Limitless Podcast - Claude Cowork Was Built By Anthropic's AI in 10 Days. Now It Can Do Your Job

Episode Date: January 16, 2026

Claude Cowork is Anthropic's revolutionary AI tool, simplifying task automation for non-coders and quickly became a billion-dollar platform used by major companies. We discuss its impact on the future of work, questioning whether it will replace jobs or foster collaboration. Join us for insights on productivity transformation and a sneak peek into our upcoming tech advancements roundup!------🌌 LIMITLESS HQ: LISTEN & FOLLOW HERE ⬇️https://limitless.bankless.com/https://x.com/LimitlessFT------TIMESTAMPS0:11 The Birth of Claude Cowork2:00 The Automation Revolution3:15 Autonomous Development Breakthrough3:59 Exploring Claude Cowork Features5:15 Demo and Tutorial10:17 Personal Use Case Insights12:14 Unique Problem Solving with AI14:35 Automating Business Processes17:28 Future Implications of AI Tools21:28 The Path to AI Adoption------RESOURCESJosh: https://x.com/JoshKaleEjaaz: https://x.com/cryptopunk7213------Not financial or tax advice. See our investment disclosures here:https://www.bankless.com/disclosures⁠

This is going to break your brain. An AI model just built another AI model, and it's probably the most important software launch in 2026 worth billions of dollars. I'm talking about Anthropic's latest tool called Claude Co-Work. It's basically vibe coding for everyone who isn't a coder. And when you realize what this means for knowledge workers, you'll start to realize why Anthropic is quietly the most important company in the world right now. The way it works is you just connect it to your desktop or your browser, and you just explain in plain simple English what you want it to do.

And the examples are pretty crazy. People have already used it to turn on their oven. People are using it to monitor their plant growing by their windowsill, as well as tidying up their desktop and even replacing their nine to five job. It's doing the entire thing for them and we'll get to that later on. But it brings into question whether this is the end of jobs as we know it today or whether this marks the start of a new era where AI and humans basically do everything together. Yeah, it's hard to understate how profound the shift is. And I related to something that's very near and dear to me, which is photography. And you could kind of relate this to how in early days the smartphone, it kind of democratized

photography where suddenly everyone was a photographer, the bar to become one was low, the cost of entry was very low, and code work is doing the same for automation. Suddenly, everyone goes from a novice at code to a full-blown system to engineer. just based on a single human-made prompt. So previous AI assistants, they were like having a really smart person on the phone. They could give you advice, but you still had to do the work.

Cloud code is kind of like having that person physically in your office with access to your files, your browser, and your tools. And then as I was going down this automation rabbit hole, and to your point about job automation, it led me to a study that shared that the average knowledge worker, they spend 28% of their work week,

managing emails, and 20% searching for information. So that's nearly half of their team work now that Cloud Code can handle. Okay, so before we get into how this tool works, there's a really important story to tell about an unexpected surprise Anthropic Discovered, which led them to building this product. So Cloud Code, it launched originally in May 2025 as a terminal coding agent, very intimidating, very difficult, very confusing for non-technical users. Less than six months later, that product was at a billion dollars of ARR. It's 10 times the usage.

It grew 10x in three months, and it went completely viral. It was the fastest startup in the world to go from zero to a billion ARR. Companies like Netflix, Salesforce, KPMG, everyone was using it, and it was the biggest thing in the world. And software developers were starting to use the cloud code for other things. They got really comfortable, so they started using it for things like canceling subscriptions, tidying up their emails, vacation research, monitoring the plant growth, and turning on remote things in their house like ovens or lights.

And it became a full stack automation software. The problem was, is that cloud code wasn't really built for this. But when users bend your product to do something, you didn't originally do. design it to, that's normally a pretty high signal indicator of a problem that you need to solve for. So that's exactly what Anthropic did. Or rather, that's what Claude Code did because it built Claude Co-work to do vibe coding for everything that isn't coding in 10 days with the team of four people written 100% autonomously. So it was this pretty amazing unlock that happened over the course the last 10 days. And here we're seeing on screen the post from Boris, the founder himself, and how he

confirmed that it actually was 100% built autonomously. And I think maybe we should double tap on that and see how exactly they were able to accomplish this. Yeah, I mean, this is the craziest part of the entire tool, by the way. It's the fact that it was built by another AI completely autonomously, or rather with the aid of four humans, but they didn't really do anything. They just kind of acted as called code supervisor and just made sure it was on the right track. And it only took a week and a half. Typically, if you wanted to build a product like this, Josh, I'm guessing it would have take like months using traditional software engineering techniques. So it is one of the coolest things that kind of came out of this.

