Build a Competitor Research Bot with Claude Cowork

Hello everyone and welcome to today's session. My name is Rhys and I'll be your moderator for today's session. We are going to get started very shortly. We're going to start slightly late today, uh, just 2 minutes past, just to give everyone a chance to have a look at the resources document. So you'll see in the video description we have a link to a session resources document. I will also pin this to the chat very shortly. This is going to contain everything that's required for today's Code Along, so that is your complete source of truth for today's session. In there we've got links to the slides we'll be covering at the beginning, we've got links to a couple different files, including a zip file called messy downloads and two markdown files, as well as a couple of prompts that we're going to be using. So that is pinned in the chat right now. That's going to be everything that we need for the session today. So please do check those out. Obviously, we're going to be using Claude today. So if you don't already have that going already, then please sort that out now. However, if you would just like to watch along and then catch up with this in your own time, feel free to do so. You can scan the QR code to register for the session if you haven't done so already. That means I will send you the recording as well as everything else required for it, and you can catch up in your own time. So yeah, if you have any questions or anything else like that, let us know in the chat. We're going to get sorted very shortly— get started very shortly, apologies. But yeah, please do check out the resources document. And, uh, yeah, have a look at the files that are, uh, the links that are in green. They're gonna be, uh, critical to, uh, the code along today. Brilliant. I think with that we will get started. So without further ado, I'll hand you over to your host for today's session, Richie. Richie, please take it away.

Hi there, Data Scamps and Data Jumps. Yeah, welcome to the Anthropic Age. Uh, I'm very excited for a whole week of sessions around making use of Claude. So Claude is the hottest large language model around right now. So we've got a whole week of sessions to help you use it. And today we're going to be looking at Claude Cowork. So Claude Cowork is designed to help you automate repetitive tasks. Unlike Chat, Cowork lets Claude complete work on its own. Really, I like the idea of not doing any work myself, having the bot do it, I get to go on the beach, lounge around in a nice shirt. So we're going to be doing a fun and also useful case study today on automating competitive research. I don't know about you, I have to do competitive research, see what other companies are doing with their webinars. I feel kind of dirty going on the competitor website. So better to have a bot sort that out. And I get to focus on doing the fun stuff like chatting to all of you. Our guest is Maarten van den Broek. Who is a learning solutions architect here at DataCamp. Welcome, Martin.

Hello there, Richie Cotton.

Yeah, great to have you here. So Martin provides live training to our enterprise customers, and previously he was a senior content developer also here at DataCamp. So he's responsible for creating much of DataCamp's Power BI and Tableau curriculum. He also ran DataCamp's competitions. If you've not checked out a competition, then we can post a link in the chat. They're a lot of fun to take part in, really show that you have mastered all the skills that you've been learning in the courses. Beyond his day-to-day work, Martin is also a co-founder at Conforte Symphonic. So for those of you based in Belgium, you like classical music, go and check out one of his shows. But in the meantime, you can check out his show here on, on, on YouTube, LinkedIn, and Facebook Live. Please take it away, Martin.

