Excel and Excel AI Hour

Hey everybody, sorry I'm running a little late. I just joined. Let me know if you're out there and we will, uh, get started here in just one second. So if you could just let me know if you can hear me and if you're out there, be great. I see a few people on the webinar as far as through, uh, StreamYard. I'm sure there's some people on LinkedIn. Um, everybody, if you haven't downloaded the file, you'll want to go to my website. I'll bring that up here real quick and show you where you can download that. We're going to keep this pretty informal, pretty relaxed today. So, uh, welcome, Del. Others, if you're out there, please let me know where you're coming from, let me know your name. And real quick, I'll show you where you can download the file if you haven't got it. And then we'll get started. So appreciate the patience. I had a call that ran a little over. I try not to schedule something right beforehand, but I did today. So if you go to my website, thefpandaguy.com, if you go over to the educational resources, there is a section called Excel Files here, and you'll see the file for today's session. You can download it right there. And let's see, we have Dale from Los Angeles, we have Benjamin from Kenya, I think I saw Leonard Sloan out there, not where he's from, and I know there's others. So again, let me know. We got Brazil, so we got a lot of different people. I'm going to ask a question as we get started. I'd love to know how many of you have used skills in any way, written a skill with, uh, like Claude, or given an instruction file to Excel in any way. Let me know in the chat if you've done that, and I'm going to share a little bit of my thoughts. Like I said, this can be pretty informal, and then I'm going to show you just a few things that I think you'll find valuable on this. So let me know in the chat. Where I want to start with is, here's what I'm seeing, and I haven't spent near as much time as I'd like with AI. I'll be the first to admit it, but I'm seeing more and more that it's just changing the way we work. It's the reality. I spent some time on my financial modeling podcast with others, but I haven't spent the hours I'd like. So Dale says, I've seen other people use Skillshare, hoping to learn how to do it. All right. And we will talk a little bit about that for sure. So one thing I want to share that I found really interesting, and I'm just going to go— I'm going to go to the— find it here. I'm going to bring it up. So this to me highlights the, the benefit of skills. I've seen it myself, and I'll show it. Lance Rubin does some really good things around financial modeling, and I want to share a post he recently did. Let me just pull it up. Let me just pull them all here. It's one of his last couple he did. I have it summarized in a deck I'm getting ready to present, but I just wanted to show it. So right here, so if we look at this one, in preparation for the upcoming session with, uh, no, here we go. Last week I did something no human can do without AI. Not even Microsoft Copilot could do it. So he took— I had 6 live Excel workbooks open at the same time, changing formatting, rewriting formulas, restructuring logic across all of them. And so what he did is he compared all 6 for a financial modeling case to see how they performed. Copilot with no guidance outside of his prompting, he, he tested a model on 7 factors: risk, complexity, structure, flexibility, validations, dashboards, and automation. They each had 10 items, so 70 in total here, and then he scored them. Copilot scored a 14 out of 70, 20%. Now, part of that, I would argue, it would have been— that would have done a lot better with instructions, but it doesn't have an easy skill thing like Claude does. So really interesting. Claude without skills didn't score much better, 18 out of 70, 26%. Now, these are complex models where you're really getting into business requirements, not, not the simple stuff of, hey, build me a model or or things like that. TabAI, which is a tool that they've been working on for a couple of years, designed around modeling, modeling, done a lot of really good work. I know he's worked with them. 50%, 35 out of 70. Here's the real kicker. Claude with skills, 64%. So, I mean, if you look at that, substantially better. His model he built, 97%. But what his point is, is, look, with and without skills, look how different it is. So what I'm here to tell you is the whole idea of skills, and I'll share a little bit on Claude. I'm mostly going to share more what I call instruction sheets, which are basically skills, can change the way AI works. For, for the longest time we've been hearing it's all about the prompt. I would argue it's all about the context, the skills, the guidelines, and the prompt is secondary. A bad prompt will get you bad results. A bad prompt with really good guidance might get you okay results. A good prompt without the guidance, the context, the skills, the instructions will get you pretty decent results. Now, you want to get really good results, you combine that prompt with the context, the skills, and you can go beyond skills, and then you start to get much closer to repeatable processes. Not 100%, nothing's perfect, but you get some really good stuff. All right, so let's jump into the, the, uh, Excel file. I want to focus on kind of skills around visualization and just show a few things. So I've done a number of sessions, you can find them all on YouTube, on visualizations. So what I want to show here, let's start— let me shrink this down just a little bit. You bigger than I wanted. And did my— oh, just— that's not what I wanted. It went really small. Let me fix this. Give me one second here. All right, so I have some data here and I want to show some differences. You can