maven about using AI for data and analytics doing AI evals. Um yeah as I mentioned this going to be a little different today than normal lightning lesson. We're not going to have any slides. We're going to use the full hour to just be live in cloud code. I'm going to be building some data visualization analyses from scratch in real time. Uh so we'll see what happens like like obviously uh LM are nondeterministic so hopefully nothing goes too uh haywire but I think it's kind of actually good like when stuff uh breaks and doesn't work it's like a very honest representation of like how these tools will react in your actual day-to-day and then we can kind of like work through that. But I kind of want to want you to see like what our workflow is fit. So before we before we get started though, um let's drop some stuff in the chat just so I can see kind of where everyone is. So in the chat, if you've never used cloud code before, like maybe you heard of it, but if you've never used it before, uh drop a one. If you have used cloud code, but you've used it for like coding and general stuff, use a two. And if you've used cloud code for data analysis or visualization, drop a three.

1:21 And let's see what's our what's our spread looking like high >> very even one and twos and threes I would say >> nice okay >> actually more ones and twos >> more ones okay dude that's cool if you're a one like uh no worries you're going to be able to follow along uh as well I think it'll good like intro if you're two or three. Um, if you want, you can actually open claude code right now and follow along from the same data set. I'll drop a link to the data set in the chat. Let me actually find that. Uh, I'm in another terminal asking Claude to give me the link to the data set to share.

2:12 >> I can drop it. Is it AI analyst? >> Uh, no. I'm going to drop uh an actual like a Spotify data set like in case people want to kind of like mess around while we're doing this too. Here you go. So, there's this Kaggle data set I'm going to work through today. If you want to try it yourself while we go through since we do are here for an hour, like download that, open up Cloud Code. Uh if you don't know how to use Cloud Code, if you're like in that one category, that's fine. And we're actually going to have a boot camp on this in nine 10 days where we're going to like set people up from scratch who've never even opened up cloud code before. Get them all the way to the point where they can build like an agentic system. Uh yeah, I mean you can follow along with VS Code and Windsor. It's going to be so I'm not going to build out. Um I mean that'd be interesting actually because I'd like to hear how it like goes differently for you. So, one of the things with cloud code you'll see today is we're going to try it with like vanilla cloud code and then like this agentic system we've built. Um, and you'll see how the visualizations change. So, it be actually be interesting to see how like even like when surfer vs or VS Code or cursor is different and yes, this is all recorded, Derek.

3:33 Okay. So, I'm going to keep going, but I'm going to pause a bunch on the way and we'll we'll we'll pause for questions and stuff. So keep dropping them in in there and and Hi and Shravia can answer but we can also answer them later. So all right. So what I just put in the chat that is the Spotify most stream song data set. That's like 950 songs uh with stuff like total streams, artist, release year, uh danceability, beats per minute, energy. It's a pretty fun data set and like everyone kind of knows Spotify so it's like nice like I don't know I feel like it's very general it's like not a bespoke to someone's hobby or interest.

4:16 Um so here's the scenario that we're going to kind of like pretend to walk through today as we use cloud code. Um so say we work at a music streaming company and our head of content walks over and says like hey we have a budget to license music for the next quarter. What kind of music should we be licensing? Like what has staying power? That's our question we're gonna try and work through for the next hour. So right here, you can see on my main screen over here is what I'm sharing. There's uh two terminals. So on the left side here, you can see there's a bunch more stuff on in the directory that I'm in here.

4:59 Um, >> so on the left side is is uh kind of like a a system we built with visualization standards, chart helpers and design and reviews baked in. This is as actually a open-sourced repo we created called the AI analyst. Um, yeah, I don't know maybe Stravia or hi you could drop that repo in the chat. So if you want if anyone on this in this group wants to like clone that repo now or later on and mess with it, they can.

5:31 Uh on the other side here, this other terminal I have on the right, um this is just an empty directory. It has that one CSV in it. We can even open it up. It's just a CSV of that uh Spotify data. And so this is like your kind of like vanilla cloud code. like there's no additional system built around it which it's like what's coming from Anthropic and it's also you know it's extremely good but there are like ways to make it even better and like um specific to your use case when you build a system. So today we're just going to experiment a little bit and see how for the first part we'll look at like how just like no agentic system versus a genic system will lead you to different results. So, I'm going to be uh adding pretty much the same prompts in these that I have copied on this other screen. So, here in our vanilla cloud code, I'm going to say, you know, load me the Spotify CSV uh and tell me like about that data set.

6:39 In our regular one, I'm actually going to tell it to create me a data uh a new folder called analysis for everything that we produce today because there's so much stuff in here. It's very easy to navigate when the directory is empty. But then I'm going to say the same thing. This Spotify data set give me a quick sum summary. How many rows columns do we have? So we start getting output here in the right um column here in the vanilla cloud code. So yeah, 953 uh rows. Uh we have some track name, artist name, artist count, release year, release month, release day. Um some platform presence stuff like is it in a charge, how many streams it has. Uh is it an Apple charts? Some of the audio features like the beats per minute mode, dancability, energy.

