We Tested Claude Cowork for a Week. Here Are the Results...

So if you weren't living under a rock over the past week, you probably heard of Claude Cowork. In summary, it's ChatGPT with the ability to remote control your computer and your browser. And when you give it a task, instead of just doing it, it creates an elaborate plan and then follows it step by step. We spent a week looking at different things you can do with this powerful tool. And in this video, we bring them all together. I show you what works, I show you what doesn't, and I give you some tips and tricks along the way. If you want to learn more about the details of Cowork and how it relates to other products in the AI space, I created a separate video on that on release. I'll link that below. But this video is going to be all about different ways you can actually put this to work. Let's begin with the first one. Organizing folders on your machine. This is one of the most basic things you'll see people do with it, but it's also so powerful and it just works. So I already showed this off in my initial video, but we kept testing it and it just keeps performing. Even if you keep it super vague, like organize my downloads folder, then all you need to do is actually select your downloads folder. In this case, mine is already super well organized, so you can see the footage from us testing it. I've been doing this on all my machines and it just works so well. This is a thing that a lot of us have been waiting for with our messy desktops and download folders, because it doesn't just organize it. It starts out by creating a categorization system and asking you some questions about how you want it done. If it's by project or by file type, things like that. And then it does a stellar job of doing it. In my case, on this machine, I had to follow up prompts twice because it created really big archive folders in some cases. And I wanted those archive folders sorted into the other categories too. And it did so. This is an absolutely incredible use case. You can use it to organize everything, hard drives, archives. But also this is the one that most of you have probably seen if you followed this story, because people have been showing this all over the place. And my hope is that this video goes a little deeper. So let's look at the next use case that you can make happen with Cloud Cowork. And this one links to the first task, but many people don't think of this. It's renaming images inside of those folders. So let me just do that. And Cloud Cowork, which by the way now is available to all Pro subscribers too, that's the $20 plan. A week ago when I created the other video, it was only the more expensive ones. And also if you haven't seen that other video, this thing is only available on the Mac app for desktop now. Windows app coming soon. It always just releases on Mac first. Let's do this. Now. So here in the Cowork tab on my Mac app, I'll just say new tab. And I say name all of the screenshots in my downloads based on what they are. And I just hit go. And that's literally it. It's going to go through, it's going to look at the images, it's going to use image recognition and rename each one of those accordingly. If you're an avid follower of the channel or generative AI in general, you might know that this is a thing that ChatGPT could already do, right? You can just upload images and let them name it. But The difference here is that it's just so darn easy. I just say what I want and it figures out a way to do it. In ChatGPT, this takes so long with moving files around that it's not worth it. And if you're more advanced and using something like Claude Code, which is a version of this for coders, if you're not aware, well, it could do these things. But even I myself, as somebody who uses that tool regularly, sometimes found that the friction of using it made me not wanna use it. And that's the bottom line of this Cowork product. That's what all of these use cases will usually point towards. As you'll see, it's just stuff that has been possible before in some way, but now it's become so easy that it's actually worth doing. Feels fluid and it feels like I have to worry less about what it does because you can see it working here right now. There's progress here in the top right where it shows you the different steps. So it analyzes screenshots. And then if we check back in a few minutes, that entire folder, which is a mess right now, I mean, what is all this stuff, is going to be handled. And all of that doable on a $20 plan is a pretty great deal. Okay, let's fast forward to the result. A few moments later. Okay, so full transparency on what happened here. It got stuck. It didn't work. I had to follow up with continue and worked for another 3 minutes. Something got stuck again. I said continue renaming all the screenshots. And then after another 5 minutes, it actually worked. So yeah, to be real, I wish I didn't have to say continue multiple times, but look at that. AI generated news anchors collage. That's a perfect name for this. AI Advantage YouTube channel, 300K subs. "Screenshot 2024-09-29 unknown." Yeah, fair enough. That's an appropriate name. I'm equally as confused by what this is. So that's totally fine. And yeah, turns out this really worked, although I had to follow up from twice. Okay, so next up, I'll show you something that I wanted to work, but that didn't actually work. So Cloud Cowork actually has some connectors built in. So if I check my settings, I'll see that this can search the web, and it can use the Chrome extension, which remote controls my browser, can also use my local files. That's another one. And then there's also some external apps, but let's stick to these for now. So as it organized my downloads folder and it renamed a lot of the screenshots, I figured that