do a prompt with me. I'm going to start with Claude, but you could do what I'm going to do with any of them. I have Claude on here. I I have ChatGPT and I have Copilot, although I've been having some problems lately. Copilot disappeared on my menu, so I'm not going to show Copilot today. I gotta contact Microsoft. I've reinstalled it. Uh, I— prompting is the most valuable skill set these days. I would actually disagree with that for most people, Arisha. Prompting is important I think it's giving it instructions and skills, and I'll show you what I mean by that. So let's take Claude. I have this data here. I'm going to highlight these 3 columns, and we'll do the same thing with ChatGPT. And I'm just going to say, build me a graph using the data in columns B6 to D16. This graph will be a presentation— we'll just say for senior management. Your management— the focus is on showing year-over-year growth. Do not use— I'm going to tell it not to use the instruction sheet in this file. Do not use the clustered column instruction sheet, but give it your— but build it using best practice data visualization techniques based on your training. Something like that. I've used different ones. We'll go with that. I can give it more context. And yeah, I could get a different result, but I just want to show something basic. And let's let it build for a minute. So it will take a minute. Well, it's doing that. What I want to show, and we'll, we'll show it in Claude, is here's my instruction sheet. So this is something I went to Claude and I asked it to build an instruction sheet. It built one, I modified it, it built me a skill. The reality is, I think it was, was it Dale was asking how to build a skill. The reality is, and let me see if I can enlarge to my screen a little bit. I realized the Claude's a little small, so let me adjust my settings just to make that a little bigger. I got to do change my resolution, so give me one moment here. All right, let me know if that should hopefully make it easier to read, but it's going to make some of it a little small. But we'll do that for now. So we're going to now look at what it did. I asked it to build that. So here's what it did. It said I'll take a look at the data first and build a chart optimized for showing year-over-year growth to senior management. All right, let me build a helper table sorted by year-over-year and then the chart. E and F columns are in use. I'll place it elsewhere. Talks about using an INDEX MATCH and it's continuing to work. So we're just gonna let that work. I'm gonna go through the instruction sheet a little bit more while that works. It says, ah, my earlier call actually partially succeeded. Don't know what it's doing. Very different than last time. A lot of that has to do with we just did a prompt without context. All right, so here it's done something. Now it's created its own chart with the year-over-year change. It's put the sales in millions. Now it's probably going to build the graph. All right, in the meantime, while it's doing that, I want to show you All right, says there's potential data loss. Fine, we'll see what happens. I'll let it do that. I can always undo it. All right, so while that's going, we'll come back to that in 2-3 minutes. The skill— what I've done here is this, this is my clustered column. This is originally what I called the skill. I could probably rename that now, not the cleanest, but I said here the purpose is to create a clear executive-ready chart, best visualization practices. I gave it my data requirements. I said look, raw data contains 3 columns, and I told it the first column is the category, the second is the newest period, the third is the oldest period. And I went through and said here's the chart creation rules: select the first 2 columns, the categories and values, insert Excel's clustered column chart, colors and styling These are required bars. All bars should be dark gray. I even gave it the hex I wanted. Data labels— show the data labels. Match the color between the bar and the data label. Do not show the vertical axis. So I don't like to show that, so I specifically told it. I went through and said, here's my layout and formatting rules. Gap width 70%. Series overlap 0. No grid lines. Don't include a chart border. Category x-axis oriented horizontally, so I don't want them to do labels vertically. Fully visible with no truncation. Resize the chart so every category label is readable. Bars are not compressed. You know, chart title rules that I kept very loose, but I can give it more instructions. That's really all a skill is doing. If I dump this, I could take this right now, I could dump it in the Claude here, or it could go out to Claude And you'll see I could even go to Claude. The reason I've created these instruction sheets like this is I can run this in Claude, I can run it against Copilot, I can run against ChatGPT, and just tell the tool to use my instruction sheet versus building an actual skill that only works with Claude. That's a markdown file. Now there are benefits to that, and you can get that if you're using just Claude. I would recommend creating it as a markdown file. So all I could say is create a skill for this clustered column chart I want to create. I could dump it in here. I could ask it to ask questions, but if I just do that, it's going to go through and create one. And so, and that, and when I'm done, all I would have to do is then load it in here. It'll even tell me how, tell me how to load it. I save it. Now, whether I'm running one in the LLM, whether I'm running something in Excel, I can always reference that file. It's literally that simple. Yeah, you can ask Claude to write macros for sure, 100%. It will write macros better than most humans will. That's code. It is fabulous at code. So