Name: Best Practices: Claude for Excel and Claude for PowerPoint
Uploaded: 2026-03-18T18:32:44.761Z
Duration: 58 min 4 s
Description: Best Practices: Claude for Excel and Claude for PowerPoint

Transcript for "Best Practices: Claude for Excel and Claude for PowerPoint": Hi, everyone. We are live from San Francisco and New York. Welcome back to another webinar with Anthropic. We are super excited for everyone to be here today. We're gonna quickly we're waiting folks to join the webinar, we're gonna quickly start off with a poll. First things first, how much of your week is spent in Excel or PowerPoint? Taking a look here. Feel free to submit your answers. Okay. Let's see what the poll results are. Okay. So a bit of a mix here. Awesome. So we're gonna go through our next poll just quickly as we're waiting for folks to join. How many of you are using Claude today? Are you brand new to Claude? You've used it a bit. You always you already use Cowork. Maybe you're already an expert and wanna dive deeper in here today. Let's see. Alright. We'll give it a second more, and then we'll run the poll for the results and see how to gauge today's session. Okay. So we got a a lot of chat only. Okay. Awesome. I think that's a great place to start. We are happy to dive deeper into Cowork and Excel and PowerPoint and how to use Claude across the board today. So, again, we are live from San Francisco and New York. We've turned we've heard a ton of feedback, that folks wanted to go deeper into Skills, Excel, and PowerPoint. We know a lot of folks on the call are doing a ton of repetitive work within spreadsheets and PowerPoints. So we're excited to talk about a way to streamline all of that, and have Cowork represent this fundamental shift in the way that folks are operating, where you're able to delegate tasks to Claude, access a completely new level of productivity. We'll walk you through the toolset, and have Claude be able to take action on your behalf. So just a personal anecdote, when I got access to Claude at the beginning of this year when we were testing it out internally, I physically sat up in my desk, sat up straight in my chair because I was blown away. I had a light bulb moment where actions were finally possible within Claude, transformative and delegating tasks. If you view like, a lot of our colleagues on the engineering team had this light bulb moment when they use Claude Code. We feel that way with Cowork. So, our goal for you today is to see have that light bulb moment yourselves. Awesome. So let's kick off with some quick introductions. My name is Emily Gray. I'm on our go to market team here at Anthropic. I specialize in working with, our customers in financial services, and we've got some folks from our beloved Applied AI and product team who will quickly introduce themselves, and then we'll jump right into the content. Great. Thank you, Emily. Hi. I'm Dan Mason. I am on the Applied AI architect team. This means we are partners to, our customers, to help them understand both how to get the most out of Claude, how to use the surfaces that we offer, and how to ultimately, implement Claude into their workflows. I'm based, outside of New York, and I work primarily with, financial services and, some other retail and other industries customers. Hey, folks. My name is Nicholas Lin. I've met some of you before. I lead our product efforts for everything financial services related, which means I've spent all of my career in Excel and PowerPoint. I am what I call a recovering investment banker, private equity investor, so very excited to be here with you all and answer any of your questions. My name is Lina. I'm part of the Applied AI team just like Dan Mason based out of New York. I'll be monitoring the chat for any questions that you might have. Awesome. Thanks, team. We're really excited to be here today. We are gonna jump to a video and just quickly show you the art of the possible because we wanna anchor the team, here today on what you'll be able to accomplish after we run through the content, after we run through the demo, and and show you a bit about Cowork, in its light. As I as excited as I am in terms of the art of the possible within Cowork. It's truly mind blowing. Let's walk through it today. First, a couple quick housekeeping items. First and foremost, yes, this session is being recorded. We will distribute a copy. You'll get an email, after we hop off today. The second piece is, this is live. We have the team here able to ask questions. Submit your questions in the Q&A a tab, and then we'll both be answering them. And you can also upvote questions, so, we'll make sure to tackle the ones that get the most votes. So if you're curious about something, make sure to upvote that. And then the last piece here, please give us feedback. We're actually doing this session because there was so much interest from the team in terms of diving deeper into Excel and PowerPoint. So please share feedback. We're always happy to hear it. And then one last, housekeeping item. There is a, a page on our website, on our docs page that has a pop up at the bottom with, access to Claude. It actually scrapes all of our documentation, everything public facing that we have. It is the quickest way to get answers. We ourselves will do as best as we can to answer everybody's questions here today, but, go to ask, go to claude.com/docs to make sure you're able to get your questions answered there. Alright. So let's run through a quick agenda. We're actually gonna start off with a very fun special announcement. Then we're gonna start with some content around terminology, what is Cowork, what's a skill, how to work out of Excel and PowerPoint. Then we're actually gonna run through a live interactive demo. Of course, we're gonna leave time at the end for Q&A, but, again, feel free to ask questions. We'll try to break up the session and make this feel as interactive as possible. So what is the special announcement? Drumroll. Everyone live on today's webinar will be getting access to a free thirty day pro trial. So everything we're gonna walk through that has access to skills, to Cowork, to excel in PowerPoint, that's all available in a pro trial. You will get a unique, link to sign up for that trial, so take a look in your inbox later today to get access. The second piece is we're running a social contest. It's always really fun to work with customers and see what they're able to build, especially with something new like Cowork and skills. So, make sure to share what you build on X, we will be amplifying those winners, directly from the official Claude X handle. Okay. So let's get into it. What is Cowork?? Claude itself, we all know and love the chat interface that we've all used for years. Cowork truly is this fundamental shift in being able to delegate tasks on your behalf. So you can think of Cowork as a teammate. You wanna give it clear instructions on where it should pull data from. You wanna give it clear access to the tools that you're using and delegate tasks on your behalf. So, again, we all know and love the chat interface. This is really that step function, into being able to do things on your behalf. If you're familiar with Claude Code. this is so the the Cowork interface is Claude Code under the hood. So there's actually code being run for you. And, again, this is a a desktop agent builder that's able to delegate tasks. If you're not familiar with the word agent, we'll dive into that. It's basically just a way for Claude as a model to complete tasks for you. And then the other piece we're gonna cover today is Cowork