And it's a sign that in the future, some of the hit products are going to be built purely by AI. But I want to get into Claude Co-Work itself now, Josh, and kind of give everyone an explanation as to what it is, the rundown of the key features, and then I really want to get into some exciting demos. The examples are going to be really fun. We have some good examples, say,

and how this actually works in production. Exactly, exactly. But before we do that, let me talk to you about, Claude Co-Work itself. The way it works is you can give Claude Co-Work access to your desktop, a folder on your desktop, or a web browser. And you basically just tell it what you want it to do in plain English. So write out your prompt, hit Enter.

Claude Co-Work comes up with a plan. It shows you the plan. It says, hey, this is what I'm going to do. Are you cool with this? And you just click, yes, no, or can you just change this bit? And then you're all good. Off you go. And then that's it.

Go for a walk, go make yourself a coffee. Hey, go watch a Netflix episode. It's all done. And you come back literally 10 minutes, an hour later. And suddenly all those tasks that you've been putting off for the last year, it's me, I'm talking about me, is completely done. And the way that all of this is enabled is pretty cool. They've got some nifty features here.

It's built on the same software stack that Claude Code is. So it works very similarly to Claude Code, but it's for non-technical people. You don't need to download a command-line interface. you don't need to know how to code, you just type in plain English, and it's done. The second thing is, it runs on a sandboxed virtual environment. What that means is when it operates your desktop,

when it operates your web browser, it's in its own environment in order to do that. So it's siloed to the particular folder that you give it access to. It can't just go rampant and start accessing anything and everything. You specify what you want it to completely. And then the final thing is,

you can pair it with your browser, which means that you can do a bunch of web-based tasks, which Claude plugin already allows you to do. And so on its face value, this sounds interesting, right? But I want to know what it can actually do in practice. And some of the examples that I've seen so far and that I'm about to show you are pretty fun to do. Yes, and I want to start with these examples,

particularly the first one in this post, which we're going to do live, which is a cleaning up of your desktop. So, Ejazz, you are notorious for having a slightly less than clean desktop. And it is normally full of contents that most people wouldn't want cluttering up their desktop. So what we have prepared is an actual prompt with cloud code that we will use to organize this desktop as a live demo and just kind of show you how co-work works in production. Yeah, so here's a live view for the unfortunate viewers who are watching on Spotify or YouTube of what this desktop looks like. And I just shared with you a prompt that I've actually worked with Claude to create. So this is the way I would advise everyone use it is instead of using your own prompts, prompt the L.

to create a full prompt that captures the entire scope of what you're looking for. What this does is it will organize your desktop into a series of different folders based on productivity, inbox, projects, finance, and it will run checks to make sure that it's doing the right thing, make sure that it doesn't write the, delete the wrong files, and it will include organized file names as a way to index everything alphabetically. So it'll keep things very organized. And he created this, I mean, it's a fairly detailed prompt. There's a lot going on here.

So if you paste this into Cloud Co-Work, you should see it start to get to work here. start actually organizing the files on your desktop without you touching anything. All right, Josh, listen, I won't lie to you. I'm a little nervous about what this is going to do to my desktop. What happens if your computer breaks? If the recording ends right here, how do I know, like, Claude's not trying to nefariously get involved kind of maliciously here? On our previous episode, we proved that, we suggested that it might be more aware than we

think. Anyway, we were looking at the scheming index in our last episode. So what you guys are looking at now is the main, interface for Claude Co-Work. If you have a Claude Max Plan, you have this neat little tab now on your desktop app, which allows you to access Claude Co-Work. And it starts off with a pretty neat prompt, which is, let's knock something off your list, and it gives you a bunch of ideas. What I'm going to do is I'm going to paste that prompt which Josh created using Claude into here to tidy up my desktop. And you'll notice this little folder section, this drop-down menu here,