see, it says, now I have enough context, let me create the skill. This is a variance chart skill. Oh, I mentioned something at the bottom that I ended up not including. It noticed that, so it's trying to create a skill for that as well, which is fine. I'll explain why. It partly had to do with— I broke this into two tasks, which I'll show here in a minute. But all right, so let's come back here. I've kept the original data untouched and it's built it. If you prefer the chart show absolute unit change instead of percentage, I can add that. All right, so here's what it built with me giving it just a simple prompt, and I did this earlier today. The title was completely different. The coloring was different. Like, see, like the Subaru right there. Why did it— it put all the labels in the middle, which doesn't make sense. All the labels should have gone to the left. And so that's— it was a decent prompt, but it didn't have my context, any of my instructions. And so that's what I got. Now watch, I could ask it the exact same prompt and I will get something different. I might not even get the same graph, and that's because it's just using probabilities. So watch, I'm going to go clean, I'm going to delete this, I'm going to delete this, and if I do the exact same thing one more time and we'll let it build, I'm going to get something that probably looks quite different. See, again, it's building a small helper table. It looks like it is going to do roughly the same, so I may stop. It's creating a horizontal bar, add data labels, and a clear title focused on the growth narrative. Okay, so I'll probably get something pretty similar, but you do it 3 or 4 times, each one's going to be a little different how it lays it out. So let's just stop this for the moment. Now I'm going to ask it, I'm going to say, we're just going to do a new prompt so it gets clean context. Please build a clustered column chart using the data in cells B6 through D16 and the instructions provided on the clustered column instructions sheet. The instructions should take precedence on building the chart. So now let it build for a minute. So, right, it reads the instructions first. That's the context. I tell it to only select the first 2 columns, and I'll explain why. Here in a minute. So you can do— if you have, you know, Claude or whatever, you can do that. All right, so now it's built it. It still has a few things to fix. Let's see. But you saw how much quicker that was than the other. The only thing it didn't get right that it normally does, and we'll see if it fixes itself— it says, hey, if you want to— is it did not do the labels right. So that tells me I should have been more clear. I should have specifically said— and we can see it under Data Labels, Show Data Labels, Data Label Color— data label format type should match, should be in my chart, should be the same as the format for my data, i.e., if the underlying data is formatted in millions, the labels should be the same. So it did everything else the way I wanted. It gave me the colors I wanted. It gave me the width I wanted. You'll notice it got rid of the outline. It got rid of the axis. It kept these straight like I said it should. It made these the same size. It created its own title. So you saw how, how much quicker it ran, and it gave me what I like as my best kind of format or style. That's the power of the skill. And so now I'm going to run this one more time. I'm going to delete this. I'm going to take the exact same prompt and now we've changed the skill. Let's go ahead and run it. And a lot of the skill, I'll be honest, I'm figuring it out as I go. It's playing with it. It's trial and error. It's asking others. You're going to start— I mean, there's a lot of GitHub skill libraries and Claude will help you build skills or instruction sheets, but using them takes your prompting to the next level. See, right away it said, right with the next one, it noticed data. The data is formatted as millions, and I'll match it. So there we go. You saw how quick that was. That was using Opus 4.7, you know. And so now the reason I built this the way I did is I like to run a second chart. So one of my favorite charts is called Actuals with Variance, and I'll show you what it looks like. And it's one I've been trying to get AI to build for me I've got 90% of the way there. So this is my— let's see if this is the one I want. Oh no, that's not quite what I want. Let me open a different file. Give me one second. Actuals with Variance completed file. So what I want to do— and this one is colorblind friendly, I'm just going to do red-green— is this is what I'm getting it to create. It's creating the top one Then it's creating a second chart, the bottom one with the instruction. So next I'm going to have AI create this. This is actually a scatter plot. When I train people on this, it probably takes 20 minutes to teach them the first time, and it takes a long time to learn. With a skill sheet, I can build the whole thing in about 3 minutes, especially with now with Opus 4.7 getting better. So what I'm going to do is we've— not this one, sorry, let me come back to what I had. We've done this, so now you'll see I've created the columns here. It will fill those in. I have instructions for building the scatter plot. Create a scatter plot to show year-over-year data. Give it the purpose. I tell it what to do with all 5 columns. I give it some creation rules. I tell it to do what to do with each data series. Um, I talk— give it data labels. All the examples previously. I even gave an example what I wanted. I ended up removing that, and I'll show you here. This was trial and error. I spent a good hour with 3 different LMs trying to do it. I could get none of them to get 100% of the way there. These are the type of