anywhere. So the goal of Claude is to meet you where you're working. So we have heard, of course, so much feedback that folks are living and breathing in Excel and PowerPoint. And now with the add in functionality, you can work directly with Claude from there. And just also to kind of anchor us on terminology, the word Claude is used kind of everywhere. It's referred to the interface. It's referred to the models. So at a baseline level, Claude is a very intelligent world class large language model, but I might use the word Claude interchangeably when I'm talking about Cowork too because it's it's all in the interface. But if you have questions, let us know. Okay. So first things first, how to get started is, of course, you need to download access. We are available in desktop applications for both Mac and Windows. If you just go to claude.com/download, you can download it from there. We'll also include the links here. And please note, Cowork is in research preview. It was built at the beginning of this year. It is in research preview. We are aiming to be in GA relatively soon, so just keep that in mind. And Cowork is available on all paid plans. So whether you're on Enterprise, Pro, Max team, you should have access. Okay. So what is truly amazing with, Claude in Excel and in PowerPoint? This functionality has been around for a bit because it is in research preview, but the fundamental shift that we're seeing now in the last couple weeks is Claude is able to understand the context of the chats that you have within your Excel interface and your PowerPoint. So it's able to understand the conversations that you're having in both open files as you work. You can read and write directly in Excel. You can edit slides directly in PowerPoint. And now you can actually take those skills and use them directly in these two interfaces. So we'll talk through how to build a skill, but, the ability to do this directly in these two places is truly amazing. Alright. So, I know vocab might be a little boring, but it's really important to know each of these terms because we're gonna be talking about this in a lot more detail. So we've talked about co work. It's a desktop agent builder. What's an agent? An agent basically just means, an LLM like Claude taking action on your behalf. So, again, you're you're gonna be able to delegate tasks for Cowork for you. So we've talked through an agent. You can kinda think of it as a teammate who does those tasks for you. Now a connector. Some of you may have heard the term MCP. If you haven't, that's totally fine. It's simply just a connection to, read and write to the tools that you use. So whether you live in Google or Microsoft or whatever tools you're using on a day to day basis, Claude is a world class model, and it is, built by some very brilliant engineers here at Anthropic. But unless it has the context and the access to access to your systems that you use on a day to day basis, that's once you provide that access, it's only gonna become more intelligent and more powerful there. Okay. Let's talk about skills and plug ins for a second here because I wanna make sure everyone understands these two terms because they're they're relatively new in the space. So a skill. It's the same concept of a human skill. We all know how to write an email. We all know how to prep for our day. I'm not an engineer. I'm on our sales team. I myself have built a skill that is a call prep skill. Every morning at 9AM when my, computer is active and I'm online, hopefully, at 9AM, Claude has known to scrape my Slack, my Gmail, my notes, my data warehouse that has access to information, and sends me a direct Slack that is a comprehensive summary of everything I need to know about my calls for the day. Now that's just a really simple example. I mean, hopefully, everyone should have a call prep skill, so you shouldn't be, like, manually researching in the different places. Claude can just run that task for you, and put it where you want it, whether it's in a Slack or an email to yourself. That's what a skill is. So same concept of human skill. We'll we'll dive deeper into that. Now a plug in. We launched our part, our plug in marketplace a few weeks ago. This is the ability essentially to bundle different skills together and different connections into a package. So it can combine multiple skills, and we have a bunch. If you go to just I think it's claude.com/plugins, you can see all the ones that we have on the plugin marketplace that you can just take for legal, HR, operations. Whatever you need to use, you can use that as a starting point and have Claude instantly take on that role. Those are two important terms. We're gonna talk through those more in-depth. If you're on this webinar, I'm I'm assuming you know Excel and PowerPoint. Probably live and breathe in those tools. Again, Claude can, natively edit what's directly in those two places. And then the last thing we'll talk about is the Chrome extension. So it's effectively a a way for Claude to read what's going on in the web and help you do your research more effectively. Okay. We'll come back to questions on this, but, again, please put them in the chat if you have any points of confusion there. So just to kind of summarize everything that we've talked about in terms of the the terminology and how to get started, first things first, can't emphasize this enough. The the first place you wanna get started after you've downloaded Claude, you get access to Cowork, is set up your connections. Again, Claude is a very, very intelligent model. It's going to be even more intelligent when it has access to your system so it can run analysis and research for you. Step two is the fun part, and this is what we'll talk about today. What are the things that you do on a day to day basis that you can teach Claude how to remember? Whether it's call follow-up, whether it's due diligence, a screening report on a client you might be working with. Teach Claude how to do that on your behalf so that you can run the skill, and Claude can do it over and over again for you. Just another example here, like a a personal one. Whenever I'm working with a a customer and trying to share the light of what a skill is, I'll take their brand guidelines, and turn it into a brand guideline skill so that every output that we're creating has, you know, the specific colors and fonts specific to that brand. So I'd probably recommend that for for your company that you're working with downloading. If you have, a brand guideline template or something like that, you can upload it and build a skill off of it. It's a really easy one to get started. And then the last place is pick your poison. Where are we starting? Is it Cowork?? Is it Excel? Is it PowerPoint? Again, Claude will have access to understanding the context from both of those those places. And so, you wanna make sure that you have access to Claude in the first place and then are able to pull up the add in. It's just a pop up on the right hand side of both Excel and PowerPoint. So with that, we covered kind of the basics. I'm gonna turn it over to Dan who's gonna walk through these steps in a little bit more detail, and then we are going to, kick off the live demo. Excellent. Thank you, Lina. So I am now going to pivot over to my local demo environment. I wanna say a couple things before I do that. First, this is a live demo, so, I hope you have your fingers and toes crossed. I certainly do. But I think we're gonna be in good shape to not, make people dizzy as I'm switching around between different environments. But there is gonna be a