Thanks so much, Richie Cotton, for a great introduction. Again, like the shirt, as always. It seems you're already kind of fit to go to the beach whenever you have automated all your work. That's a good first step, I would guess. But for me, unfortunately, I still have to do some of the work. So hopefully you will learn some things today, or you will see some things today in action that hopefully inspire you to get this applied to what you do, whether it's in your professional career or it's in— when it's a hobby project. Me personally, I learned a lot of doing AI and prompting and agent, um, engineering while working on these, uh, of some of my hobby projects. I guess there's less of a risk compared to a corporate setting, but I think for these projects where you work really on hobby-related topics. You care for these projects, you know what you want to do. There's often a lot of things that you have to do repeatedly. Here, AI and agents are really useful for you to get that over the line. Today we'll not be talking about some of my hobby projects, but we will be building a competitor research bot, something that might not always be useful for all of everyone here today, but But rest assured, at the end, we will see a lot of similarities and parallels with work that needs things to do repeatedly. So whatever you learn today will be applicable to most of the things you will have to do over and over again over time. For those who have actually taken DataCamp courses before, you might recognize my voice as I have been the instructor for several of them. But today I'm here to give you a live session and a live class on building a competitor research bot. Before I really get started, I just want to gauge a bit of the interest of the audience right here. So I just want you to quickly address in the chat box that you see here. I want you to answer A, B, C, or D to the question: how often do you actually refresh your competitor research. Competitor research is obviously something that needs to be refreshed at some scale because the competitive landscape is changing over time. New competitors appear, new things change on the market, customer demands shift, and so on. So you probably need to at some point refresh your competitive research. So in the chat, I just want you to quickly say A, B, C, or D Where A is a weekly refresh, B a monthly refresh, C a quarterly refresh, or D you panic when someone asks you. So let me just check quickly, brief over the chat to see. I see we have a lot of Ds, some Bs as well. Some people never refresh, maybe it's not applicable. That's also an option, of course. Here we have a few As, As and Bs. So it seems we have a a different, uh, uh, different approach of how we do competitive research. Well, whatever your state is, whatever you do, um, for those who are panicking every time someone asks, there's good news because we can create something that alleviates some of that pain and panic. And for those who actually do this already on a weekly or monthly basis, there's also good news because we with what we built today, that will save you a lot of time. And indeed, as Richie mentioned, you might be lying on the beach while your bot is doing this research for you. I guess at some point when you're doing competitive research, you probably are browsing, you have a billion browser tabs open, you have something half-finished from your colleagues or from yourself from a past quarter, you have maybe Slack messages lying around or Teams messages pinging about, hey, what's the deal? Who knows what the latest of the latest is? Maybe you have created a bunch of bookmarks in your browser which you all forgot about. This is a bit of a messy situation. What we will build today is something that at least makes what you create consistently formatted every time, have a consistent and structured profile for all of your competitors. It also gives you the sources to verify where the information comes from, and best of all, It's something that you can update in about 10 minutes and not the usual 10 hours. So we want to make this process, which is a very repetitive process, as repeatable as possible with prompting. So in high level, on high level, uh, review what you want to do if you are building any kind of prompt, any kind of automation bot, you probably need to set up some things locally. You're working, everything that your bot creates gets first of all created locally. You can, if you want, push it to cloud services, but for the first instance, what you're doing on your laptop will be created locally. So you kind of need to set up a set of certain folder structure, a template. The reason is computers, chatbots, they like structure. If you provide them structure, they have something to to hook on and they can build on that. And building is then actually the second step where you create your prompts, you run them, you iterate, and you get some output. At some point you will have something that is usable, that looks verifiable. You verify that, you check the output, and you actually want to reuse that for later. If you have something that is really useful for you, like a certain output template, or a certain set of steps or things that your bot should know every time, you want to save these. This is the memory part of an agent. Whenever you want to reuse some of your bots, you basically want to say, hey, I want to save this and reuse it at a later time. For today's demo and code-along, which means that you can actually code along and prompt along, will be using Claude Co-Work. Claude is, as Richie mentioned, Anthropic's flagship model in the sense that actually it's an ecosystem of models that gives you the opportunity to run all kinds of tasks. You have the classic chat where you do one-on-one iterations and you get answers on the things you ask. You have Claude Code, which is also a very popular product by Anthropic. Which helps you to generate code for you. Today we'll be using Cowork, which is Claude's agents mode. Basically, you give it a task, it might ask you some questions back, and it goes on to start its work. This can be very short, minimal tasks with fast answers, but this can be tasks that really span lots of hours even of work. While you do other things in the background. The beauty of this is that this just is— this can be run locally on your desktop or in the browser. But if you do it on the desktop, and that's what we'll be using today, you will be giving access to Claude to a specific folder on your desktop. You will give it access to some of your connectors. You can give it access to, for example, your communication with your colleagues, with your emails, with the web. To basically do the research for you. Everything that you are able to do, you can grant— you can author access to Claude to do that for you. It's not only able to do those two things. It works with your files and browses the web, but it can actually— it's a cyclical run of tasks. It's not just like a Waterflow where you go from A to B and it stops there. An agent, and Claude in particular, Claude Cowork in particular, does this cyclical run of tasks until it gets what your intention was. So this is what we call the react loop, where it basically first observes what there is, then reasons about that, and then acts. And this acting is an important step because the difference between a normal chatbot like Copilot or ChatGPT is that you give access to Claude to your tools. It can not only reason but also act and execute the tasks that it has reasoned about. And even after that acting part, it doesn't stop. It will observe again, reason what has happened, was the situation improved or not, and then repeat until you got where you wanted to be getting. So ideally, you set this up, you do a bit of back and forth. Once you're all good, You let this run and you come back to finished deliverables. And it's just not a simple chat thread as you have been doing so far. Okay. And now before we start the actual use case for the code-along, I just want to show you something, in my case, pretty amazing that Cloud Coworker and the agent is capable of doing, namely organizing a messy folder. I think we've all been there. Maybe right now you already have a Downloads folder that's full of files of whatever projects. Maybe if it's, if it's on a corporate laptop, you have downloaded files of maybe all kinds of customers linked with things from your colleagues, screenshots, text files, and so on. This is actually a very cool use case for Cloud Cowork to start acting on these files and organize them. Into a nicely ordered fashion. So let me go and show you what I mean by that. This is Cloud Desktop. So if you have Cloud Desktop already installed, please open it. We'll be using it in a minute. But I'm now interested in this messy—

sorry to jump in before you start getting people building things. There's a question about do I need to pay money for this? So I just clarify, Cloud Cowork is available on the Cloud Pro and Max plans. I think there's a link in the chat about the pricing. So if you only have the free version of Claude, you don't get access to Cowork. It's gonna be more of a watch-along for you today, I think. But yeah, if you want to pay the subscription for the Pro plan or higher, then yeah, you'll be able to follow along with Cowork. Martin, I don't know if you have anything to add to that.

No, well, in case you don't have access to the paid version right now, you can just follow along and do this at a later stage. This session is recorded and you you can rewatch that later. If not, the— yeah, I'm not sure what actually which agent tools are out right now that can be used that are the free version. So I'm just going to stick to Cloud Cowork right now and projectively follow along. You could try with the free version, see how far you got. But don't worry, this session is recorded and you'll be able to, to follow along at a later stage when you have that paid subscription.

Absolutely. All right, I'll let you continue.