where it says, work in a folder. I'm going to specify that it works in my desktop. And I'm going to say, allow for this particular case because they don't quite trust it just yet. And let's go. Okay, so now, how does this think so we can kind of walk through how the best ways to interface with this are, which is plan, execute, and update. So the first stage that we did was the planning stage where I kind of prompted Claude to create a larger prompt. That was planning that outlined the scope of things. And then it's going to execute, which is the phase that we're seeing now. And then you can actually decide to update. And you could push the update live to your machine. and we might want to take a peek at your desktop

because it might be doing it right now. And that's the loop that happens with this. That's the recursive loop. It's the plan, execute, and update, which mirrors how a good human assistant works. They don't just do things. They show their work. They check in and they course correct as they go

to make sure that the model doesn't drift as it goes on these long complex chain of thought processes. And what I'd like about this step in particular is it doesn't just plan and then execute immediately. It gets the humans not of approval first, which is very important towards like AI working with humans in the future. A few things I want to point out. It understood the prompt.

It created a plan and it did this in its command line interface and it's coded. Identified what it needs to do. And it gives you me a little summary here. It says you have 27 screenshots, many older than 90 days, several PDFs, 15 folders, images and media files and an installer. Claude, would you like me to proceed with the dry run as Josh.eats prompt described? So it even knows that you send this to me. That's frigging hilarious. So I'm just going to be like, yes.

Run it. We'll see what happens. And it works pretty quickly. I think we're going to see this happen in a second. Really quickly. And while it does this, maybe I could describe a different example that I used personally recently with Claude Co-work. And it's, I use a tool called Obsidian, which is a markdown note-taking application. So the app is basically a harness for markdown files that organizes them in a really unique way that connects them via backlinks. And it's kind of messy. There's just a lot of notes kind of spread out everywhere. But because Claude can ingest markdown files, it was able to ingest all of the notes that I've made over the last three or four years and then draw unique insights from them and organize them in a unique way. And it's a really nice file management

system in a way that I would never have tried to tackle myself. So that's been fun. If we're looking at the process here, what is going on? It's doing a lot of stuff. It's, and it's doing it, if you noticed, in rapid speed. So basically, it's identified all the files and folders that are on my desktop and it's created a bunch of categories that were prompted by the process. You might wind up with a lot for this. I think you're just going to make my life better, but also worse. And so we have a bunch of file moves that it's already actively doing active

projects. You can see here what it's actually doing. And what I like about it is it is just obliterating this. It's given me a few that it needs to review. So this is a cool point to make, which is typically when people are uncertain about using these AI products, it's because they don't trust it and that they think

that the AI model is going to do something that they don't want it to. Claude Co-work has context about what those things might be. And instead of doing it, prompts you to say, hey, you need to give me an idea of what you want to do with this, right? There's a movie folder that you have a bunch of movies that you watch on the plane with. Is there something you want me to do with this? Right. So type of proof to proceed with these changes. You may not want to approve this because this is going to create 16 folders. So perhaps you want to prompt it to compress that into four or

five of the most common categories. That way, it will actually organize it into a way that doesn't add more folders, it actually compresses things. So maybe, yeah, choose the four or five most common themes among the folders. And it will apply some more smart thinking to that and generate you something a little more neat than 16 separate folders. Actually, I've decided to disagree with you. Approve. Let's see what happened. Let's get this bad boy going.

Come on, I want to see this done live. So while we're waiting, what other fun use cases have you seen? Okay, so another fun use case that I've seen, and I wish I had the example for me on here, is someone, to his wife's disdain, lost their wedding photos. And then he found it, but it was in a file that was corrupt and unreadable. So he decided... This story was amazing. It was an amazing story.

So what he did was he went to the original photography company and they said, hey, this is the software that we officially use, download it, and just run it. to get your wedding photos out. One problem. It was a Windows-only app, and he has a MacBook, so it never worked in the first place. So he was, like, distraught,

and then he realized, oh, I could use Claude Co-Work to do this. So he asked Claude Co-Work to identify the file in his browser and then code up, sorry, not in his browser, in his desktop, and code up a custom-made app

just so he could access the photographs. And Josh, I kid you not. Two minutes later, he had it running slicker and better than the actual official software that the wedding provider gave it to him himself. It's pretty insane. Okay, just talk organization complete. I'll be ready for the grand reveal. Yeah, swipe for the grand reveal. Let's take a peek. Drum roll, please. Nice. Wait, that, dude, what the hell? That is actually insane. And what's

the best part about that is all of you just got to saw that live. Pop into something. Let's see what, let's see how it did. Let's see. We got to check. What are we feeling? It's wrong. Why is my inbox here? Okay, let's go into inbox. Okay, it's got some random screenshots in there. Don't know why that's done that. Projects. Okay, yes.