it might be a good idea to actually upload that to my Google Drive so it's backed up. Now thinking back, that should have been my initial step before I gave this AI agent access to my local drive, right? In case it deletes something, I still have it in my drive. Anyway, I tried it at this point in time. And this is the thing that did not work. Because what I prompted for is to upload the files to my Google Drive. But as it engaged the browser extension, the Claude for Chrome extension, it couldn't successfully navigate its way through the interface of Google Drive that is quite simple, but there were just some browser restrictions where it didn't let it press certain buttons to upload files. And this is just one example, but it is representative of a whole host of problems because often when you use external apps, you will just hit these walls where it just can't progress. And that's just the reality of an internet that is currently still built for humans. The Google Drive website is not built for an AI to be remote controlled by, right? That will change in the future. And we already see integrations like these connectors up here that are a bit flaky. And actually saying they're a bit flaky is an understatement. They're very flaky. And I'm not a big fan of them because they just don't work reliably when I need something to get done. And that's why I don't use them extensively in this use case video. We're focusing on things that are achievable with these built-in connectors like the web search or the remote control of the web. I suppose we tried. Again, it's not perfect. So this is a thing that did not work. Don't worry, I have more use cases that did work, but I did want to point out its limitations, which, yeah, remote controlling your browser is a big one. Okay, so for the next use case, I want to show you something that people have been doing since the dawn of ChatGPT, and that's creating reports. Now, it might be obvious to you that AI can create reports. But this application, Plot Cowork, creating reports takes it to a different level. Because all of a sudden you can really turn it into a multi-step process. And usually the best reports aren't just, hey, here's information A and let me format it into a report. Usually it's like, okay, here's kind of something I want to express with the report. And then there's going to be multiple steps along the way that I need to outline. Maybe there's multiple sources and I want to compile all of that into a report. So the person reading the report doesn't have to go through all of the steps, right? That's the point of a report. So what we figured in our testing here is that what if it created a report on what just went wrong with uploading files to Google Drive? And the result was actually stunning. It took the entire chat history, took all the problems, all the error logs that it encountered along the way to failing at the past task, and it compiled a report in Google Docs. And this is a big thing that I really want to highlight and have you take away from this video. Its ability to take multi-step processes and digesting them into Google Doc or a Google Sheet is immense. And you might notice that this report shows all strengths of the newest Claude model that is actually really good at creating these documents. That's an update that came around about a month or two ago. But yeah, and actually going out there, creating this file that is ready to share with anybody from a multi-step process is impressive. Let's see what's next. Okay, so at this point I started feeling confident. It did an amazing report. It reorganized all the chaos that was in my folders, even renamed the screenshots. Great. I wanted to push it. I wanted to break it a little bit. So I thought to myself, what if we take all the research that we have on this video and we turn that into a slide deck and we do it dynamically? So I have all the research for this video, everything that we prepared here in one Notion page. What if I try using the Notion connector, even though I mentioned that these connectors are flaky sometimes? But what if I give this a shot? What if I let it retrieve all that information, process it however it wants, and then turn that into a presentation, into a slide deck. Well, I did exactly that. And let me tell you, it just worked. The connector worked. Retrieving the information from Notion worked. I just had to give it the name of the page and it hit it perfectly. And then it took about 10 minutes to create the slides, but I don't care. I did other things meanwhile. You can even run multiple tasks in parallel inside of Cowork. And it came up with this. So here I'm viewing the entire presentation in PowerPoint. And look at that. It's pretty good. Matter of fact, this would be a great way for me to walk through the video. I might even start using this on every video. If I have video research or an outline, heck, why not just turn it into a slide deck, which then I can keep on my screen or a second screen while I'm presenting. And again, it's not to say that AI could not create slide decks before. It's just the ability for it to go into my Notion where it sits by itself and to work for 10 minutes independently and create something of this quality level. That's the difference. It's the ease of use of bringing these tools together and it just works. So I thought about this. Why did the connector here work whereas in other cases it really didn't? Well, I noticed that a lot of my connector uses, I had large expectations of them, right? So when I connected Notion, when I connected Google Drive, I wanted it to find certain files. I wanted it to analyze maybe a big chunk of what's there. And there's a lot there. In this case, what I did is I gave a specific page name. And it did that well. So that would be an evolution