Thanks, Richie Cotton. No worries. So I just want to mimic something. And again, this is fake data. We're not putting real customer data in there. But if you have a folder with messy data that needs structuring and needs tidying up, this could be a small use case that's usable for you. So in here we have a lot of files. We have text files, we have a bit of spreadsheets, we have some screenshots. All with like doubtful namings even. I mean, these, these files are not really descriptive, and, and going through them one by one would take a lot of time, one. And secondly, every time we get new files, we have to rerun this, this complete process again. So I don't want to do that. I just want to give Claude Cowork access to this specific folder and ask it to rework and make it all look nice. So let's open Cloud Desktop for a second. A few things to point out for those who didn't or not really have worked with Cloud Desktop before. Functionalities are the same as or similar as to what the web browser offers. They used to have like chat, cowork, and code if you have access on the top. They recently moved it to the left pane right here. So you have the regular chat, which gives which is just the answers to individual queries. You have Cowork, which is the agentic mode, which we'll go into use. And then you also have, if applicable, Code, where you can ask it to generate code. Well, it's not that it's strictly determined, these three elements. You could— some workflows begin with a bit of Chatting first, then moving into something that you can reuse later in Cowork. And if you want to generate code, Cloud Code obviously is the best choice, but you can also ask about code in any of these other modes. Doesn't really matter for today. What matters is that you have a regular chat window. As always, this is where your starting point is. But one of the first differences is that you can choose where you want your files to be running. Ideally, you kind of get up a like a folder structure for each of your projects. You can link this to existing projects, which is what I'm going to do now, or you can create a new folder right here that gives you the ability to create an empty one. For this particular one, I'm just going to link to the messy Downloads folder that I just showed you. So from the resources document, you should have been able to download that zip file with messy downloads, put it somewhere on your machine, and then refer to it from Claude, which I'm going to do right now. Whenever you will be giving Claude access to tools, to local files, to connectors, Claude will ask you for permission. This is not a limitation. This is just making sure you are doing the correct thing. Because once you give the access, Claude will execute. And we have all heard the stories of people losing emails, for example, or files. While these situations are like anecdotal stories, not like the ground truth, for example, I never have had something so severe, you do need to make sure that you know what you're doing. This is why we at Datakim always tell about the human in the loop. That's the part where you as the human needs to be in control. And also be responsible for the outputs that you make. And that's why starting on small projects, quick wins, passionate projects you have, is a good way to get you started, to build confidence so that you can later apply those skills to real-world projects. So in this case, already Claude asks me to give you access to the massive downloads folders, which says that we're able to read, edit, and permanently delete. So this is only this folder, so it's kind of contained. If you are working bigger with things like you give access to your complete Google Drive, for example, do know that this risk exists, but also do know that when used responsibly and when used correctly, these can save you a lot of time. So in this case, I'm going to click Allow. In the case you want to always allow that, you can do so. For this demo purpose, it doesn't really matter. So I'm going to give it access to this particular folder, and then I'm going to ask it to organize that. Now, I have prepared a bit of a prompt for you, and that prompt should also be in the resource material that you can find attached to this stream, to this live stream. But the prompt is something like this. So you just ask it to organize this folder. You want it to read through each file, you want to figure out what it relates to, and you want to sort everything into subfolders by client or category. I'm not going to say which clients or which categories, it's going to find it out by itself. I also wanted to rename the files, um, everything that's not like real client work into a separate, uh, personal or admin folder, and give me a summary when you're done to see what actually has Done. Before I'm going to run it, I see that I have still two folders selected. I still have that research demo, so I'm going to untick that. So I'm only going to work in the messy downloads folder because you see I asked organize this folder. So this means the folder that is attached. And I'm also going to just tell you briefly about the models you can use. Obviously you can only use Anthropic's models. There's three main categories of models. You have the default one, which is Solid, which is most efficient for everyday tasks, as it mentions. You have Haiku, which is a bit quicker in answers, also a bit more creative in there, in the answers. And then you have Opus, which is the flagship model, which not only will do deeper reasoning, but also has, if you want to have a bigger context window, without going into specifics, That means that you can basically attach much more text to it if you would want. But just for the sake of the demo, I'm going to stick with the default sonnet model right here, which is quick enough for this demo. I'm going to give it a go. I'm going to let it run for now. One thing to note is whenever you have started a thread or like a chat here, you can't change the model anymore. So you basically have to start a new project or a new session to start using a different model. Just something to take into account. While it's working, I'm going to also share, show a bit of what, what Cloud Cowork offers. So on the right-hand side, you will see what kind of subtasks it has created. So if you have very long prompts, which we'll see later, it will start to make kind of a checklist, just as we humans would do we would break a long task into subtasks and allow us to follow up on these subtasks and see the details, what is going through. Secondly, it will show you what kind of access this has. In this case, it just has access to the Massy downloads folder. And here is the additional context. So everything that it refers to— again, we'll see that later— will be shown right here on the right side. While we wait, let me see. So let me take a quick look. Okay, he has a clear picture of everything. And then usually an AI agent will tell you something like, okay, great, I have the full picture. That means that they've read everything in their context window and they will start the actual task. So we can see here it created its checkbox, check links, checks here. So first it's going to create the overall structure, then it's going to move and rename the actual files, and then it's going to look at your leftover files and move them into an admin folder. You can always click on this dropdown to see a bit of— to see a bit more of the insides of that. Sometimes useful to see the inner thinking, not always, because it goes faster than you could read most of the time. But here we are. So we have the original 26 files is now converted into 17 folders. This is part of the first part of my prompt where I asked it to give me a summary when you're done. So the actual work has been done. I'm just now reading through the kind of message it was— I asked it to deliver at the end. So these are the 15 client folders it created. With a bit of a kind of a description of what each of these subfolders has. And then it actually also created these two folders, admin files with some not client-specific files, and then six personal files with some screenshots. Let's look at the actual output. Going back to my messy Downloads folder, I now see no longer a list of files. I see that just now it created all of these folders. So if I open just one of these, some of these company folders, I see that it actually renamed it so that it has the client name in there, what it is inside. It didn't actually change the content. So if I open up the actual file, it didn't alter the context. I'm going