things you'll really run into as you try to do this on your own. My goal today, and you can try to follow along if you have a tool, is really to show you it's a new— this there's a new skill we have to develop, and it's all about giving the instructions, the prompting to make the LLM better. That's what a lot of this software is doing. And there are some real benefits in selecting software tools. I have my software showcase. If someone wants to see some great LLM tools out there, I believe Arsha that's joined works for one of them. There's Sapien, there's Concourse, there's some tools that's on my website. Before we're done, I'll show where you can sign up if you're interested. But there are cases where we need to learn how to do it. It's not 100% going to be with these tools is my take. There's real benefit in getting better and better at it. And to do that, you really have to know how to build the chart in the first place. So let me run the scatterplot now. So what I'm going to do is I'm going to come back here. I'm going to highlight this right here. And I'm going to say using the data highlighted, I'm going to get a little lazy this time because it's already showing it's selected. I like to implicitly type it or explicitly versus being implicit. As a general rule, anytime you can be explicit, you get better results. That's because you're being very clear to the computer versus letting it interpret. And so that's why I usually tend to be more explicit with exactly what I want. But we're going to do this. I think it's structured enough. It's simple enough. It will be fine. Using the data highlighted, build a scatterplot below the column chart on this page using the scatterplot instructions sheet. All right, so I'm going to let that run here for a minute, and I think that should be good enough. What, what happens is— what I've learned is sometimes the, uh, the backend doesn't have an automated way to do certain things. That's what I've learned with these scatter plots. What you're going to find is what we're going to— what it's going to do at the end is not going to look like what you would expect. I've spent an hour, I've tried it with 3 different LLMs. They've tried all different approaches. None of them can quite get one part right. 4Seven tells me, says, hey, I can find no way to do it. And we'll see that here in a minute. And so it's going to run it. And there it's going, is creating this scatterplot. I'm going to allow again, dangerously, always allow. I'm going to live dangerously. Feel free to ask questions as we're going through this, please. Yeah, this was a— this is a hard one to train right now, especially kind of virtually where we don't have the opportunity to chat. But I wanted to show some things I think it will be helpful, and I've spent most my time on visualization, so that's what I wanted to do. All right, so it's working. It's going to take it a few minutes. To figure things out. And so now you can see I have the dots that I want. Still doesn't look anything like a scatter plot. It's like, hey, the series exists, I'm going to keep working. And you'll see how close it gets. So while that's running, anyone have any good Excel jokes for me, or I may tell one? Let's see, does anyone know what a newborn spreadsheet needs? What every newborn spreadsheet needs? Formula. Not a good joke, but it's a joke. Let's see. Does anyone know why are spreadsheets like prison? So Bridget, good question. Yeah, you could transfer these to PowerPoint for sure. You can, you could do a paste link. You could probably even ask Claude. Claude will work across tools. I haven't done that much yet. That's on my list to test more. So you notice it says all the settings look correct, but you're not getting what you want. It's saying the instructions say change the— so let me— ah, so I have in my instructions— I want to show you one thing, see if it's here. Yeah, so what, what it's struggling with— I want to show this. I know I was answering a question here, but this is worth seeing. So what it's trying to figure out is right now, if you look at my data— let me just bring that a little bit bigger— you'll notice all you're seeing is a little tiny line that doesn't look like a, a column at all. And it's saying, hey, I can't figure that out. And And finally, here's what it told me. It says I cannot set percent to 100%, the bars. And so yes, because they have too many cells is the reason for the prison. So bars, yep. So watch this. I spent probably an hour trying to figure this out, and this is as close as I could get to what I wanted. And but watch this, look how simple it is to finish this thing if you know what you're doing. I'm going to come in here, I'm going to select these error bars, I'm going to hit Ctrl+1, I'm going to come over here, I'm going to select my Y error bar. Now watch this, the percentage— it could not change the percentage, it shows as 0.1 there. Watch what happens when I manually change it to 100. All of a sudden I now have the error bars I want. And you get more control with these below and above because you have separate series, which allows you to easily set separate colors. You, uh, don't have to worry about as much as if you try to do this with just a column chart. You get a lot more flexibility. This is one I learned from Lee Ligerani. And so they're two separate charts. Okay, so now I can just make this a little bigger. You know, kind of match it all up. And the last thing I have to do myself is just add my data labels. So what I'm going to do is select the first series, and let me shrink this down just a little bit. Come over here, select my data labels. Okay, these ones here don't look right, so I'll hit Ctrl+1 and I'll put these below. And because they're different series I could just come and