7:28 Um yeah, so kind of like just high level. Basically the same thing here in our other cloud code terminal that has a gentic system. You know I think that's the only difference here is that they created a folder called analysis for us to put stuff in. Um yeah so at the end so I'm not going to paste my prompts in the chat because they'll just take well I guess if you're following along yeah I should be pasting those. I will also share.

8:00 Yeah, good call out. Sigh. I will also share at the end of this. I'm gonna have all the prompts in a doc and I'll email those out to everyone. But yeah, I'll I'll put paste them in as we go, too. Um, okay. So, where were we? Yeah, both loaded the same 950 songs. Let's make some charts. I'm going to ask the same questions to both Windows and we shall see what comes out. Uh where is my question here?

8:33 All right, same question. Both terminals. Let me drop it in the chat for you. Boom. There you go in the chat. And we're going to drop these into both. I'm gonna drop it into the Aentic R system first and then I'll put it into vanilla here. So my question was which artists have the most total streams? Visualize the top 10. Save the chart in some folder. And so you can see here like um I'm not telling it what chart to make. not telling it any guidance really beyond like, hey, here's a question that I have. Like this is kind of like a a question someone would probably really have, right? Like people don't necessarily be like, hey, uh, you know, give me a bar chart of XYZ. They have more like a question they want to solve.

9:33 All right, so looks like it has completed an AI analyst. Let's find out. There it is. All right. The weekend Dominic's total streams on Spotify. The weekend Bad Bunny. Let's see what we get in the other one. Okay, so like off the bat, you can see like they're like same data here. It all ties out pretty similar. Um, one thing you'll notice is it is slightly different in terms of like the the AI analysts one has a uh it says like in the in the header here like what's the actual takeaway.

10:24 There's a little like cleaner in terms of like not having these added uh added framing around it. In this case, I actually wanted something better out of the AI analyst. And so, let me show you a little bit more about like what the AI analyst is supposed to do because it it actually did not follow the workflow I wanted to. So, um I'm going to tell it, hey, you didn't follow the chart skills and agents workflow. Can you show me what that's supposed to do in an ASKI diagram?

11:03 All right. And so sometimes I find when you build like these kind of agentic like harnesses around Claude uh if you don't orchestrate it like there's just like some rammedness to like is it going to follow the orchestrator properly or is it kind of like gonna go off on its own. So even though it is like a slightly nicer chart than what vanilla call code I wanted to make be much better. So right now what it's going to do for me is it's going to go and read through my repo here, check out all of the agent files and the scale files and give me an accurate picture of like what it should have done. And this is something I do like pretty often when I'm kind of priming cloud code with my whatever system I built to make sure like kind of like double confirm that it's going to follow those steps along the way. So here like what it should have done is make a chart of X. It should have then loaded visualization pattern skills that learn from a bunch of the themes and chart types and storytelling with data rules that we have in those skills. It should have kicked off a chart maker agent that has a bunch of steps around like loading and validating data and um generating uh decluttering the data like removing spines, grid lines. It did some of this, right? Um, legends, if it's cool down here, it actually like tells you what it skipped. And then it goes through some validation and review periods with like um visual design critic agent like which will say if you have more fixes that you need to go through and do again if it needs revision or it's approved. So it's kind of like a score at the end. So it knows that it skipped, right? So it says it didn't call the storytelling with data style from it chart helpers. It didn't use the highlight bar function. Um, I have like a bunch of themes in here around like a minimalist theme, a New York Times themes, an economist themes, a corporate theme. It didn't use any of those. Um, I like it to have like kind of like off-white backgrounds because it's not as jarring. Didn't do that. So, yes. Want me to redo the top chart following the workflow properly? Yes, please.

13:28 Um, in these cases, I'm not going to do it today because I don't want it to waste our time with it writing a lot of code. But when you have these uh systems that like kind of like deviate from their or like escape their harness, then what I would probably do actually in reality if we weren't doing a demo is I'd say, "Hey, can you tell me why you didn't follow the workflow as expected and can you tell me where the gaps are in the like existing code that caused that? And then can you fill those gaps?

14:00 All right. So, let's redoing the the chart. While it does that, I am going to pull another prompt out. Oh, yeah. See, this just looks better to me. So, it's like now you have kind of like this one bar that's highlighted. The rest are grayed out. The title tells you the the takeaway. There's like more direct labels. I just feel like they're like the right size. They're not they're not cluttered and like the data between these two is is identical, right? The weekend Taylor Swift assurance all around 14 billion streams, but there's a story here like has the same screen as Taylor. I don't know. We'll we'll get into it later. So like this is the kind of the difference um between where you get like just going with like raw cloud code versus like trying to make it customized to whatever like your style is the use case um you're working with is the visualization skill is in our uh repo that is open sourced so that maybe if there's a link already shared hire Shavia probably above Okay. So, where were we? All right.