fair amount going on here. So I am gonna stop sharing and switch over to my local, and we will get going. Okay. With any luck, we are all seeing the exact same slide we left off on. Looks, good. what I wanna start with here is is just to walk through what we mean by each of these steps. And as we go, I'm gonna do my best to give you a sense of, how to get things done in Cowork using these tools, and then ultimately how to take these things into the Excel and the PowerPoint surface. These things are all connected. And and, you know, in the Cowork environment, it is really connected in the sense that Cowork understands Excel and PowerPoint and Word and PDF and other file formats you work with all the time and is able to both produce them and edit them. And so Cowork is a great surface for when you wanna do things across, you know, a variety of those those different use cases. When you wanna focus, then go straight to Excel or PowerPoint. Right? That's a better place most likely to to do fine tuning and to evolve, you know, things that you already have. But I'm gonna try to give you a sense of both of those, and the first thing we'll start with is MCP. MCP is is just a protocol that we developed, in, I believe, the last eighteen months or so. Anthropic has donated it actually, I think, the Linux Foundation. It's an open standard now. But, fundamentally, what we have, developed is a way for large language models to talk to data, right, and to specify data sources that they wanna use. And so one is that if you do wanna use a certain data source, you should say so. Right? Claude will otherwise use whatever it has, and it will make the the best judgments that it can. And and Claude is usually pretty good about knowing what data lives in which sources. But you, as an analyst or as, you know, overall the the driver of this process, may have a better idea, and you should share that idea. Right? Giving Claude that direction will make it better at what it does. And, also, you wanna not overwhelm Claude. Right? If you provide connectors for everything under the sun, Claude may take a while to figure out which one to use, and it may not be the one that you know to be the best fit. So generally speaking, you know, these are just two best practices for setting up these kind of data connections. Very briefly, and I realized a lot on this slide, these are things that I think have been particularly relevant for us in the finance space, but but you can imagine that there is a a universe of connectors outside of these finance focused data sources. But we offer things that are both, you know, in the productivity space. That's what this left column is here, with GitHub, Google Workspace, Microsoft 365. The connectors in the middle there really are, you know, professional financial data sources, all partners of ours. And then you can imagine that there's also a world of of custom connectors. These could be things that you develop with Claude Code. These could be things that an enterprise builds, you know, and deploys and and manages themselves. Right? There there's a variety of different ways that could work, and there are also public MCPs that might be of use. So, what I'm gonna do is I'm gonna show you what this looks like inside of the Claude application. Right? So in this case, I'm sitting inside Cowork. We'll get to this use case. This actually is the one that I wanna walk through with essentially, around an earnings update for a fictional financial services firm. But I wanna start with the customize tab. So this is actually gonna cover cover both skills, connectors, and and plug ins. So connectors is what we mean by MCP. I'll get to that in a second. But these are all the the additional ways that you can provide context to Claude. Oh, so I see our chat popping up there. I will ignore those questions for the moment, but, Emily Gray can pop in with them if you'd like. So the basic outline here, though, is that these are all things we're providing, Claude, in in addition to its basic knowledge, in addition to the data that you're providing. And so let's go through a couple of these just to see what's here. And I'll cover skills a little bit too because it'd be relevant for what we're gonna show. Skills are things that either are gonna come to you from Anthropic. We we provide a handful of skills that are available by default. It's also stuff that we can preload with our specific corporate knowledge. So I've got some custom built skills here for Silvern, which we'll talk about a little bit. And so all of this is available to Claude. You can expressly invoke it. You can have, like, a slash, you know, Claude Code or slash slash command like in Claude Code. or these are things that Claude can just know are available and use to its own discretion. Right? If you're saying I'm doing an earnings update and it knows you work for Silvern and it's probably gonna pick the Silvern and earnings update skill. Right? So those are both possible entry points for that work. You also can then collect these skills into plug ins, and connectors can also be part of plug ins as well. I've preloaded a handful here, you know, for Silvern because, you know, imagine Silvern's a bigger company. Right? There's more than just, you know, my function, which might be financial analysis. There are plug ins around, you know, legal marketing. It might be that I need some of these things to do pieces of of my job or to feed something to those teams, and these teams obviously might focus on them for their own purposes. And I can also enable and disable these. Right? So I don't have to be using all of these plug ins and skills and connectors all at once. Right? I have the flexibility to turn them on and off as appropriate. And you can see I've disabled some here that I think are not relevant, you know, to to what I'm doing in my job. And so, you know, that packaging is a way for you to, again, collect this context and collect these capabilities in such a way that Claude can make the best use of them. So now for NCP and for for the connectors that we have here. So I've enabled one specific connector for the work we're gonna do here, is S&P Global. I know that I can get the data that I'm looking for in this example from, S S&P's Kensho service as an example. But you can see there's a variety of other things available here. Again, a mix of productivity and design and and other things. And so this setup is gonna come to you clean out of the box. You you will have to enable and will it will want to enable the things that you want to connect to. You know, Claude, on its own is gonna suggest some things. It'll help you set these things up, but this will all come as a clean slate when you start. So, that's MCP configuration. So what I'd like to do now is actually just show you, and I think I've got this still queued up, what it would look like to do something like this using these this context and these connectors, with Claude. So, I've got a a use case here that is very similar, I think, to what a lot of, you know, equity analysts do. And in this case, you know, let's say that earnings have just come out. It actually came out in late January for Apple. But earnings have come out for Apple, and you can see, you know, that there's an update for fiscal Q1, which is actually the last three months of 2025. I'm gonna direct Claude to use a skill that I've provided. Right? In this case, a workflow for actually running earnings updates the way that Silvern wants to do it. Right? So I could say, you know, you know, give me an earnings update, and Claude would do its best, you