change the color. I could match it exactly. I'm going to guess saying, oh, I think it's this green here. Make it bold, maybe make it a little bigger, which I could have done. I could give my instructions, say use font size 10 with labels. The reason I didn't put in automated labels is because it won't put them in the right place because it can't change this to 100%. I could never get them to look the way I wanted unless I did it at the end. But look how long that took me. So, so if you don't know what chart you want— so this is really when you know. So what I'm going to go two things here. What I'm showing right now is once you've learned how to do good data visualization and you know certain charts you want, that's where instruction sheets become incredibly valuable. If you have no idea what kind of chart you want, there's a few things you can do. First, when you're deciding what chart— and I, I teach this in my visualization course— it starts with what's the message you're trying to accomplish. Certain charts work best for certain types of data, but there are usually multiple charts that can work. So the first question you have to ask is, what's the message I'm trying to accomplish? Then, what helps does it? This one here, this chart in my opinion, the message here it works really well for— and I would change this title because they just use 2025 sales— is to quickly understand what's happened year over year, right? This is much better to me than if you put Toyota here, Toyota next to it with their numbers, right? So it's a little small to get that all to fit, but if I just showed 1.91 and 1.84, it's like, okay, I can see there's a little difference. But here I can quickly go, okay, this is my biggest. It was pretty flat year over year. I can quickly see, oh, I got a big problem with Chevrolet. Honda and Kia have grown quick, but they have a smaller base. Then if I wanted to focus on, hey, which ones are growing fastest or which one lost, I could always put a red shape around them. And so it's all about the message. If you don't know much data visualization, you don't have a good idea of what graphs you want, and you just want AI to pick, there's a couple things you can do. One, you can ask it to pick, tell it to use best practice data visualization. Also create a general skill sheet saying, for my graphs, I don't want grid lines, or I want light grid lines, tell it the hex. I want legends, I don't want legends. I want a certain font size. I like dark grays or light blue scales. Don't use a lot of color. Give it an instruction sheet for general graphs is what I would recommend. So very similar to this, I would just go in and say, okay, skill preferences for best, you know, for, um, for data visualization. Select the chart you think is best based on what I tell you the message is. So the chart should match the message. Then give it any data requirements that you, you know, hey, my data is always going to be here. If you do, if you don't, say I'll specify in my prompt where you locate the data. Then give it any rules saying, hey, coloring and styling, I always like here, I do like this, provide no data labels for the— and this one is what I told it here for this, this one. Do not show the vertical axis. Or let's take this one here. I told it data labels should match. I gave it my gap width, my series, my grid lines. So the more of those instructions you'll get— you give them, the more you'll get your style. Like if you have color branding, can it produce a commentary on the variances? Yeah, yes, it could. I mean, you could have— you could ask it to write, you know, checkboxes. There's different ways you could do it. Um, I haven't played a lot with adding the commentary, but for sure it can do it. You may run into some challenges, kind of like I did with this column where I had to do a few things manually, but I was only able to do that because I know how to build it. So next, let's just see a couple other examples of how good it's getting. And this one I find really interesting. I've built this a lot of times. So I have this data here. I built my skill sheet for a waterfall chart. What I'm going to do is we're gonna— we'll go ahead and use Claude again. We could do ChatGPT, but I, I like Claude. It's working. I'm going to come up here Move this out of the way. I always— when I— once I build a new chart, I always go to a new chat, clear out the context window unless it needs that information, because it will just— one, it will erase your context, so you'll run out of tokens quicker. Two, it's confusing. If it's not adding to what you're asking it to do, then don't include it. So only keep the conversation going if it's helpful. So let's look at a waterfall chart example here. So I'm going to start with just say, build me a waterfall chart using the highlighted data. I want to see how much better Force7 has got. Well, it's doing that. Let me see, I'm going to open up a file, waterfall Claude. Here's what I built last time without skills. So without me giving it a skill— and I should have told it not to use the skill sheet, we may go back and do that— here's what I got. Not, not bad, but there's a couple problems. One, it put actuals first, put budget last. It did it, you know, horizontal instead of vertical, and it built this right here. So what you're looking at if you go to the chart type, is it used stacked bar charts. It didn't even use the waterfall chart type, right? And so me letting it on its own, I got something very different from what I wanted. This time it built a waterfall because I specifically said build a waterfall chart. But even then, sometimes I don't get that. So, right, pretty good, not bad. You'll notice it couldn't set— it said one thing to check. Excel's waterfall chart doesn't always auto-detect