15:22 So, next one. We need to know what's trending. Like, I want to know not just who is at the top right now, but I want to know like how's popular music been changing. So, I'm going to ask this question. How have the audio characteristics of popular music changed from 2010 to 2023? Look at danceability, energy, and balance over time. Only years with at least five songs. Visualize it. Save it as a PNG.

16:02 And they're off. Okay. Okay. So, we have our audio trends over here and we got them over here. Okay. So, left side here obviously like this is kind of like again like your I mean it's pretty impressive that it it made this. Um but you can see there's like definitely some differences here. Um so left side you have like these three colored lines. you have markers on every point. Um, kind of like got the grid lines everywhere. I'd say like what I do like about what it did where it was which is a little annoying with what the Gentic system did that I might fix is like it kept it at least to on this 0 to 100 scale. I don't know that I would actually have it at this 30 to 80% scale that the Gentic system did. I might want it at like I don't know. I'd probably still want it at a zero to 80, but I get what I was trying to do. I was trying to zoom into what's going on so you could see like the trends. But the main takeaway here is on the right like you don't really know necessarily like what to uh look at here. Whereas on the left you have like one line that's highlighted.

17:21 The other two are there but they're grayed out. So you're go your eye goes like straight to the story and the title tells you um what happened. Um so you don't have to figure it out by yourself. So in this case it says like popular music is getting sadder. Veilance dropped 18 points since 2014. Um, and it also adds, you know, I like that it adds uh some stuff around like, hey, what actually is this graph is like what's the data set it pulls from.

17:56 There is another thing here that I do kind of wish I had highlighted though about it being newer music being more dancable. Um but the cool thing is like you know I didn't tell it which feature to highlight. I asked a question the system decided that this is in this case surveillance is like the most interesting line. Um when you run this each time it might change. I actually ran this yesterday in the story it was telling me was more around like this big jump in dance ability but it's actually trying to find a story for you. you have to work with it and tune it to find the right story. But um I find it really like I wouldn't necessarily just like spit out a chart, send it to a stakeholder, but I find this like really nice for like EDA to be like hey find the stories that are in this data and then like present those visually to me maybe in multiple charts rather than just like give me the chart and I have to find the story immediately. So it's kind of like a nice partner in that sense.

19:02 I don't know. What do you guys think? How any any first reaction in terms of like how these two charts compare? Like what are the differences between the two of them while I ceue up my next prompt chats? night and day.

19:33 Okay. And like Yeah, I guess like the whole thing with this too, Nathan, is like you can like um you know, I haven't really worked on this version of our repo for like three weeks. But like at like my day job, I spend a lot more time like tuning these to like what I know my stakeholders are looking for and even like the brand that we have. But um it's like pretty cool like you I mean we could just do it right now actually like hey is there a here I got to like say it instead story around dancability instead of veilance could you reframe the chart to emphasize that okay so just like we'll see what happens but I'm pretty sure it's just going to gray out this veilance one and then highlight the dancability one.

20:28 Um, that's kind of the other thing too, right? It's like even if it doesn't get like exact story right away that you want to tell or that you think is the most important, it's really easy to switch around. So, there you go. 2015, there's a dip in dance ability. What a shame. All right, let's go for another prompt. It can make the whole thing Mia at once. You can definitely oneshot like you could oneshot a deck, oneshot an entire document. You can make a whole dashboard kind of thing at once in HTML. You could have it output a bunch of PGs. Um, but uh yeah, kind of depends what you want to do. I've actually lately like like going more like one by one because I kind of want to interact with it and also validate some things. But actually what what I did do was pretty cool at my job the other day. I went one by one building out this analysis. put it all had it autoput everything into a document for me and then I met with my team and I had like screwed up and I added like I didn't add some filter to the data basically and they're like oh can you do this again um but like not with that data and like it wasn't a dashboard right it was like a document with all these like specific images in it but because I did it all in cloud code and I had cloud code like like journal the process we had gone through. I was able to be like, "Hey, can you rerun everything we ran but like apply this filter and like refresh refresh everything and it took like minutes whereas I don't know." So, so that was nice. Like I was comfortable doing a one shot there because I'd already gone step by step.

22:20 Okay, let me I think I already posted this in there, but I'll put it in there again. Here's the next prompt. Okay. So, do more dable songs get more streams? I want to understand the relationship between danceability and total streams. Call out any songs that are surprising. High streams below danceability. Visualize it and save it as a PNG. Yeah. So, this will be our last kind of comparison here, but I kind of like Yeah. want to understand like what matters for a licensing decision. If we think back to like our original question from the head of content who like has some budget for licensing like like danceability like has increased a bunch was like does that actually matter? So we'll see.