of my thoughts on all these connectors. If you use it on something specific, it has a chance to help you. If you expect it to know your entire Drive, it will not help you. It will fail. In this case, specific Notion page to slide deck, I give this a solid 8 out of 10. And I gave it no instructions on how I want the slides, right? So if I told it to be a bit more concise or maybe to be speaker slides to support me presenting, it would have came up with something different. This is so good that it makes me want to use it regularly. And I hope you at least try this too, because this really works well. Okay, so for the next use case, it's a combination of using internet search and its ability to create dashboards, websites. Both of those fall in this category of what Anthropic calls artifacts. And Claude could already do that really well. But here you can link it really well to some other steps to find information. In my case, what I did is I wanted a subscriber tracking dashboard on my route to a million subscribers. That's a big goal, but it's something that I set for myself recently. And I wanted to see if Claude Cowork could come up with a dashboard to track this. So I told it, create a YouTube progress dashboard towards 1 million subscribers for the AI Advantage YouTube channel. And then I gave it a channel URL to make it really clear that this is the channel that I'm referring to. And I actually found that the web browser encountered some issues along the way to find the right type of data. What it ended up doing is it couldn't look at the YouTube page directly. And it referred to a bunch of other sources and it took data from January 2026, which is fair. But this monthly growth is completely off. The community here is growing by 4,000 to 10,000 subscribers a month right now, not 990. And without further ado, here's the dashboard it created. As per usual with Claude, it's an amazing dashboard. This is great. The milestone progress is beautiful. This is the one graph that it didn't populate. But other than that, This is fantastic. It even gave me info on average video length. And I think just like in Claude, this is a powerful thing that I want you to be aware of. It's creating dashboards, it's creating websites, it's creating mini custom applications. It's just something it can do really well. And now you can combine it with the web browsing tools in here. And just because I list this as a separate use case doesn't mean you have to use this separately. You could be doing another task and just at the end you say, okay, now turn all of this into a dashboard. And it will create a little plan on what that should look like and execute on it with all the tools it has for you, resulting in a dashboard like this at the end of some other interaction. So as for all of these use cases, they could be linked together or used in sequence. I'm just showing them off one by one. And hey, if you're enjoying this video, make sure to subscribe. We make practical AI tutorials for all levels. So make sure to subscribe to get us closer to that million. And if you're a regular follower of the channel, you might notice the new set. This is my new living room. Some videos might happen from here. And let's see, maybe we'll even turn on that fireplace at at some point. All right, let's continue. So this task is something that we tested and we found that it works because also people online are reporting about it, but I wanted to do a live test to show you if this actually works. As I told you, I clean up this downloads folder and there's thousands of files in there, organized now, but it's still thousands of files. And I have one little video clip from a few weeks ago here in Lisbon. It's just in this beautiful park called Jardim da Estrela with yellow leaves falling down. And I just wanted to find that clip. I want to see if it can retrieve a specific video clip that is somewhere on my drive in my downloads folder. So I'll just ask it, find a video of a tree with yellow leaves. And then I'm just gonna set the downloads folder as the source so it doesn't have to look for everything on my drive. Let's see if it gets this. Okay, why is Claude going to youtube.com tree with yellow leaves? Oh, well, I need to take responsibility for doing this because I told it to find a video of a tree with yellow leaves. I didn't specify that it's my downloads folder, even though I linked it. Let's try again. And now it has it. So that was really quick. It said it found it. That is not the video. Okay, let's give it a second chance. That's not it. Try again. It's a phone video. Oh, it got it. And to be fair, this was a bit of a curved ball because the video was already renamed. It was Lisbon in January.mov. And yeah, sure enough, that is the video I was referring to. Okay, so that worked-ish. Now let's continue with this video file that we found. Here's another use case. You can use its ability to run code and do all sorts of operations like turning this into a GIF. So let's just continue our conversation here by doing that. Now, here's a pro tip. This package that it's using here, FFmpeg, is something that can do various operations with videos. It can trim them, it can cut them up, it can do very basic video editing for you. So you could essentially prompt for very basic edits. You could make it black and white, you could trim the video, you could turn it into GIF like here. And if you want to learn more, you know, you just ask, what could FFmpeg do for me with video files, and it'll help you. But for now, let's just have a look at this GIF. Yep, that's a GIF of the tree. This is, yeah, perfect. Need a little workflow and just another thing you can do here really easily. Okay, so if you're at this point in the video, you're like, okay, I get it, but none of this is like really impressive