the first and last bars as total. They'll render as floating increases if the budget and actual bars aren't showing a solid total sitting on the axis. Right-click each of these bars and set as total. So I've never been able to get In Excel, without do— I know one person who's got to work without doing some really complex stuff. It never sets this second one as a total, and so I have to go in and manually set it. But all right, so there's what I got once I set it. Yeah, not a huge fan of that. So now let's see what happens when I build it using the instruction sheet. So Build a waterfall chart using the waterfall skill sheet for your instructions. Please follow the instructions as close as possible. It will run into one challenge, but you'll see the difference here. So keep your questions coming as that runs for a minute. And so really, like I said, my goal today— this isn't the typical, hey, here's this formula, here's that— it's to show— I, I hope you take, I would say, two things away. One, if you don't know fundamentals well, you're, you're gonna leave yourself at the mercy of the LLM to produce good data. Which will bite you at some point. You need to know what you're doing well, whether that be data visualization, whether that be modeling, whether that be data cleaning, whatever it might be. The better you understand that principle, the easier it will be to get good results. So that's the first message. I'm able to get what I want because I know what to— I know what good looks like, and I can continue to work with the LLM until it gives me what I want. That's true in Excel, that's true with other things. The second one is the takeaway as this finishes, is you need to learn— we need to get better at writing instructions, whether it be process workflow, whether it be charts, whatever it might be. Instructions naturally lead themselves to skills. I will show you here in a minute, I can get very similar results to this using ChatGPT, using Copilot. That's why I did an instruction sheet versus an actual skill in Claude, because I can't port that skill from Claude to Copilot or ChatGPT, and everybody is going to be using different LLMs. I will say, I think right now the best LLM out there is Claude. All right, so again, you'll notice one thing it struggled with. So now I got the colors I wanted, it gave me the exact colors I wanted, I got the width. I think I forgot to tell it to not give me connector lines, so it gave me those, which I generally don't like, but I could add that to the instructions. Now I set that as a total, and you can see there that to me was much better than what we got before. The workbook's already shared, you just have to go to my website to get it. All the instruction sheets are there, I share all the workbooks, so you can get it there. Now you'll notice one thing here it did, because it didn't set this as a total, this $0.1 million doesn't show because it's above— it kind of ended up being slightly above the limit because it didn't quite quite set this right because that was a total. So I might go manually change this. I might even get rid of it. I probably would. So there's always going to be little quirks because of the age of Excel, if you're building in Excel, and the fact that sometimes the LLM just isn't able to get certain things to work. And that often has to do with more what access they have to writing scripts in Java, against Java or the OpenOffice or whatever the office application. It's not the LLM at this point. It's often Excel. So, you know, I could come in and do my little bit of cleanup. I might make, like, I might come in and say, oh, maybe I want this to be 11 million. And so I do that and— oh, I did 111 million, that's not what I wanted. Oops, make that 7. See, now the 0.1 million shows up. I might get rid of this and maybe that's what I want. And so the reason that was not really showing is it didn't resize it quite like I wanted because this total was up here. And so it was trying to show everything, and that shrunk down the 100,000. If it was able to set this second total like it should, it would have done that exactly right. So I had to go in and do that manually, but still a lot quicker than me doing everything. Now, in all honesty, you know, I could probably build many of these graphs because I've done it so much, and, you know, pretty close to what the LLM can do. And what I want to show is just how good 4.7 is getting. So one of my favorite— and I'll show this chart, I'll show you what it looks like. I have it in a different file. I'm going to show this one. This is not one that's in the— is out there on Excel. This is part of one of my courses. But one of the things I teach, and we're going to do this with Excel here in a minute, we're going to have it do it. Is I, I teach how you can use checkboxes and formulas. So Leonard, the instruction sheets are in the Excel file on my website. So let me just take a minute and go back to that real quick. So if you go to my website, go to thefpandaguy.com, go to Services, and while you're there, buy a few courses. But we'll say— I mean, go to Educational Resources come to Excel files, download today's file. It's right there. The video will get up there in a little while. The video will be on YouTube. You can come back and watch any of these. That will give you all those instruction sheets. All right, so this last one I did not do an instruction sheet, but I just wanted to show it. So this is something I teach people how to do, is how to build a dynamic chart where I just click the checkboxes and everything but the cells goes away. I click on, they show up. All you use is checkboxes and a simple IF formula and an NA, and you can build that. I teach that in other courses. Thank you, Del, for copying in the link. Um, you