23:11 You think music and you think it's going to be because people want to dance, but okay. So, we have over here we have one over here. So, yes, this is kind of interesting. So, on the right side, um these are actually pretty close in terms of like what they did. Like I'm I'm glad like when I ran this yesterday actually it's kind of different every time you run it all these dots were green and it kind of had the labels on it but like it had made it look like every single thing was the same. So I do like how it did assign this cluster with the red dots even just in like the kind of uh vanilla version.

23:59 Um yeah, I mean the labels are overlapping a little bit for sure since it's like basic map plot. Um here on the left, uh yeah, most dots are gray. This is like kind of like the one of like the main things if you're doing data storytelling is like figure out the story like highlight that and just gray everything else out. That doesn't matter. Like so you don't have to focus on that. And it's really easy to do that when you just like have that as a directive in a skill. So most are great.

24:36 The outliers like really pop here. Um what do we have? And it puts okay it has like it has pretty much the similar kind of like information in each of them like the percent densibility and and the number of of streams. Um, and then the other thing we see just even in the terminal is that both of these are are outputting some uh takeaways. So I think they've kind of we'll read them, right? So on vanilla it says the takeaways many of the biggest hits in this data set are emotional ballads, rock anthems and moody tracks, not dance floor bangers.

25:20 Um, emotional resonances seem to be what drives repeat lessons at scale. And the key finding over here that kind of bulleted out. So I kind of do like that like this like it it gives like the high level and then goes into detail and it actually looks like it did some additional analysis here around correlation. So dancability doesn't predict streams. The correlation is essentially zero. R equals.11. This Pearson score. If anything that's a slight negative relationship. Less danceable songs do marginally better. So you can see it's like going the extra kind of uh it's not just like visualizing and saying the story. It's now backing it up by with some statistics. So we have a bunch of like you know we have dozen agents in here and a few dozen skills. I think those are in the doc cloud here and a lot of these are getting into like some like driver analysis, statistical analysis.

26:19 This is probably pulling from that a bit here. Yeah, takeaway is like the the same, right? Making a song more dable won't make it popular. Emotional resonance seems to matter more than rhythm for mega hits. Okay, so where are we at time? We're about 30 minutes through. Oh, I'm right on my mark. That's pretty good. Um, so that's all to say like same cloud code, same exact same model, exact same data, exact same questions. The only difference here is there there's definitely there's always some difference in terms of randomness because we're dealing with non-deterministic systems. But you can see like the harness that we have put around even just these data visualization skills um like creates totally kind of different graphs than you would just with the vanilla stuff. And um this kind of goes for everything like we're talking about data viz but anything you want to do like creating agents and scales and python helpers and having them in there and having an orchestrator that knows when to uh implement each of those is going to make the way you use cloud code especially for uh data analysis way more helpful. Um, all right. So, we're going to continue with our analysis, but we are just going to do it on the uh AI analyst now. We don't need to be doing it.

27:47 >> Hey, Sean, one question probably in a lot of people's mind is how do you validate the numbers? >> Oh, yeah. Okay. So, let me see if I did this in here actually. Nice. That's a really good question because I added it in this prompt. So we will test this a little bit. Okay. So for fullon understanding of how to validate numbers like you probably want to build a evaluation system in our boot camp. We're going to spend a lot a whole like hour and a half just on validation of figures like kind of the best practices here. This session is more just around the capability of data visualization. the TLDDR. Uh my philosophy around leveraging AI or cloud code for any sort of data analysis is that the human is accountable. So no one gets to run the analysis and send it to a stakeholder and and it and it screws up and be like, "Oh, AI screwed up."

28:45 It's like, no, like I screwed up because I didn't validate. Just like if I ran some SQL query or ran a dashboard or anything else like non AI related and I messed something up like I'm accountable for that like things can be automated and still get messed up right there could be bugs in your SQL query there could be bugs in pipelines there could be bugs in dashboard so still accountable for that I think the you know the easiest way is uh when you're building your system run it on a bunch of analysis you had already done in the past where you know the numbers like you can actually like even look at the old analysis. Um and then you can confirm those side by side to see if they tie out. If they don't tie out, you can basically add more guardrails to your system and nurture it so you get like the guardrails around it to tie out.

29:40 Sometimes they don't tie out because like you screwed up in the past. That's happened to me a few times where I've ran it on analysis I've done before and I actually found AI found errors that I had had. Um, so that's back testing, right? You can actually do you can build AI evaluation system that does a ton of back testing for you and back test not only your work but the work of your entire company. That's like a whole another field of work.

30:06 Um, and then the other thing I do, which we're going to try and do today, I have not tested this out in Google Docs yet, but we'll try it at the end of this session if we have time, is every time I'm using cloud to do SQL or do analysis, I have it save all of the code it ran. So, never do I have it like create numbers. The LM never creates numbers. the LLM, the agent um runs Python scripts or SQL to create numbers.