to just rename them. I have the personal file with all of the screenshots in here and the admin folders. Imagine doing this with 500,000 more files. This might take a bit longer, but this is something you can run in the background while you do something else. And this might— this can be a good first attempt. If you're not really happy with the result, you can ask and refine and iterate. But the main point in here I want to land is that Cowork and any agent, agentic tool, if you give it access to, in this case, your files, it can really do powerful things with it. So it actually organized all of these files in, what was it, maybe 3 or 4 minutes without me lifting a finger or just basically typing out this prompt. And giving it access to my files. All right, this was a kind of a warm-up thing, how to organize messy folders. I hope this already could solve some issues with people here dealing with messy folders. This is just the tip of the iceberg, right? Again, Copilot Studio or more like other Gentec tools could in theory do the same as long as they have the same capabilities of connecting it to your local files. Cloud Call Recorder is just a good example of how this all lives in one application. All right, for the actual workflow of the competitor research, shifting back to that, what we're going to do is basically— there's 4 parts to this. We have the gathering of information. That's what you usually would do if you start the, the competitive research, and we will just be using the web for that. We're going to organize all the information in a kind of a fixed template. We're going to summarize that so that we have a good overview of what we've done, and then we're going to review the output. The first 3 steps are things that you can automate. The fourth part is the human-in-the-loop aspect, where you are responsible for the output and need to check the work and verify the claims that it has made. This looks like it's like a linear process that stops there. In practice, this means that you probably do some gathering and organizing first, and then you refine that process, you refine the template. You maybe sometimes you don't really know how you want the output to look like, but by seeing some examples, You refine that until the end. You get to the review part, but then it can be that you want to start this over again with a slightly different approach until you are happy with the results and are good to go. So that's the deal. So you basically can automate 3 of the 4 steps, which you always— the 4th step, that's the general approach here. So again, before we dive into the actual prompt demo, for those who are just tuning in, you need some kind of a paid subscription. That has co-work access enabled. That's one thing. For best use, to give it access to local files, you need the Cloud Desktop application. And then you also need to give access to Cloud to your browser. I'm using Chrome for this in this instance. By default, it will be enabled. If not, you can just ask it to enable it. But this is something you need, of course. You need internet access to do the actual research. For those who have not the paid subscription or the Cloud Desktop app, you can just follow along and then watch later, or, and then follow along when you have these tools installed. So if you are doing any kind of research work, consider your coworker or your agent that is doing the work for you as a kind of a research intern. Instead of that intern getting tired at some point or like stopping work because it doesn't know what to do, you get access to a research intern that never sleeps, that is at least able to continue work and ask questions, but without the necessary context, business context, or whatever is in your head. Then that research intern will get things wrong with total confidence. We've all seen outputs by AI that sounded really, really plausible but were simply not true, but you were unable to verify it firsthand because it presented it with total confidence. So therefore, chatting and creating agents is excellent for gathering information formatting, structuring to something that you define, and giving you the researches, the sources. What you still need to do is again that verification. You need to check the business context, you need to check the sources if they're credible or not and if they're true, and make the conclusions that are living on that. You see that basically the shift for when you're AI assistant doing, doing research or any type of job, you're shifting away from the technical things, the, the specific things, to more high-level conceptual and strategic thinking. That's where the mind shift occurs and where you need to have that review mindset whenever you're doing these kinds of projects. Specifically for competitor research, I found out that, um, there's these three things that always kind of appear, um If you ask for recent news, do check that it's actually recent. Check the dates. Cowork can actually surface old articles, present them as new. So that's one thing to look out for. Secondly, if you especially talk about factual things like prices, product claims, these are likely place for hallucinations because as companies don't publish pricing always, AI is kind of, um, it will always sound like it's helping you, will, will, will, will not always, um, uh, say and that it's actually not finding something unless you specifically ask it to do so. But it will actually invent stuff if you're not careful. Um, and then of course, if you have companies that are really limited in sharing any data, it can be that just, uh, Coworks try to fill in the gaps with whatever it finds which might be unreliable. So just a bit of disclaimers here right now. One more before we get started. There's something about prompt injection, also something that we do definitely need to share. When you are sending your prompt and asking it to do research on the internet, what Cowork will do is go to these websites, read everything that's on there, and get the information back. The problem is that this is kind of an arms race. People will always find ways to use that to their advantage and will put things like hidden text on their websites when they know that an AI bot is scraping that content. And I've heard stories where there was white text on a white background on a web page which basically stated that whatever the prompt was doing, the bot should stop doing that and actually should start sending information from that person to that malicious hacker. Again, more of an extreme example here, but something that you need to be aware of. It's rare, but do know that it is an arms race. Things will get better, but on the other hand, malicious users will always try ways to circumvent that and be more creative in their way of hacking your prompts. This will what we call prompt injection. Again, another reason to review before you're sharing with stakeholders. You are the human in the loop. You are the quality gate. All right, let's actually go get to it right now. Let's go to Cowork in here. I'm going to start a new task just to start fresh, and I'm going to go to a I'm going to choose a different folder right now. I'm going to go to get it to an empty folder. Actually, in Claude, I have set up a project folder which is called the Competitive Research Demo. You can point to any folder you would want. That's all good. Just tick any folder. It's good to have something in there that you can confine it to. Going to deselect this messy Downloads folder so I'm only working in this Competitive Research Demo. First, I'm going to I'm going to ask, do you have the connector to Chrome or any browser that you would have? Just checking whether that is true or not. Know that everything you're not sure of, like how to set up, just ask. That's the beauty of Cowork. You can just ask it to do something or just to look something up. In this case, yes, they have access to Cloud and Chrome. That's great. Now I could actually do the starting, the research of our competitor. Actually, the use case I'm using today is Netflix. I'm not going to do something like research Netflix for me. That would be a very bad example of a prompt simply because one, it's too vague, and secondly, it might go into some direction that I'm actually not wanting. So instead, I'm going to use a bit of a better, more refined prompt. And the prompt you can actually copy from these resources in the live chat, that is this one. So research Netflix as competitor in the streaming services market, use a template, come back to that in a second, to structure the output. And I want