can find that on YouTube. I've covered the NA and the checkboxes before, so I'm not going to cover those in detail well today, but here's the prompt I wrote. So I'll walk you through my prompt, and then we're going to run it, and then we may run it— I think we'll have enough time to run it once in ChatGPT. I haven't tried that yet, and we'll see the difference. So I say, create a column chart using the data on the checkboxes data tab in cells B3 to F15. Column B contains the dates, column C is the sales and should always appear on the graph. The other metrics— COGS, gross profit, and customers should be tied to a checkbox so I can select the checkbox and they appear on the graph. When I turn off the checkbox, they do not appear on the graph. So that's the instructions I gave it. Good instructions, probably not great, but let's just see what happens. But again, I knew what kind of chart I wanted, and that gets back to you need to know what you really want, what makes sense. And so here I would, I would get better results with instructions, but I just want to show you how good it's getting because a few months ago we couldn't have built this with it. I don't— I, I think it would have struggled. All right, so we're going to let it build for a minute. Feel free to ask a question while that builds. And it may not build it right this time. Last time it did. This will be— oh yeah, okay, it's, it's still figuring it out. It's struggling a little bit. Say, hey, let me try a different approach. Now it's adding the checkboxes. It's figuring out I need them. There we go. So a little messy what it built here. You can see it kind of sits on top of itself. I need to move this You know, so again, instructions would help with something like that, but I would have definitely done some things differently here. But look, got the checkboxes right, use the right formulas, use the IF. One thing you'll notice this time it did is it didn't— the labels are working. Last time when I asked it to build it it made the labels an NA as well. So when I got rid of one of these, what would happen— oops, let me undo that— what would happen is down here— and it's still working, I think— down here on the legends, when I clicked off on COGS, this would show NA. When I clicked off on gross profit, this would show NA. So you never know exactly what you're going to get when you let it build on its own. But look at that, I basically just told it. Now I could clean it up. I could have gave it my general instructions, say, hey, don't show this, label things if I wanted. Here I probably wouldn't do labeling because I think it's— there's too many data points to label them all. If you have all 4 on, I would probably use the axis. But I could have gave it that instructions. Hey, if I have more than 2 series, give it an axis. If I have less than 2, Use data labels. And so there we go. So now let's try that once. We're going to try the exact same thing. Just so you can get to get that idea. I'm going to go to ChatGPT. You'll get very similar results with the instructions. I can tell it to use any instructions. I just want to show you what we get if we ask it to do the column. And that's not what I wanted. Oh, it's— why are you treating that as— that should not be treated. All right, there we go. Let's see if it will now. Uh, it's just— it's gonna load it in as an image. That's fine. Let's see what happens. Oh, it already noticed. I— let me redo this. So I'm going to stop it. Because I forgot to delete the helper stuff. I don't want that there. I want it to build all that. I just deleted the chart. All right, so now we got a fresh one. We're going to copy this and give it to it again and let it build. I left it on fast. Maybe I'll go with standard and let's see what we get. Well, I only see a dark blank image. Okay, this time it wasn't able to read it, so fine. I'm going to copy it over here. I'm going to copy this, put it in, and let's see if it reads it. It should be reading it. I want to type it in. There we go. I think it's reading it now. Ah, it's not. See the difference between the two? It's not reading this right. Let me try something here. I'm going to try copying— I copying this into a— I've never had this problem before, so if it can't read it, that's where I would need to manually type it, which I don't really want to do. So I'm going to try one more thing here. There we go. I copied it in from Microsoft Word. It didn't— it tried to copy it in as an image from Excel. So you can already see which one's doing better. So let's copy this. Let's start over. I'm gonna get rid of that. I don't need that there. All right. Yeah, it didn't find existing checkboxes, which it shouldn't have, so it's going to link them. You can see it's already a little messy. The, the other one created a cleaner version for sure, but let's keep seeing what we get. So it says always on chart series checkboxes. It says there's checkboxes here. Those aren't checkboxes, so I did not find the actual checkbox controls on the sheet. So what it did here, which I don't love— you can see it really struggled with this exercise. If I change this to false, that goes away, but it didn't label things right. So what tool you use— now, if I gave it an instruction sheet, I could probably get there even with ChatGPT. But right now Claude is just far better. So I think we only have a couple— we have a few minutes left. I know this was a little different than what I normally do, but I hope that's helpful in how to start thinking about some of these things and start building. And if you're using Copilot or ChatGPT You're not going to get as good natural results as Claude. Instructions are going to be even more important and your prompt. Claude does a better job interpreting right now. So feel free to ask