know, to to sort of just conjure up what it thinks would be appropriate for your situation. But you can give it this kind of very structured direction, and Emma will walk through what that does. I'm also providing it some input data in the form of a model in this case, which is basically last time's model. Right? So this is a financial model that I had, you know, from the last time that I did this work. Right? And I've got, you know, things like income statements and, you know, overall balance sheet and cash flow. This data is not necessary for Claude to do the earnings update because it just needs to go and pull new data essentially, but I want it to actually modify this for me. Right? I want it to bring this model forward. And so that whole process is something that's also encoded in the skill. How do you do this, right, you know, is is one thing that we're trying to encode there, and we'll get to sort of how skills, are built in just a second. But I'm providing all this as input, and then I'm saying, you know, to Claude, use Kensho data. Right? You know, that that's that's the data source I want you to use. I could have said, search it from the web. You know, the the data is largely gonna be public, but I want it from an authoritative source. And I could you know, and I want it cited in this case. Right? I wanna know where this data came from. So that process, right, is similar, I think, to what people would do, you know, with working working with Claude in a chat context, but we're gonna be able to accomplish a lot more in the cohort context. Cohort is gonna be able to create these things and run through this entire workflow in a way that would be very, I think, much more difficult to specify without skills and without connectors and without cohort. So I'm gonna kick this off. And I've already got this prerun. I'm not gonna force us to sort of wait and and watch through this whole thing, but you'll get a sense of what's going on here. Claude is now exploring the question that I've asked it and actually looking over here at these skills. Right? So it loaded this silver earnings update skill. It's now gonna look at the files in the folder that I've given it. Right? It is now gonna ask me some questions. Right? This is one of the best parts about this flow is that Claude can clarify before it goes off and does all the work. Right? In a chat context, Claude might just stop and sort of say, hey. You gotta answer these questions before I keep going. Now we've got a structured way to answer all of these things. So you can see here, I've got, you know, a a question about, which quarter. It's like, yes. You're you're correct. The fiscal Q1 is ending December 2025, so that's the right quarter. I do want you to run the full workflow. And, again, this is specified where I defined this earnings update workflow to do a beat miss analysis to roll forward the financial model. Emily, go ahead. Yeah. Yeah. Sorry. I just wanna pause here because there's a a question that is relevant for this workflow. When you're building this skill and you're giving a Cowork a task to complete on your behalf, what's really amazing is Claude will actually pause and ask you these clarifying questions. So it will make sure that it's building things appropriately in terms of, like, whatever the output is, where it's pulling the data from. So you're not just, giving Claude a task and it's just running it on your behalf. You're able to to see how Claude is thinking through how to complete that task, and then it's asking you the clarifying. question. Absolutely. So I just wanted to to double click on that. is intentionally interactive with you. Right? If if this was something that you wanted to run, you know, entirely headless, you could start it and walk away and and tell Claude, you know, go until you're done. But I think it's very much a better pattern to work with Claude. Right? Treat Claude as a Coworker. Treat Claude as a partner. This is also true for coding. Yep. Right? In coding, we call it pair programming. This really is the, you know, tactical equivalent for knowledge work. So I'm gonna say, yes. I want you to do this whole workflow, which is gonna generate a handful of different files. And, you know, I I could give it some directives here, but I'm just gonna say the standard KPIs are fine. That is enough for Claude to keep going, right, and to ultimately get through, you know, my entire workflow on my behalf. So this would probably run, I'm gonna guess, maybe for ten minutes. What I'm gonna do is I'm gonna show you a completed version of this so we can look through what it actually looks like when it's done. So you can see a very similar, you know, beginning here. Right? Not exactly the same things. Right? In this case, it asked me about my rating and price target. So, you know, Claude may be curious about different things depending on, you know, the specific circumstances that it it finds. But then when it starts running through stuff, this is where, you know, Claude is running code. And and you don't have to know the details of the code. I just think it's important for you to know that it's in here. Claude is actually writing scripts to read my financial model. It is then, you know, talking to itself and saying, well, this is what I think I know is going on here. It's then giving you back, you know, essentially some some continuation of of its thought process. So as you're working here, you don't have to be in the details, but they're there, right, if you need them. Now it is able to then go and plug into my connectors. So you can see here the S&P Global connector is available. You can see there are errors. So Claude tried to do something here that actually failed in the moment. Right? It tried to get some line items and statements. Now Claude correctly analyzes that, well, it turns out this data isn't available, right, in this particular way that I was trying to get it. Let me try a different way. And now this time, it succeeds. Right? You can see that these results aren't aren't red. So now I've got the data that I need at a high level, and then I'm gonna go and I'm gonna actually do a combination in this case. I think of can show data polls and consensus earnings estimates from a web search. Web search is something you can enable or disable as you see fit. In this case, I I'm happy to have a combination of the two. But all of this is is coming from this context that I've allowed Claude to gather. Right? I gave it MCP connectors. I gave it the workflow that I wanted to run. It is asking me when it needs clarification, but otherwise, it's just doing this stuff. Right? The this is this is interactive at a higher level than what you would consider to be kind of the command and control where, you know, you have to put everything in the right formulas. You have to go to the web searches yourself. You can give Claude the ability to do that on your behalf. So this is gonna go on for a little bit while just to sort of get the the three phases going. Right now, it's doing this model roll forward. It's then gonna do a a report. Right? Actually, building out a word document. When the three phases are complete, you'll both get a little pros. Right? You know, Claude giving you this update over here, and you're gonna get some actual files to look at. So I'll I'll shrink these down a little bit so you can see what this looks like. So this is all available for you to view and iterate in Cowork, but it also mirrors what's going on on disk. Right? These files actually exist. Right? So if I go over here, I think I've got this sitting