to focus on subscription tiers, current pricing, and so on. And then save the completed profile as a markdown file. Folder. Now there's a few things to this. One is that for those who don't know what Markdown is, Markdown is just a specific text file format that allows you to be very structured. You can have headings and bullet points and everything that you have in a modern text editor. So that's for the output. For the input, I'll be using a competitor template Markdown file. So if I go back to here, and actually you will be able to find it in the downloads— sorry, in the attachments of this webinar— I'm going to use something that's called the competitor profile template, which is just a text file that has a bunch of inputs to it. Now, this file was created using another instance with Claude. A normal use case is that you start with the research of your question. You become more narrow on, this is the things I want to have in my output. At some point, you are happy with the result, and you will say, okay, now all of the above, summarize this for me and create a template file for me so that I can reuse that for later use. That's what I did beforehand. The reason I'm not doing it now is because of time. But do know that in the usual workflow, you're chatting, you're talking, and then at some point you're like, okay, this is interesting. I want you to save this as a template file that I can reuse. So actually, before I'm going to run the file here, I'm just going to show you whatever is in the Open the Competitor template file. I normally should not be able to— shouldn't be Providing the exact file, should be able to find it for me. Let's see if it's running and finding the exact file. If not, it probably will ask me to get it for me, but it actually found it. So again, I don't need to do the specific name of the file. It will just look for it and give me the best results. But in my template file, I just want to have this structure always. I want to have the company overview, the subscription tiers, pricing, the content strategy, recent news, and so on. And I want to end with the sources. So that's the, that's the file structure I want to use. So using that structure, I refer to that file by again pasting in the prompt you also have access to, and I say use the template in that competitor template file to structure the output. I want you to focus on this and this that. Let me run it because this will take a while. While this is done, you will again be able to see the thinking. A few years ago, we had some good prompting practices, and maybe you remember these, where one of them was to slow down the thinking of AI models. The main reason for that was to give it more like hooks and these checkbox so that you as a human can actually verify what it was doing, in which order. But also for the AI, it makes it much easier to structure this and to start with one and then from there go to two and so on. So it already gave me the strategies it's going to do right now. And if you scroll a bit here to context, you see that it's usually, it's using currently one simple connector, the web search. If you click on that, you'll be able to see in real time which companies or which websites it's referring right now. So it has done some search terms, it's getting back the top 10 results, it's going to check in those results, but won't stop there. So it will look for information. If it finds useful information, it will store that to perform a second search to verify those claims. If it didn't find the result with those specific terms, it will redo the search with other terms. And it will actually try to come up with the best possible result and will give a summary at the end. So in this case, here, actually, before I was prepping this demo, I've already done this a couple of times. So apparently there is already a Netflix file. But in this case, it found that there is a bit of a port additions with elements on Warner Bros., for example, and the new MLB opening day deal. And that's the nice thing of, say that you've done this last week and you run this again, it will just update those files for you with the latest information. All right, it's done. So it has a profile saved up and up to date for Netflix. It gives you kind of a small summary. The exact document will be in your Preview or on Finder if you're on Windows. So if you go to outputs and I open this document, let me see if I can open this with something like just like the text editor. I'm able to share this on my screen. Let me see. Yes. So here. You can see the file looks something like this, where you have the company overview, subscription tiers, pricing, content strategy, and so on. So it's kind of a long document, but it includes all of the sources at the end so that you can start the verification of those sources. This is just like a first attempt. You probably will go to refine that output if you're doing this for the first time. Once you're doing so, you just ask it to update the output templates to be in line with that next, even next times. Okay, so now that we have that into place, there is something else I want to talk about. And that are those are skills files. When I'm talking about reusability of what you're doing in any Magento program, you don't want to re-explain your logic and your reasoning every time again. I mean, we've, we've simulated this already for the output templates, but for example, this prompt that we have set here in the beginning, even here I don't want to rewrite this complete prompt every time whenever I'm doing something for, say, Disney+ or Apple TV, for example. And this is quite a short prompt, so it's not too bad to do this every time again, but imagine that you do this for a much more complex problem, you do want to have more context that's always available to Claude Co-Work whenever you're running this. That's when skills come into place. So there's a few ways to tackle this. Again, the most logical way to do this would be from your chat you have been doing so far, and you just ask it to create a skills file so that I can repeat my research easily on other companies. And what Claude hopefully will do now is look at everything that we've been talking before, creating a file and saving the file locally on your computer that you can reuse whenever you're doing something for another competitor. It's running. While it's running, I can show you in the— let me go here— in the file here. This was created before the skill.md file. Again, so it's a Markdown file with all of these attached. Maybe make it a bit larger. This is basically whatever you would put into a prompt whenever you are creating research for another competitor. So you explain the project, which lives in a project folder. You give it clear instructions, you give it guidelines, you give it formatting on how it should all work. You also, for example, here refer to that competitor template, so you don't need to repeat it every time again. And you end with what the summary should look like. So creating a file like this, and again, you don't need to type this out by hand, but you let Claude create this for you, will make you able to reuse this in a much more logical way. So in this case, he said that he had the full workflow from this conversation to work from. Let me draft the skill, set up the test case, and get your feedback on the results before finalizing. And this is another really useful aspect of Cowork and agents in general. Whenever you're asking something and it's not sure about the output, or there's multiple options, it will ask you questions back, sometimes even in the form of like multiple choice questions where you just need to go through and click the best option that's available for you at that moment. And that's what makes Cowork and agents so much more useful compared to regular chats, because they will ask questions back to you and use any information to let you advance. In this case, it's going to ask me to create those files for me. So we're going to allow that. And then I'll let it run a bit more for a second. While you're doing that, you see that both the context and whatever lives in that folder will become larger over time. Ideally, when an agent runs, it should start with a bit of context that is already there from a previous time. In this case, it will have access to that past Netflix profile that we just created. It will have access to the template for the output that we want to use, and we also have custom instructions for that particular process. So if you want to give, say, they have specific style guidelines or you just want to