any questions. I'll hang around for a couple of minutes and let you guys ask some— ask any questions you have. I'm glad you found it helpful, Dale. Also, I'll let everybody know, just so— is there's questions coming— if you go to my website, I do have several training courses. I have my digital courses, I have corporate training, I have a free data storytelling, I have my FP&A best practice course coming up. You know, love to have you guys join me in any of those. You can see my Modern Excel course, my ultimate bundle, Leveraging AI for Better Data Visuals. I'm going to re— I'm going to add more and more to this course. I'm already— I'm going to start adding instruction sheets in the future so you can learn about those. And then finally, I'll drop this in the chat. For any of you that want to attend my upcoming showcase where I'm going to be demoing— where vendors are going to be demoing their tools, both planning tools and analyst tools, you can join that. I like to see, kind of see the future, see these tools side by side in demos and some amazing things they're doing. How does one add the Claude add-in to Excel? Great question, Del. So first, you are going to have to have a paid account of— so to use— so is company data privacy protected? You can— great question. So let me talk just in general. Let me go back to Excel. If you want to add in Claude, if you wanted to add in ChatGPT, you can go to Add-ins, you can type Claude, you can type ChatGPT, you can type whatever you need. And add it. It's all you need to do. Now, to use it, you have to have a paid account to use the, the Claude add-in in Excel, to use the Copilot add-in in Excel, to use ChatGPT, whichever one you're using. You will need a paid account. Now let's talk security for a second. If you have the basic account, it is not SOC 2 compliant, but you can turn off use this tool to train my data. I always recommend you check with your company for its governance policy around sharing. If you have company-wide, you're all able to use it, that probably means you're on a Teams or an enterprise account— you're on an enterprise Teams account and it has SOC 2 compliance. If you're just paying the $20 a month, you're on your own. You're not going to get SOC 2 security compliance but you can turn off— there is a switch where you can say, don't use my data to train. So that's what I'll say on privacy. I'm always careful there to make sure you check with your company. So, you know, you use, use caution there. You can add ChatGPT as well. Copilot should automatically appear. These are other tools. TraceLight, HyperPerfect, Melder are very similar to Claude and ChatGPT, Elkar. Those are different ones I've tried over the last year. So that's what I'll say. Other questions? We have another 2-3 minutes before I have to drop, so if anyone else has questions, I'll hang on for another minute or two. I'm going with my Excel hat today. That's what the FX is, it's for— it's a Microsoft Excel hat. Glad you enjoyed the section, the, uh Session. Good, John. Thank you. Some other comments. I see an interesting here. Dale commenting, this was very helpful. So thank you. The reality is we all got to just get in, figure it out together, figure out how to use it. Some are ahead of others. If you don't, you're just going to get left behind. It's like when the computer came. If you chose not to use the computer, it was just a matter of time till you got left behind. The same is true with AI. But one of the best parts is if you're wanting to create a skill, just ask AI what you want to do. Honestly, if you ask AI, you're telling what you're trying to do and write your prompts for you. You'll often write the prompts better than you will. Yeah, you might have to tweak them. If you ask it what kind of context it wants, it will tell you. AI is in many ways almost the ultimate instruction Like, I could have just gone here and said, I— let's just give one last quick example. Help me create a general instruction sheet for all visuals I create in Excel. What should I include? Walk me through step step creating one. Ask me each question with options for an answer, and you can start building one. So my website, the name for my website, it's thefpnaguy.com. It's just— dropped it in the chat there. That's the website. The F Panda Guy. What should the sheet cover? Chart design principles, data prep standards, chart type selection. So it's asking me what I want. I can say, well, I think I want chart type selection. Which chart types do you use most often? Uh, I tend to use waterfall and combo. Okay, and it just took that. It will start something. Now that's a beginning. From there you're going to continue to tweak it and play with it and learn. But if you don't know where to start, see next. What's your preferred color approach? I like monochrome and accent. How should titles and labels be handled? Minimalize, minimize grid lines. Always show data labels. Direct labels over legends and descriptive titles. I like all of those. What about axis and formatting? Start y-axis at zero. Minimal axis decoration. Consistent number formatting. And it's going to give me something I could then edit. Now, is it going to be perfect? Far from it, but gets me a start. Now it's asking, hey, should the COVID instructions cover when to use a waterfall chart, color conventions, common pitfalls? And you can just keep going. When to use a combo versus two chart primary versus secondary, mixing chart type, avoiding. And then I just continue to keep giving it to it. And I gotta jump now, so I'm gonna go ahead and end this. I do have a meeting right at the top of the hour, but thank you for joining. I hope you guys found that beneficial. Just go in and start, start learning is the best advice I can give. Thank you, everyone.