in another folder. These files, right, you know, are the same files that I was just, you know, showing you in Cowork. Right? And so, you know, this this is mirroring what's going on, you know, in your local environment, and you can continue to iterate these things here. Right? So now comes a bit of a fork in the road, and I'll jump back into to sort of talking about how to think about that. You've made these new files. You've put them into your local environment where you can work with them, and you can choose to either keep iterating them and and and working with them in Cowork. Cowork is capable of doing that, but you also could switch into these other surfaces. In particular, Excel is what we're gonna talk about first. Yep. And and just adding to this because I I saw a question in the chat, so I just wanna also clarify. Claude and Word doesn't exist today. But as you can tell what, Dan. just built here, you're able to export it into a Word document. So it'll it you can either download it as a PDF file, take it directly to Microsoft Word. You can upload it into Google Sheets. or sorry. Google Docs if that's what you use. But as of today, it's just download only, Absolutely. And and, but, great question reason. that I think Excel and PowerPoint are particularly compelling is how tricky they are as formats. Right? And I don't mean to to, you know, talk down about Word. Word documents can be very complicated. But the idea of the structure and the formulas and the and the the various connections that happen inside Excel and the visual nature of PowerPoint. Right? Those are two really interesting things, that Claude has has gotten very, very good at managing. So all that said, I'm gonna jump back over here for a second into the deck and, you know, just remind us that what got us here, right, was this ability to connect data, to, connect skills, which we're gonna talk about in in more detail in just a second, and to integrate all of that into a a request to Claude. Right? Not a series of steps, you know, that that Claude has to follow that's encoded in this in this skill knowledge, But just a request, hey. Here's what I want you to produce. You know, here's some resources. So let's talk about skills. What I wanna do now is is talk you through what it would be like to add a skill to the collection of skills that I've already got with this workflow. Right? So I got the result that I showed you, with the earnings update with a a collection of skills, making a beat miss misanalysis, rolling a model forward, generating an earnings note. Let's say that, you know, I'm looking to do something more, and the thing I'm gonna gonna show you is how to generate an earnings update deck, so we can bridge it to PowerPoint. A few things to note about skills at first. Claude will use these skills in whatever surface you provide. Now there are some skills that are gonna be fairly specific. Right? Some skills say, make me a PowerPoint. Right? And so if you ask it to make a PowerPoint in Excel, Claude will do the best it can, but that's probably not a great match. If you say, Claude, I want you to research, you know, this topic and, you know, I want you to, you know, produce me, you know, a sort of the the the most well reasoned, well sourced, you know, kind of output you can get. That will have applicability in these different surfaces. Right? Claude will will either make choices or ask you for for your preference, or it will do the best it can in the environment that it has. And as you'll see, these things can also apply across surfaces. So we we can run a skill in one surface and have it actually, execute in another, either from Excel to PowerPoint or vice versa. Now skills do work better in certain surfaces. Right? So that's definitely another thing to note. And so what we're gonna do here, and I'll jump back out of the deck, is to show you what building a skill looks like. So in Cowork, and this also applies to Claude Code for those who are especially adventurous. In Cowork, what this looks like is the same kind of request, but now we're actually giving it directives to to make a skill. And I'm gonna say create a new skill called Silvern and earnings deck. It generates a PowerPoint and news deck from this data. Now I'm giving it enough context to understand, a, the workflow I just ran because I see this as an extension of the workflow that I just ran, and I'm gonna give it some additional resources. Right? So I'm gonna say, the skill should do all of these things and and create this stuff, but I've got these brand reference materials. So let's take a look at what that looks like. So I've provided a template, right, for an earnings update. Now this is pretty specific, and you'll see what the output we're gonna get is gonna be very, very similar to this. Right? I could provide something a lot less specific. I could provide just brand guidelines, which I actually have done in this brand style guide document. So here, I'm giving primary colors and giving accent colors, you know, typography, things like that. So both of these things are inputs to the skill that I wanna build. Right? And, you know, most companies have these. I assume your company probably does too. You drop these in, and you tell Claude, that's what I'm looking for. Right? Just match what I've got. Here's a handful of high level things that I wanna include in this. Right? But, you know, fundamentally, that's what the skill needs to do, is to take, you know, whatever the actual data is, whatever the actual earnings report you're running for whatever company, for whatever period, and create a deck that goes with it. And so there's a couple things to note here. One is that we have this skill creator skill. We could talk in an an entire hour about this, but at a high level, the skill creator is kind of the meta thing that sits, and I'll just go to the very top here, the meta thing that sits on top of, this entire process. The skill creator is a way for Claude to produce skills that will ultimately meet your needs. Right? And that will encode the business logic and the knowledge that you are interested in encoding. It also is going to help you test them. And so that's one thing I wanna point out here is that creating tests for a skill to see how well it works is is an important part of the process, and it's also part of how you can iterate the skill yourself to get to the result that you want. This is great, Dan. Sure. Just wanna pause to clarify a couple questions here. So I think just the first question, we we wanna simplify this. One of the questions is how is a skill different from prompting Claude to, like, run the same set of instructions? I just wanna call out that a skill is a way that you're teaching Claude to. remember how to do the same task over and over again. So you'll give it a name, like a brand guideline skill. And anytime you prompt the brand guideline skill, it will remember all of the instructions you gave it in this markdown file, to run that task over and over again. So I just wanna clarify the the difference there. And then we we did have a question here. Is a skill output reproducible? If I run the same skill on the same connector and the same data, would the result always be the same in. the same structure, And that's, a great, question. The answer? to that is most of the time, yes. Because Claude, especially as it's gotten more intelligent, is much better at reproducing, you know, especially very specific templates and workflows. But not exactly. Right? You saw in the the the example that I just showed around building the earnings update, I