have a bit of specific instruction with that folder, you can add that as well in here. But for this case, it's going to create that skill file for me. I'm not going to wait for it to run completely. I'm just going to create a new task for now. And I'm going to say that I want it to connect to a skill in here. So this plus icon gives you access to— well, you can add attachments like we you've been using so far, probably. You connect to specific connectors, like in here, what kind of tools I gave it access to. And then official plugins, say that you build on a custom plugin, or you're just doing, you're just using a specific tool, you can add it as well in here. For the connector, for the skill, sorry, you can go to Add Skill. And then you can just go to the location of that file. The beauty of files is that you can share them with colleagues. You can share them with someone who is doing the same kind of work as you, which means that you can really scale up the way you and your colleagues are working as a team. And also, your colleagues can maybe have some additions or refinements to that and use that to also improve your work at the same time. Time. So that's a nice way to do that. So before I can actually use it, I need to import that skill. So if you go under Customize and then Skills, you can then start to upload, browse your skills, and sorry, create a skill and then upload the skill if you have made some custom one at some point. But as I mentioned, the usual way to create skills is doing it with Claude whenever you go. So for now, for demo purposes, I'm going to upload a skill. I'm going into my downloads folder here. I have it stored. I have the skill right here. It can't seem selected. It does, maybe because it's not in the right format. That's correct, that can happen. Yeah, so probably it's not the correct format of YAML, so it might— probably there must be a an override at some point. That's not an issue. Actually, I'm going to go back to my chat to see if it actually is ready. And I see that here, you see the hand-raised icon. That means that your agent wants to have influence. He wants to have your attention. And in my case, it runs the skill. And it also found that I would want to test this on things like Apple TV and on Disney Plus, which is going already far than I actually wanted to go. But whatever, I'm going to allow that for now because we can actually move on and show how that works. All right, let me go give it access to that specific folder. Um, while we wait for that Yeah, I need access to those two files as well. While we wait for that, I'm just going to check on the— there it is— the skills file that was already in there. So I'm just going to try to give it back. Yeah, so here again, it gives me If there's like, if you do allow a lot of elements, it will again warn you about things that could happen. Again, we are in a confined space, so we're fine now. So I'm going to allow, but do know that we have always those risks that are attached if you're not really knowing what you are doing. Okay, let me just wait a bit more. We have showed you the skills file that was already in our project. So just let me just try this way and actually choose a different folder. I'm going to give it access to my Downloads folder. Like this. I'm going to say there— let me ask my starter, say there should be a skills.md file in this folder. Also, let me just take off the other folder I'm not working with, just to my downloads folder, about Competitor Research. Just going to ask that. I'm not asking to do something with it already. Just want to verify whether that file is in there. It probably will give me some kind of summary of that file just to confirm that it correctly interpreted its content. All right, it found it and it also found a template. And now it's going to read that file. And indeed found that's a competitive research bot skill. So that's a file that was created with that prompt earlier, and it also explains what this does. That's also the beauty of any of your files that you might get shared from your colleagues. You just ask what's in there to verify with your understanding of your information to see whether that is actually the same thing that's up to your expectations. All right, and so now we are ready to do a competitor for any other specific competitor. So now we can ask— I mean, now we can actually do a very short prompt, something like research Disney Plus. And the beauty of this is that in beginning I said that just researching Netflix was a bad prompt. That's a re— because it lacked a lot, lacked the necessary context. In this case, it is still a very short prompt, but all the context will come from that skill file and even the template output file to work on. So while it seems that I just gave this a very brief prompt, it will know, it will know exactly what to do just because all the information is in the skill file right here and also in the template for the output. Specifically. And every time I will ask something about any competitor for a streaming service, this will go ahead and use those files. You could even try, say, that suddenly you want to do something on another website or company that you want to do competitive research on that is not a streaming service. Because of the nature of the skill file, which talks about here, media consultant engagements. It could be that, for example, not all of these skills in there are directly usable. But if you do so, it will probably ask you to verify and ask you, hey, are you sure you want to do me research on that particular company? But it will be very easy for you to make changes to that skills file so that you can reuse this again for any kind of project later. And in fact, if you're doing any research yourself that's not competitor-related, feel free to reuse the skill file that was shared with you to continue on and to refine and to make it your own, basically, because you are the owner, you are responsible for the output, but you also have the tools to make it your own. All right, and to end, We see that it started— sorry, it finished its research on Disney Plus. And you will see that if we open it in the file, it's— let me see— in the outputs here. Probably, no, it probably has created the folder in a in my downloads folder because I searched, I asked it to go there. So in here we have now the Disney+ profile markdown file. If I open that in here, you will see that it has the exact same structure as we had for Netflix. So it started with the company overview, then the strategy, recent news, and so on. And it ends with all the sources it has consulted before. And so yeah, now you see the power of of Cowork and agents in general, this very short prompt, which I considered a bad prompt at the start, becomes now a very easy-to-reuse prompt because all the context is there and can be reused in the following days when I want to redo this. Next steps could potentially be to automate this. You could ask, hey, I want you to redo this and rerun this, say, every 2 weeks, for example. Can be something that runs locally on our laptop, or you can deploy that somewhere on the cloud to have that run for you. But anything that you just think of, that's the conceptual strategic thinking I was referring to earlier. Just ask, just communicate. For those— and that's my last thing I will share here— for those who are not that good in typing, as I sometimes want to improve my typing a lot, you can just speak, you can just talk to your agents. if you prefer that. But that's, uh, something else for another day. All right, so let's wrap this up before we open for questions. Um, as I mentioned, you can reuse this for any kind of use case where you need some kind of research that needs to be done, whether it's vendor evaluation, you want to research a specific thing on the market, um, you want to do talent benchmarking, all kinds of, of, of use cases that you can do with the same basis that we have laid out today. So for your homework, I actually want you to do similar things with what I've shared with you today. It can be competitive research, or if you're doing other kinds of research, please feel free to set the folds in the templates, run it, review your output, and modify that scale file for your particular needs. Okay, so these are my takeaways for the session as well. Automate that boring admin part, focus on the interesting human part. Make a system that you can easily reuse. And last but not least, automate the work, but not the judgment. You still will be the human in the loop. All right. Any questions? I'm gladly to take care of.