30:40 So all the number creation is in code. That's how you get rid of hallucination. You still have incorrect numbers if it runs the wrong code. So I like to have it output that code along with each chart and then I can kind of run through it and eyeball it. I can run it elsewhere into like manually somewhere else. I really like it in we're going to try it in Google. I really like it in notion. It has a really nice MCP where I'll have like a chart and then right under the chart it'll have a toggle that says like code and logic for creating this chart. You can toggle down and in the notion doc it'll have all the SQL right there. Um it just makes it a lot faster for reviewing. Um, yeah, lots of other stuff for validation. You could get into, you know, building out golden data sets and kind of like tiered different data sets you trust, but that would be the whatever five minute spiel.

31:40 Uh, for the full validation kind of work, like we'll cover that in our boot camp next weekend. Okay. How did this do? All right. I created another chart. What did I even ask? Ah, yes. All right. Our head of content wants to know what kind of music to license from Squorter. We know this. So, I'm kind of like restarting the analysis a little bit here. Compare the audio profile the top 50 most streamed songs against the rest of the data set.

32:07 Show those average dimensions. Tell me what's the different about the biggest hit. And then so here's here's what I was talking about for validation. In this prompt, I said save the chart in the Python code as analysis audio profile code.p.py. So now before where I was just outputting the PNGs, um in this case I also had it save the code it ran right here. Um and it does some validation on itself checking NAS and stuff but you can see everything it actually ran to make these charts. I am not going to go through and validate that right now but like that's part of the job, right? In your day job, you'll want to go and like read that code. Very similar to how when people are pushing production code to build a product, a lot of the work of engineers are now on the review side rather than the coding side. Same thing with analytics. You're going to spend more time reviewing than writing out some of this code. But like it is going to go much faster and then the quality of stuff you can produce is going to be a lot higher. you produce a bunch of stuff at the same time and when you see errors in the code, cloud code is really good at learning those errors if you call them out.

33:31 All right, so this is yeah, this is like the context step. Before we recommend anything, we need to know what popular actually looks like. So most people assume, I mean, we already tal talked about this, that his songs are dable and stuff, but the most the top 50 most songs are actually less dable uh than the rest, which is pretty interesting. They're less energetic, they're lower balance, which means they're moodier, less happy sounding, and they're more acoustic. That's the only thing that they are more on these dimensions than the remaining 92 songs, more acoustic.

34:05 So, the abyss the Yeah, I mean, we saw this in the last stretch, right? Like the biggest hits on Spotify are are very emotional songs. They're not the bangers, I guess. Or the bangers are emotional. I don't know. Um, yeah. I don't know. That's like probably not what most people expect. Maybe some people expect that. I would have thought it would have been something different. I listen to a lot of music while I run, so like I like more of like the happy stuff, but I get it.

34:35 Um, all right. Let's let's let's keep going. I'm g drop another prompt in here. I know someone just dropped something in the chat and I immediately put the prompt right after they said it and hit their chat. Sorry. All right. So, you can see I've done this analysis before. So, I have all my prompts created. Um, so yeah. So now I want to know the biggest hits are dancable. Is it shifting? Show me how dancable energy and valance have trended by release year. Um yeah, we just did this trend over time.

35:15 But the reason I'm running it again is because I want to save this Python code this time. But this should look pretty similar to our last one. Yep. trend over times down here. So, here we go. Dability surge 20 points from 2015 to 2023. The biggest shift.

35:46 All right, I need to move a little faster through this stuff. So this is interesting because this is like h this is not like the popularity of the streams like the last uh graph we had up here audio trends this is like what people are uh creating. So it's kind of like the industry is chasing danceability like this but the songs with like the two three billion streams they're not the danceable ones. So like new music being created is dancier but the songs the most people are listening to is not dancier. So the staying power is like emotional and acoustic. So that's that's kind of interesting right? Like if you're someone who's trying to figure out what music do I license, like naturally you would be going towards this graph of like what's what are people creating, but it's actually not even what people are creating. It's like what people are are listening to.

36:55 We'll do another outlier graph and then we'll get to a recommendation and and start wrapping up a bit. I'll drop that in chat, too. I think this should be similar to the one we just ran. But I want to run it again because I want to get those uh the Python code. But yeah, so from the top 100 most streamed, which ones have unusually low danceability? This is a little different than we looked at before. So under 55% or unusually low energy? Um in this case, I'm going to actually tell it kind of what chart I want. scatter plot dancability or streams with those outliers called out by song name and artist.

37:46 Okay, so 40 outliers. So this is interesting, right? It found 40 outliers and it even says like, hey, I have a bunch of stuff in here that's like you got to cleanly label stuff. I don't want crap cluttering every everything that's like in our kind of skills in in in the system here. And so obviously like it has gone through the workflow and probably done that review step and found too many to label cleanly. Let me pick the most notable ones annotate and show the rest as highlighted dots which is not something you're going to necessarily get out of if you like you know create this on your own first of all just like in Python and Mapl, but also if you run it in probably the vanilla version.