got slightly different questions. Claude was interested in different things to try to fill out, you know, the earnings update in the correct way. You will get some, you know, differences in possibly the data that it finds and searches, right, either from approved data sources or from web search. So there is definitely a possibility of that kind of difference. And at the same time, for most well understood workflows, especially with good evals on the skills, you can get very, very close, right, each time. Right? We do think skills are are largely reproducible. Yep. And I would say that Claude as a as a model is very good at following instructions. So same thing as a Coworker. You wanna be very clear on the instructions you're giving Claude, where to pull data from, what you want that output to look like so that. you can teach it how to remember. Okay. I will continue. So you can see here that what's happening is the same thing as before. Right? Claude running some code, reading some some skills, reading some some elements that it's gonna need. It's going to look at the documents that I provided. Right? So it's looking at this brand style guide, running a script to actually extract that data. And now it's gonna do its best to create the script and then test it. Right? And it's gonna look at actually converting to images, right, and actually visually looking at these things. Claude, it does have vision and is able to to see things that it's able to produce. You can see here Claude then finds some errors itself. Right? So there is some amount of, in the skill creator in particular, some amount of logic that helps Claude do a a first pass of this by itself. It's not, you know, just putting out something for you that you have to correct all the errors. It's gonna do some of that correction, you know, on its own. It fixes formatting issues. It, you know, changes some visibility issues. So now you've got this other idea of the eval viewer. I'll show you this just as an example to to give you a sense of the way that Claude is thinking about this. This is something that then as you build more skills, you you can also use. But now I've got this skill. Right? And now I can see both that it made me an example deck. And, actually, I'll go I'll show you the eval viewer first. That's probably better. It also gives you a sense of, like, well, hey. This is what, you know, I did to generate these things. Right? I generated, you know, this this template. Here are the numbers that I plugged in given the data that was provided. Right? So this is a combination of filling in gaps and reasoning over the the overall request and producing, you know, what ultimately I think is a a pretty reasonable deck. Now this is also something that I can immediately give feedback on. So what I did here is I actually said, hey. I like this one, but let's make two changes. I've got this price target here. I just kinda wanna see it in a different Right? It it it, you would pop a little bit if I saw that in red. And in fact, I like the order a little bit better if I push the financial highlights above the beat miss summary. That that's just a preference. Right? You know, that didn't have to be the way that we did it, but, you know, I thought twice after generating the skill in the first place. You know, that's that's my preference. Claude will then regenerate the skill and the deck, right, and now give you a better look at, you know, what that new layout's gonna be. So what I should see here, I hope, is, yep, a red price target and a different order. Right? So these kinds of changes are are obviously things you could also do. Right? But in this case, we're encoding Claude with this knowledge. Right? Because what we really want is, again, for this to be reproducible. So then I can say, that's great. Help me install this skill, right, so that I can use it. And so I have this button here that actually would copy to my skills. And if I go back to the customize tab, right, I'm gonna see now under skills, I've got this silver and earnings deck skill. So this is just closing the loop on me sharing my knowledge with Claude and trying to encode these workflows that are gonna save me time. Right? And, you know, in this case, it's very specific to this use case. You know, this earnings deck depends on the other things that are going on here. You can also imagine much more general versions of it. Like, hey. You know, brand guidelines is a great one, but, you know, there are other cases where you might have a skill that you could use across a variety of different things. This is this is really something for you and your organization to think about. You know, what are the things that you wanna encode and distribute, and what are things that are just gonna save either individual people or individual functions, you know, some time. So, all of that said, now I've got this collection of skills. I've got this collection of outputs. If I go back here, you can see at the bottom. Remember, I got these things generated plus this new deck. Now what I'd like to do, if I go back over here, is to talk about the surfaces. So working across apps. The way to think about this is really just that, you know, you should pick your surface. And then if you have things across multiple surfaces, open Claude in all of them, and and they can connect. And there are some particularly interesting things that that allows you to do. So, I will now show you what it looks like to be in Claude in Excel. So this is essentially the same thing that I produce as as part of the skill output. Right? And what I'm gonna show you here is that I got the roll forward that I wanted, so it actually correctly filled out Q1 twenty six. But, hey, I don't like this formatting. Right? This formatting is strange. And it was doing some estimates, and, you know, obviously, there's some things here that it needs to clean up. So I can do some very basic things, again, with Claude as a as a as a pair. Right? And I can just say, hey, Claude. This doesn't look right. Check out the formatting on the right side of income statement. That's not telling Claude exactly what to do. I'm just pointing out the problem. I could have been more pedantic, and I could have said make, you know, the bars and lines eight and thirteen blue. But Claude can see that part. Right? And so a lot of what we're depending on here is Claude's ability to actually interact with the document. Right? And so you can see here, Claude looks at this and says, yep. I see that columns m and n. Right? These two right here have inconsistent formatting in a bunch of ways. You know? And here's what's wrong. This is an existing document. Right? I I didn't give Claude some template. Right? I gave it the actual thing. And by giving it the actual thing, I'm enabling Claude to actually interact with the specific state of this and then and do it in a more limited sort of targeted way. So I can just say, yes, please. Those fixes sound great. And Claude is now going to, you know, do its best to to get these fixes right. Now if it's still not right, I can ask for more changes or I can change it myself. Right? I I have that option. Right? I am not forced to go through Claude for all these changes, but frankly, it's just kind of a better use of time. Claude is gonna be faster at this than I would be. Now some of you, I'm sure, could differ with that. I know there are some serious Excel jockeys on the call. But, fundamentally, this is the kind of thing that you can use Claude and Excel to do at the at the sort of very lowest end, and you can do much, much more