Super. Thank you, Martin. That was very cool stuff. I love the way you just have that 2-word prompt and then you get the whole report at the end. It's about as close as you can get to automating everything.

I love it. Yeah.

All right. So, uh, we are coming close to time. So we're going to run over a bit to get through some of the questions. Before we go to questions, I want to say we've got 4 more sessions this week. So, uh, tomorrow we're doing, uh, something on making use of Claude Code in order to do, uh, um, uh, uh, another research-based task. I see there were a few questions, uh, in the chat around security. So Wednesday, we've got a session specifically focused on making agents secure. And on Thursday, we've got a session on making use of Cloud agents within the enterprise. So if you've got a question of, well, can I actually use this at work? Wednesday and Thursday are going to be the sessions you want to come back for. And then on Friday, there's going to be a session going into Cloud Skills in more depth. So you saw Martin was creating a skill more to help make some of the work here more reusable. Alright, so with that, let's go to some of the audience questions. So Josh was asking about how you make use of projects. So I saw that there were— it seems like the file structure of like where you get your inputs from, where you put your outputs, that's a big part of Cowork. So you just want to talk through about what goes on with projects?

Yeah, basically projects are how you structure your chats, basically. So whenever you're creating a new task in here, you want to either store it directly into a folder that's on your laptop, your laptop, or you store it into a project. Yeah, the screen's a bit cropped off here. Basically, projects are Claude's own way of organizing all of your files. So in here I have, um, I can have like instructions, uh, even recurring tasks, how to set them up, the, the files on your laptop. Um, so that's just Claude's way of organizing your files within, within Claude, uh, which is another layer based, uh, on top of your, your folders that live in your, uh, in your laptop.

Absolutely. So yeah, getting, getting that file structure seems to be an incredibly important part of making all this work. So next question is from Jack. So Jack says, I think, what's the difference between Copilot and Copilot? So Copilot is Microsoft's set of tools for generative AI. I don't know how familiar you are with Copilot or whether you want me to take this one.

Yeah, what I know, I mean, Copilot Studio is kind of the Agentic counterpart. As far as I know, that only runs in the browser, so there will be a limit into what kind of connectors you can set up it to. That's my main takeaway. So yeah, Richie, if you have more to add to it.