38:30 All right. Uh is this okay? This is pretty interesting. Where did it put it? Outlier. All right. So, these are all the outliers only. It's not all the rest of them. So, 40 of the top 100. So, it's the top 100 most top 100 songs. You can barely see them actually, but they're they're in here. They're gray. Um, but 40 of the top 100 songs are outliers. So, they they aren't like edge cases. It's not like an edge case that like this uh relationship between um like dance stuff is driving big hits.

39:11 It's like nearly half the biggest hits on the platform cluster into like pretty clear archetypes like emotional ballads, moody pop, classic rock, soft acoustic. And so yeah, I selected a few of these Nice. I remember Blinding Light by The Weekend was like two years in a row my top Spotify song and then I never listened to it again. It's burned out.

39:43 Um, okay. We are I think getting close to time and I want to wrap this up. So these like you know look at these outliers. Um, like these are the songs people like play over and over for years. Like I just said, they're not the song of the summer. They're like the song of the five years or the decade. And those are like the licensing wild cards, the songs that don't fit the trend, but massively outperform everything else. So based on all this analysis, now we want to go get our recommendation to our stakeholder.

40:26 So, based on everything we found, give me a three sentence recommendation for a head of content about what kind of music to license next quarter. Just tell me in the chat. No need for a chart. Um, so it says, "Don't overindex on high energy tracks. 40% of the 100 most streamed songs have below average dance ability or energy. And the correlation between danceability and stream is essentially zero." Pearson score of.11. The biggest hits skew towards emotionally resonant mid-tempo tracks like moody pop, acoustic ballads and atmospheric production with danceability in the 45 to 65% range and veilance below 50%. Have you heard of people talk about music like this before? I wonder if people talk like this at at Spotify when they're like like if they listen to their songs like oh yes I'd say that Danceability is in the 45 to 65% range.

41:18 Um, that said, actually, hey, I don't know if I'd want to work at Spotify. I feel like it would ruin music for me. Um, that said, new releases are converging at 71% dancability, up 20 percentage points since 2015. So, the sweet spot is tracks that are rhythmically modern, but emotionally weighty. Think blinding lights territory, not club bangers. Sorry for the clubbers in the audience. So, there's your answer. Um you know kind of how went through these four steps of of storytelling like provided some context um and these are actually like four steps we have we work into the agents workflows context tension outliers recommendation. So it provided context at the beginning in that first question.

42:03 Then it looked at some tension of like how do these relationships like not fit like we think they would. Then it looked at outliers. Um, and then I made a recommendation based off all that. So each chart was built on the last and each one moves a story forward. This is kind of why I like to go step by step because you can like the system can do that. Like we do have it set up so like if you did a one shot, it would look at what it did on the next step, last step before it goes to the next step. But I kind of like being along for like the journey of it in case I want to steer in a different direction.

42:39 Okay. Now, what I wanted to do next, and then we can do Q&A actually while this goes, but all right. I don't know if this will work. I was literally doing this at 11:57 right before our course started. And it kind of ran, but we'll see. I'm going to try and get it to make this into a report, but no promises. It might tell me the MCP is not connected.

43:17 If it does, then I'll try and troubleshoot it while we answer Q&A. Um, so what I told it here is let's package this up and analysis into an existing Google doc. I gave it this this document ID earlier and this Google Google doc I had Google Doc I had open. You could just tell it to create a Google doc, too, but I don't want you to guys like see my Google Drive on a recording live with like a few hundred people or whatever. Uh, so I just open this one. Um, yeah. So, I told it to write a a full analysis into the doc executive summary um section for each analysis with findings each chart we made. Under each chart, this is I don't know if this will work. This works in uh this works in notion. I don't know if it'll work in Google, but under each chart, add a subheading called code used. Pull the code from the saves file and include the data source. H I think I want it to do something different with that, but we'll we'll see. I wanted to do like a toggle thing. So, what it's doing right now, it's taking all these images we just made throughout the past 46 minutes, and it's uploading them all to my Google Drive right now. And so, I'm connected to Google through an MCP.

44:29 And so, that's what all this stuff is. So, it's going through uploading images. Now, it's creating a document. It says all the images are uploaded. Now, let me set the title. And you can kind of watch this over here. It's going to set the title to Spotify licensing analysis. What music should be licensed next? Come on, change the title. It's got the document ID. Ah, it created a d a new doc by mistake.

45:04 But look, it corrected itself. Let me use the existing document you specified instead. I'll write all the content there. All right, there we go. It's starting to populate. That's cool. Okay, there we go. Executive summary. It's got the title in. It's added date. It's saying the data set and source. It's got our graph here for the first one. It's doing some weird stuff. We'll try and fix this. But it's like for some reason it's making all this bold, but that's a quick fix. It's still adding.

45:39 Now it's adding our our next question we audio trends. It's it's written a couple lines about the key findings, the thing that we're kind of talking about earlier. Um he could be concerned about enthropic singing. what's in my Google Drive that will just that will just make the AI analyst stronger. Um, so yeah, but now it's like a good question around um security.