interesting things, right, as as you sort of, you know, think about, more complicated use cases. And as long as Claude has access to the data, access to skills, access to the connectors that it needs, it'll be able to do the same kinds of things you just saw in Cowork. So you can see here the formatting here, you know, looks basically right, and it's actually pointing out why I left the yellow yellow, you know, for a couple of reasons, and, you know, we can agree or disagree. But now I wanna point out two things, and I realize we're getting to a time we wanna open up for q and a. I have skills here. Right? The skills that exist in Cowork and exist in my Claude environment are also available here in Excel and in PowerPoint. Right? So if I wanted to run, as an example, the roll forward scale, let's say that I had this model before, you I'd actually done the roll forward and I wanted to use the scale to do that. I could run silver and roll forward from here. Right? And I would be able to see my connectors. And in this case, I've got the S and P Global connector here just the same as it was in Cowork. Right? So the surface you're in doesn't matter as much as those things being available in your Claude environment. Right? You you can use these wherever they're appropriate. But I do have that ability to run skills. So that that's one important part. And I also have the ability to talk across documents. So let's take a look at PowerPoint. So you can see here, this is the earnings deck. I'll blow it up a little bit, as well. This is the earnings deck that we looked at inside Cowork. And one key thing here is that I've got this little connected files. So you can see the connected file here is the Apple financial model. Right? So I can now see through to Excel. And if I look here, actually, at the top, I can see that my connected file is the Apple deck. So what I can do is you know, and there's a few different ways we could make this happen, but I can say from the deck, hey. I noticed that in the PowerPoint deck connected to this one, I think there was actually something missing here. So if I look at these financial highlights, I'm missing the actual graphs. Right? It just turns out that, you know, I didn't get those filled in in the first place. So but the data, right, for these graphs for revenue and earnings per share, guess where it lives? Right here. So what I can do is I can say, you know, the slide that shows revenue and EPS graphs is missing data. Can you fill in those graphs using the data we have here? And so now what Claude's gonna do is it is going to connect to that PowerPoint and, in fact, go over and look at it. You can see ran code here is just a way of saying, essentially, Claude is now gonna start talking, right, into this PowerPoint deck. And let me just double check that it's actually going. Yep. So it looks like it is building it's basically pulling stuff over and then trying to share with the PowerPoint agent. Right? So Claude can do this back and forth across these surfaces. You know, the details are really up to Claude. You can make certain directives about how you want it to work. But, you know, fundamentally, now I can see over in the PowerPoint. I've actually got this, you know, request. Right? Excel is asking, find the slide that has revenue and EPS charts. Right? Because we're gonna replace those charts with this real data. And Claude's gonna have to think about this a little bit. Right? You know, it may look at some visuals, and there you go. It's found the slide. This is the kind of thing. Right? I could have done this in Cowork as well. Right? That that that is possible. But I kinda like the idea. Well, I'm working on these two things. These are the two things that I need to make perfect. I'm gonna do some of it. Claude's gonna do some of it. And the ability for Claude to talk back and forth across these files is a really impressive and important capability. So I'm gonna pause here. We have about ten minutes left. I know others have been monitoring the questions. I have definitely not. Is there anything, Emily, that we should surface or dig deeper on? Yeah. Well, first of all, thank you so much for running through this. Just one thing, we have never had a webinar this large. I think there are over 10,000. people live here, so I think we broke the chat. But I can still see some of questions that are coming in, so I'll make sure to touch on those. Just for a second of humor, Tony asked if we are. real people. I'm obsessed with that question. We are. My I went to my, grandmother's nursing home over the holidays, and she was telling her friends I work in AI, and they asked me if I was. a robot. So love the question. We are real people. The number one question that Dan asked, out of NCP connection skills plug ins, are they secure and safe, or is it an open marketplace where other people can create and add their. own features that may or may not be trusted? brief answer on that, but, Nicholas, feel free to jump in also. The the ecosystem around, what we expose in Claude is is very much trusted partners, you know, things that Anthropic itself exposes and and builds. We build these as references. So, you know, there there is a lot of reason for you to make your own customized versions of some of the plug ins and and, you know, skills that we offer. But, fundamentally, yes, there's also an open marketplace. Right? So if you were to go off and add NCP connectors of your choice, you know, those are things that that then, you know, you you and your IT groups are gonna have to work out. But we do control what is available, essentially out of the box inside Claude. Then I think the beauty of MCP is that it is an open standard. So Cowork and Claude Code can create MCPs in a matter of minutes to hours. So if you see an MCP server that you would like and is not live today, anything that has an open API that you can access, you can tell Claude to wrap it up in an MCP server and help you turn it into a local or remote one. So that's really the beauty of it, but we work really closely with a large group of our ecosystem partners across many industries, life financial services, life sciences, and others to really create really high quality better ones, and those are the ones that show up on our connectors directory. Awesome. Thanks, guys. We'll do another one. Is it better to complete many simple skills, or is it fine to complete one complex skill? For example, is if the execution is 80% the same, but there are nuances, company analysis across industries. Is it better for one complex with multiple industry archetype or break it out into definitely have, I think, skills? some materials we could probably share on this also. But but in terms of the general, approach, you know, things like brand guidelines, right, are are clearly things that are applicable to, you know, many, many use cases. You know, things like rolling a financial model forward may belong primarily inside of a larger workflow that's very specific to one team or one group. Right? And so a lot of that is just how many how many people will use this? Why will they use it? You know, how much do you want a perfectly standard version of something versus for, you know, individuals and groups to to have their own flavor. Right? Because, you know, we we know, you know, in financial services, you've got a lot of different ways that different groups like to do their models. Right? We wouldn't wanna push out a skill that removed that freedom of choice. Right? We think this needs to be a coworker for you in whatever the context is. We can provide