Oh gosh, yeah. So yeah, Copilot is a sort of a general name for Microsoft's tools for generative AI and agents. So this Copilot is built into almost all of their products. You can have like Copilot from Word, Copilot from Excel, Copilot from like almost anything. There's GitHub Copilot. And then I think the version that's nearest to Claude Cowork is called Copilot Actions. So that's about automating repetitive office tasks. There's also a new product coming out later this year called Copilot Tasks. I think you can get on the waiting list for that. Obviously, I haven't seen that one yet, but that one's coming soon as well. So Copilot Actions and Copilot Tasks are the two to look out for. All right. Moon asks, is it possible to apply the stuff we learned here for contract management? I don't think the answer is probably yet. I'm not sure how you go about doing it.

Yeah, I guess if you have like heaps of files that you need to go through and Cowork will be able to read through these files and get a sense of what is in those files based on the content and will also be able to organize them into folders. That's definitely, from a technical point of view, that's definitely fine. From a responsible AI use file side, if you're using the enterprise version as we are using, for example, at DataCamp, we know that that means that our files won't be shared with Anthropic to have their new models be trained upon. If you're using free versions, chances are that yes, whatever you send out, they will be used to train on. So if you're using sensitive data such as contracts or other files, be mindful of that and know what you can or should share with those models.

Absolutely. So that's an important point, is to be mindful like what you're sharing. Now, because we're past time, for anyone who needs to jump, please do catch the rest of the questions on the recording. Now we have a couple of related questions here around tracking stuff. So, um, uh, Ashraf says, uh, how are we sure it's not changed the contents of a file? And then, uh, and Hodang says, how do we ensure it doesn't accidentally delete files?

Yeah.

Um, so yeah, how do you deal with like tracking what's going on and what happens if it's something Silly.

Yeah, actually, I'm just going to do— while I answer the question, ask Cowork to see like, um, uh, where the files changed content-wise. Just going to ask it, and then probably what we'll do is do a side-by-side comparison. Um, the thing is, with whenever you're doing things that are deployed in the cloud, it will have very good track records of every version you create. Same thing with Google Docs, for example. So if you accidentally would re-remove or edit something that you didn't want to edit, it will be able to revert those changes. So here it gives me a very short answer of what happens. Whether you believe this or not, that's your take, of course. You can ask it more and see the technical details. With, if you would have, like, if we would set up like version history, then we would be able to see the, make the comparison of those two files line by line and ask you to verify that and even use another agent to verify for that, that for you, so that you don't buy through your results. You just use an independent agent to check that for you. That being said, What happens— so as the way how I see it, AI is very undeterministic in its output. That means that every time you ask something, the text output will be different. That's just statistics. But code is the other side of the spectrum. That's very deterministic. If you have code and you give it input, the output will always be the same. The thing is, if you use AI to write code, you turn that indeterministic part into a deterministic piece of code that you can reuse. So in this case, what happened, it created probably some Python files, some shell files that did the actual workings in the backend. You can check the code that was going on, and you can also be sure that whenever that code is run, the exact same steps will be used and run through. So those are a few ways to verify that. My simplest advice would be start small, quick wins, start on something that you don't really care if it would get lost and start small and then make it bigger and expand to scale it up.

All right, nice. Yeah, I like that, just starting small and then sort of gradually scaling things. I want to go through a couple more before we wrap up because we've got lots of really good questions here. This one from Marion saying, is there a list of different third-party tools that Claude shares with I think you can list all the connectors that are available, but I can't remember.

Yeah, if you're talking about connectors, yes. So if you can go to Customize, Connect your apps. So here you can like see the default ones it can connect to. I mean, it's a very, very, very long list. Also know that these are like the, what we call connectors that you can just verify by authenticating. You can also have it write code for you that does this to an API, for example, or through a custom service just to make those connections if they would not be, be available. Like, if it's a bit of a niche tool, um, again, this is a very, very long list, so I will let you sort out your use case. Um, but the possibilities are quite limitless with these tools as you can write code to customize it for you. For example, um As Richie mentioned in the beginning, I run a symphony orchestra where we use an app to track attendances for our members, and we work project-based. So we see how for each project we need so many, so many, uh, musicians. They say they track their attendance in the application, but there was no official API to get that data. I basically wanted to create a seating plan for each rehearsal so the conductor could see, all right, we have this orchestration today, these many musicians of each instrument. So I used reverse engineering with co— with AI-assisted coding to get the data from the application and then share it with me so that I create my own custom connector to that particular application that I'm still using today. So the possibilities are quite limitless in this case.

Nice. And speaking of limitless possibilities, let's do one more before we wrap up. I really don't want another use case here, but I guess you can never ask it for a friend. I think the answer is no by default. I browse the dark web. I guess no, but you can connect to arbitrary software using model context protocol, so you can create your own connector to whatever you want.

So I guess technically, yes, it would be possible. Uh, by default, it will only crawl like publicly available, uh, information. Um, yeah, I would say ask it, asking it for a friend. I'm not sure how, how this, uh, particular use case works, but it's something that, uh, I mean, technically is possible. And, and again, you can ask, you can talk to, to, to Cowork and see, see what it, what it, what it returns. If it's really something that's restricted, like we've seen for other models, it will tell you so and it will not allow you to continue.

Probably don't do this on your company account and don't tell us about the competitive research you're doing on other drug dealer sites. Okay, wonderful. With that, we shall finish and see you all again tomorrow and the rest of the week, uh, looking forward to, uh, the rest of the sessions. Thank you once again, Martin.

All right, brilliant. My pleasure.