46:13 I think like for what I'm doing here like I don't really care but for uh your day job that's like a pretty serious question but it's like any software you work with uh you know work with your security and legal teams through that process you can negotiate certain types of contracts with those uh companies. Um we'll probably talk a bit about that in the boot camp or at least in or maybe in the fiveweek course. Um but that is definitely a major um hurdle. So we even like will do work where we go into companies and help them like work through that whole process of working with security and legal and being able to do this in a safe way. Um but uh I didn't do that for my personal Google.

47:01 Yeah, it is kind of scary but it's also fast. It is fascinating. Okay, so this isn't exactly what I wanted, right? like it has this code use thing here, but it's just telling me where the Python script is. Well, you guys get you guys get the gist. So, I I might stop it and try and see if I can get it to do the code thing I want and then we can ask answer questions while it works on that. So, let me just get this stopped. Oh, shoot. I exited.

47:38 It's okay. Let's go back in. I need to resume my last session. That's fine. Okay. So, I'll say, "Hey, it looks like all of the text in here is bold. That's fine for the headers, but I don't want all the body text bold. If you can fix that." And then I wanted code under each graph, but I wanted it in like a sort of toggle. So, it's not showing unless I toggle that in Google Docs. Can Google Docs do that? And can you do that?

48:14 And I find like you'll see like, yeah, I ask it some questions at the end before it does it. I don't really like going like, "Hey, go do go do the thing." If I'm doing something for like the first time or it's not in my harness already, I'll like kind of like prime it, make sure it like goes and does some research to see how I do it. All right. 10 minutes though. Um, I'm using Whisper Flow. Yeah.

48:39 Um, I will start answering some questions. Hi. Are there any kinds of questions that came up or should I just kind of like be going through these ones that are popping up now? Are there any themes? >> Um, I think the uh, yeah, we can just answer kind of like the ones that are popping up right now. I think they're all pretty diverse. >> Um, >> yeah. Let me read them.

49:09 >> Yeah. Do you want to read them and I can kind of like mess around with this while you're doing that? >> Cool. Uh, so Kikica asked, uh, this comes off as a little overwhelming to keep up with all the advancements going on. Could you advice suggest how to get started to build a good foundation first? >> Um, two-day boot camp next weekend. Like that would be that's a good like we're gonna get people from like never used cloud code to actually build this like AI analyst for their own use case from scratch. That's going to be pretty awesome. But if you don't if you want something free then like we have a repo AI analyst that I think's been shared earlier in the chat probably. It's the same repo and like honestly like open it up and just start talking to Claude to ask what's going on. I think um follow people who are kind of working in this space on like Substack and LinkedIn just trying to keep up with like the influx of news. But the honestly the best thing you can do is just go start working with it. Like don't worry about going and doing like weeks of research to do it. That's why we're doing this boot camp on the weekend that's hands-on because like a a big hurdle is just like getting your hands dirty and working with it. And I I guarantee you like two days like seriously two days and you will come up with so many more ideas of how to use this like better than we can as well. Um, but it's really critical, I think. Like I took two weeks off my day job, honestly, to find time to just kind of like get that catalyst going. But like I'd say like a two-day whether it's like on a weekend or take a couple days off work if if your work's not like kind of like empowering you to do this is all it takes. And then you can kind of like start automating out a bunch of your work like through this analysis. And then that earns you time to go do deeper research and learning. And then also don't have to be too hard on yourself.

51:14 Like stuff's changing every week. You don't need to keep up every week. Like if you go on vacation for two weeks and you miss some stuff, like you can catch up when you get back. Now if you go on vacation for two months, I don't know, we may be in the Matrix at that point and just like the batteries for the robots. So, you know, I don't want to go away for too long, but uh yeah, don't be too hard on yourself about it, too.

51:40 >> Cool. Uh let's see. Greg asked, "Maybe out of scope for today if head of content wanted to base decisions on both song characteristics and song licensing cost from separate files. How easy to prompt clot code to find an optimal solution?" >> Um from separate files. Yeah, that should be >> Yeah. Yeah, that should be totally fine. I mean, like I have it do stuff that's like joining a bunch of SQL queries across a database, but Hi, you've done stuff with like your with your Airbnb like so Hi does a bunch of like real estate stuff where you've gone cross file a bunch, right?

52:20 >> Yeah. Yeah. So, Claude is really smart. As long as you sort of like uh give it a direction, ask it to give it a high level goal, it will come back with a plan and you iterate on it. So in the so anything that it doesn't know how to do, you can ask it to ask you and then uh and then you can you can really just uh multiply from there. Okay, it did it made a new draft of the analysis when I gave it that feedback around like the the boldness stuff. So, he was like, "Yeah, this looks cleaner for sure.

53:03 Still didn't quite get what I wanted to do." So, now at least it embedded the code in here. I'm gonna just tell Sorry, I'll get back to my question. Uh, this is better. I see that you embedded the code. I really want a toggle for that code. So, it only shows when I untoggle it. I honestly don't know if you can do this in I I feel like we've done this in Google before, but I don't know if it can figure out with MCP. With notion, it does it great. All right, another question. Hi.