these templates. Right? So then you can customize, then hopefully, that's sort of the best of both worlds. Awesome. Next question. This is an important one we get often. Where does my data go? Is it used for training? Do I need to strip client PII before using this? Yep. Nick, do you wanna jump in on that one? I I can certainly give a couple of blanket answers, but please. Yeah. For sure. So, we have a very robust trust center with all of our privacy articles, and the treatments are a little bit different, between the consumer and the enterprise policies. So I would really make sure that you all look through the trusted center articles very closely there. But in general, we take safety very seriously. This is really our entire brand at Anthropic, so we want to make sure that all of your data is extremely safe with us. Great. I just, linked the trust center in the chat here. Awesome. Let's jump to so we covered a lot within Cowork, Excel, PowerPoint. What would you guys recommend? When should I use Cowork versus the Excel PowerPoint add in versus a project versus Claude. Code? I think there's some nuances here, but. That's that's a great question. walk us, through best jump in. there. So, the the the the difference between Cowork and Claude Code is primarily user comfort. Right? They do many of the same things. And you can expose, you know, the same notion of Claude Code with the appropriate skills can make Excel files. Right? Now a lot of this is just, you know, if your role does not involve coding, if that's not your primary surface, you will probably be happier and more comfortable in Cowork, but they are functionally very close to equivalent. Excel, PowerPoint are a little bit different. Right? Those are add ins in the context of, individual Excel workbooks and PowerPoint decks. Right? Those have, you know, a different approach, you know, and and fundamentally a much more efficient and faster approach in a lot of ways to making changes to those files versus what Claude Code or Cowork would have to do. So a lot of it is really your comfort combined with how broad the change you wanna make is. Right? It is across services more than Excel and PowerPoint. You know, is it are you making 10 Excel files? Right? There's a bunch of reasons why, you know, using Claude Code or Cowork would just be a better way to do those sorts of jobs. But if you are working inside, you know, a a couple of files, Claude for Excel and PowerPoint can be really, really powerful. Yeah. I think my mental model here is that are you working on open files, or would you like for Claude to work on a bunch of, you know, save down, close files in your file directory? So, you know, if you're working on open files, it's probably that you're spending a lot of time on these files directly, and you want Claude to really be an iterative collaborator with you so that you can make some changes, Claude can make some changes at the same time. So that's where we've seen, you know, Claude in Excel and Claude in PowerPoint be really powerful. You can see all of the changes that happen, as they come, and you can also ask Claude to validate and explain all of his actions as well. Whereas, Cowork is a really a ton autonomous agent that is able to reach across many files on your local desktop. You you know, see it work for a couple minutes and then produce the the results for you, which is fantastic very often for a first draft. But for any sort of deeper refinement and iteration, Yep. I I also realized I didn't totally answer the project question. PowerPoint. Projects are primarily, at this point, a chat construct. So the main difference there is that projects are for collaboration in the Claude chat environment. Whereas with Cowork, Excel, PowerPoint, Claude Code. right, there there can just be other ways of distributing essentially the the outputs. And so you should think about it as, you know, if this is something where, you know, my primary output is something that I am producing and owning and sort of putting into another system of record, then Cowork and Claude Code are great. If you need to collaborate in either real time or close to real time with others, right now, projects is a good answer for that. Great. Thanks, y'all. Another popular question that we're seeing is, does this work within Google Sheets, slides, docs, etcetera? Not today, but these are obviously really important surfaces, and, Anthropic, we build really quickly. So we're excited to think about what is to come. Yeah. And I would add to that. I use Google Slides myself. The workaround right now is just to download whatever Coworker is able to output for you and then just upload it into Google Slides. Actually, I have another workaround, which is Google. Drive better. has a application locally so. that you can sync all of your local files with Google Drive directly. That's actually primarily how I use Claude and Excel and PowerPoint today, and the same files can can be collaborated on and shared with your Coworkers directly online as well. And, there was a related question. Our applications, Claude Excel and PowerPoint, don't just work on the desktop. They also support the online Microsoft three sixty five version as well. Good point. Thanks, Nick. I knew you'd have a better workaround. Awesome. This is a great question from Pedro. Can you call skills within skills? I e, can you, quote, unquote, use a brand template for formatting, quote, unquote, within a wider presentation is yes. a skill? And let me just show really quickly. I think it's this one. So you can see, you know, this skill, sort of definition here. In fact, where is it? Oh, you know what? I think I did it differently in this one. But you can explicitly say, and this is actually just generally part of how progressive disclosure and skills work. You can specifically say, this skill needs to call these other skills. Right? You will need the Excel skill. You will need the silver, you know, earnings analysis skill, right, to complete your job. So, yes, skills can be collections of other skills. You can either do that explicitly, or you can do it by telling Claude that its job is to to accomplish a certain outcome and then giving it tools. It may it will also potentially use those skills on its own. Great. Those are very exciting. Let's do one more question. This is very common, and I I wanna make sure that everyone understands this. The question is from Jack around, is it best practice to use Opus or Sonnet or Haiku with an Excel, PowerPoint, and even in Cowork?? My recommendation is to try to use Opus as much as we can for Excel and PowerPoint workflows. The background here is that these are fairly new capabilities to the market. We've started training on them in the past couple of months, and we've seen these capabilities get really, really good. But Opus right now is our best model for both of those services. Great. Awesome. Well, thank you everyone from the Anthropic team for joining, and thank you for all the folks live on the webinar or if you are watching this recording. We're we are super open to feedback. If you have other sessions you want us to do or if you have questions, we can lean on our dear friend, Claude. Again, you will get a free trial. If you were live on today's call, we will send you a link. And then it's so fun to see what customers built. So, again, if you wanna post it onto x, we will pick some winners and repost them from the clawed handle. But thank you so much everyone for joining. We really appreciate it. Please. share your feedback you all. other questions. Thanks, everyone. Thanks, guys.