or useful or new. Then oh boy, do I have something for you. Let's go deeper and let me show you something a bit more complex here. Because What this can do is take an entire folder of files, right? I showed you the downloads folder with 1,000 files in there where it worked well. And then I also showed you the ability for it to create spreadsheets. Now, what if we combine those two, but instead of the downloads folder, we give it a folder of a bunch of information? In this case, the information is going to be benchmarks of different AI models. So what we did here is we took a bunch of these screenshots from various blog posts, and we just threw them into a folder. Then we gave Claude Cowork access to that folder, and then followed up with a prompt that gives context on the screenshots and what's inside of them. And then the goal of having a spreadsheet that compares all of these fairly. And let me tell you, the result here is really impressive. It made sense of all the screenshots. It created a new spreadsheet detailing all of it with multiple tabs. Now, to be fair, we did include something in the prompt that says push yourself to impress me. So it really went a little overboard as Claude usually does so well. But look at this. Finally, we have the Anthropic models compared with the OpenAI and the Google models all in one place. And then the thing that really excites me about this is whenever a new model comes out, I'll just put that screenshot into this chat and I'll let Claude Cowork massage this into this existing sheet. Heck, it even created our own tab on strengths and weaknesses. Based on the data that it saw with the benchmarks. I thought that was really impressive. And it really goes to show that you can throw a lot of variety at this and it will just figure out what to do because it has this ability to do multi-step plans before it actually acts. So it's not just more powerful on the acting side with all the tools to, you know, remote control your browser or your local file system, but also the ability to plan and reason before before it actually goes into doing is incredible in a case like this. So this one was one that I didn't expect to work as well, but it really did. Just throw a variety of information at it, and it'll compile all of it into any one of the formats that I showed you in this video. And as we're talking about transforming one thing into another, for the next use case, here's another format you can consider. So we started with a PDF file, concretely the Gemini 3 Pro model card. This is their PDF that they release along with the models that contains a lot of scientific information about the model. And then we asked it to turn that PDF into a video outline. And what can I say except that this just worked. Created a beautiful video outline. And when I followed up and asked it to actually turn that into PDF for me, it did that too. So yeah, its ability to work with files and turn one format into another is really amazing. And even an upgrade to everything you can do in competing products because it has the this planning capability and this kind of, if it faces a problem, it comes up with a way to solve the problem system built into it, which something like ChatGPT doesn't have as of now. Okay, so there's one more thing that I want to share with you in the same line of, you know, working with multiple files and then producing impressive results. Concretely, what we do on this channel is create videos to teach you and to show you how these things work, to open your eyes and maybe even inspire you to what's possible. And in that sense, we wanted to kind of have Claude Cowork do the same thing for us. So what I did here is pull the past 5 transcripts of this YouTube channel. I put them into a folder, gave Claude Cowork access to that folder as per usual, and then gave it this prompt. I told it about the folder and I asked it to analyze them. What advice would, in this case, MrBeast have for me? And it looked into all the files. It went through a multi-step process of actually identifying common themes, making recommendations, and it ended in a document full of recommendations, sort of a report again. Now, I personally found this report to be really useful. But what I want you to take away from this is that its ability to work with folders of files, and then doing stuff with that is incredible, even if the folder is not well organized. But ultimately, this last use case, even though useful to me, shows its limitations. The fact that I had to manually pull these transcripts, put them into a folder, give it access are still manual steps. This clearly shows that direction that this tool will grow into, which is really automating these processes end to end. But that doesn't take away from the fact that a lot of opportunities are opening up right here in front of us. A year ago, you still needed to go into ChatGPT and upload it file by file and prompt it, reprompt it. You had to have awareness of the different tools that it had and you had to ask for them. Now you still might have to compile the folder of the files, but the rest happens sort of automatically. So if you have awareness of these different use cases that this video gave you now, of the different results that you can achieve, then the question really remains is what is the sort of data or what are the screenshots, what are the reports, what are the documents that you could give it so it can think over them and turn them into something new, something better, something more concise, something more visually appealing that will help you achieve your goals. My hope is that all these examples made that a little bit easier for you. And that's pretty much everything I have for today. And with that being said, my name is Igor Pogany. And I hope you have a wonderful day.