53:37 >> Yes. So, have you thought about the output being an HTML website with interactive charts or an editable PowerPoint? >> Yep. Yep. Yeah. We had a presentation on our team with one of our colleagues yesterday where he had basically done a very similar vibe, but it's a PowerPoint. Actually, when we originally created the AI analyst stuff, you'll see it in the workflow. It output a deck. I kind of ended up moving over to Docs because I want to be able to validate the code and stuff. So, I kind of wanted more uh uh real estate for me to put that in there rather than having it in a deck.

54:16 But, you can do both. And then HTML for sure. I do HTML a lot when I'm like doing my own exploration and like trying to like if I'm like doing analysis work and I just want like to see how things are going as I go. It's really nice with HTML, too, because you can tell it to like do some wrong long running analysis and just like update the HTML as it goes, whereas like here it have to output a bunch of PGs and you have to wait till the end for that. HTML you can just have that open in your browser. So, and it's so it's so clean too like the it's really good at HTML and CSS because that's like all these coding tools are like that's what they're trained on first, right? the internet and you can make it hand you make dashboards like dynamic and stuff. I don't know. I don't know what's going to happen to all these BI companies to be honest. Um what's another question?

55:09 Uh I think you answered the other ones. Um so let's see. Uh oh, Anna just dropped another one. Could you please explain how skills work? Do you only use them in clock code? >> I think so. I mean, I haven't used them. I honestly haven't used cloud code other stuff since skills came out. Um, I don't know. Hi, you maybe you want to take that one because I know you've done a lot with you were making slash commands before >> they were skills, >> right? Yeah. So, skills has gone through many iterations in the claude uh kind of like very short history. Um, the TLDDR is think of skills as something reusable. Like if you do something more than once then you can create it as a skill. A skill is nothing more than a file, a markdown file. So something that a computer knows how to read but it's in plain English. Um skills also exist in all the entire anthropic ecosystem. So if you use claude co-work chat, you can create a skill. These are just reusable instructions for Claude to know what to do in certain scenarios. So in our case, AI analyst, there's a bunch of skills around how do you tell great stories, how do you um uh you know what chart types to to to bring up for in what occasions, what should be the look and feel, uh how do you validate and stuff like that. So that's that's kind of like the the high level. Hopefully that makes sense.

56:43 Um, Maxim, I don't uh I don't know what the industry is going to do, but I would say like if you asked me a year ago, I was more like, oh, they're leaning towards senior talent. And now I think at least for me my perspective is like I think that like um more entry level I think there's I think there's opportunity there cuz like I've been working with some folks who like are out of college in the past couple years and like they learned how to code with AI and it's actually the creativity they have around using stuff that's on the frontier that's changing so often I feel like is a little is like on average, not everyone, but on average like a lot more flexible and forward thinking than like folks who have been like kind of doing this for like 20 years, but I don't know what the industry will do.

57:40 And how what's next? I know we're at time here. I'm gonna answer >> you can answer one more for folks who want to keep asking questions. find us in the uh in our Slack channel which I'm going to drop the link to. >> Yeah, let me see. Uh how much how Okay, it's a good question. How should you think about how much you need to break down a skill into small components? We'll cover this in our boot camp in whatever it is nine or 10 days as well. how to decide what's agent skill helper when to break it down because also sometimes cloud code there's things that could be agents where the same thing could be done with like Python right if you're like just like routing things around um a lot of the time I will if it's like skills if it's like things that will often correlate together like you will often use do these things together put it in the same skill because it's it's effectively one thing. But if it's things where it's like I'm going to leverage this like uh component or this skill like independently or in different workflows all the time without the other pieces then break it down into a smaller component. That's kind of how I think about it usually. And I also ask Claude a lot about it too. Um if things go off the rails some I usually will say like can you review your skills and agents and blah blah blah. Did any of that like push you off the rails?

59:18 Okay, that's all the questions I can answer for today, but obviously join us in Slack. We're going to have the boot camp in 10 days. That's going to be really fun. We're going to have a lot more time to answer questions. We're going to go really deep into a lot of the aspects of this. Um, and then one more cool thing, we have a fiveweek course that's like more endtoend how to do analytics with AI. It's not just cloud code, it's with chat. Um, it's with uh other third party tools as well. Um, and if you do the boot camp, like basically like if you take that fiveweek course, you get the boot camp for free. So like it's like if you do this like cloud code intensive boot camp then we'll just knock whatever you paid off that from the fiveweek course. So if you want way more extended like access and we'll have like office hours every week multiple times a week to that too for the people who join that then um that's another place where you get a lot more insight and questions answered too. Otherwise I'll see you in Slack. But I appreciate everyone coming.

60:31 Hopefully uh this was fun. >> Thanks everybody. >> See you later.