I love this because, okay, this is pretty cool because it's identified separate hobby and projects that I actually work on in my spare time. And what's interesting about this is I have other folders that were on my desktop that weren't projects, but it was able to somehow identify that these were personal hobbies

and do this for myself. That's freaking awesome. We're one for one right now. We're one for two. Pick one more, Josh. You see anyone that's picked? Let's go with that second one. Media?

Yes, let's go with the media tab. What's in there? Oh, Josh. Oh, look how adorable those guys are. Look at us. I mean, come on. So it worked. That's a mission success. That is a really next demo.

And you could kind of think about how this scales out to larger corporations, right? There's this one crazy stat where the average employee spends two hours per month on expense reports, which is kind of an outrageous thing. For a company of 1,000 employees, that's 2,000 hours of productivity lost monthly. And something like Claude handles this with just OCR to read receipts, categorize expense based on vendors and the amounts. And there's so many ways in which you could automate a business. And there's this interesting example that I wanted to highlight, which was the spreadsheet and accountants in general, where if you remember not too long ago, before computers and calculators became a big thing, there were buildings full of people whose job it was to actually do math, to do long-form math to calculate the, accounting of a business. And that entire building of people got replaced by a single spreadsheet.

And what's funny about these companies, and you think about microhard with XAI, their entire goal is to have photons in and then bits out in the terms of like building native and unique software. So if you think about a building full of humans in a spreadsheet today, the spreadsheet is far more efficient. But not only that, if you compare a spreadsheet that is 100% automated versus a spreadsheet that has a single cell that a human has to monitor, well, the single cell that a human has to monitor is going to compute much slower, it's going to be a lot less accurate. It is going to just give you problems that it will not.

And I think it's kind of a testament to how future companies will be structured. Because if you choose to do something like, let's say, expense reporting, and you don't use 100% AI for this, you don't use Claude Code or Cloud Co-Work to do something like this, well, anyone who does automate that part of the business will immediately unlock two, additional productive hours. And this scales across the scope of pretty much any business that exists today, where so long as there is a verifiable outcome that you're looking towards, you can automate that in the world of bits at least. And this kind of plays out,

you could see how this plays out over a long period of time. We're starting with bits where anything that you have to do on a computer, any bit of software can be automated so long as it has a verifiable outcome. Then the world of atoms comes next, where, well, if you have to move physical objects. The cost of that comes down with human-ed robots as we scale. And you start to see this job replacement type argument become a little more real. The other side of this, the flip side, is that accountants since computers have gone up, I think, 50 times. Like the amount of accountants that exist post-computer versus pre-computer has multiplied so much because of the Jevin's paradox. I mean, now that there is an abundance of accounting capability, a lot more companies want this

for their books. A lot more people want to, a lot more companies one exist, a lot more companies two are able to actually keep an accurate record of account. And the net net is a positive. So while this probably reshuffles things around, I don't think the net is going to be as negative as a lot of people perceive. Well, with that prompt, I kind of want to get on to the final part of this episode, which is kind of like the bigger picture. Where does this tool lead us? And I want to run through a few examples with you now. So number one, I think this tool is going to be useful for millions of people. And I really mean that versus a fad. We've been sold a few lies from the likes of OpenAI and Google's agents, which sound good in theory, but haven't really been productive. This is something

that my mom or sister can access today, and it will add net new value to their lives. And some of the use cases that we've discussed already on the show is just the tip of the iceberg. We haven't even fully explored what this tool can actually do at great lengths. And if we were to hypothesize here, and if it went the way that Claude Coe did, and I'm showing a tweet from Boris Cheney where he described the creator of Claude Code, where he describes the journey that Claude Code took, it started off slowly and then super, super quickly. My hypothesis is going to be an AI literally just wrote 100% of a tool that is going to be responsible for GDP's worth of economic value that is built on top of that. And I don't think people have fully grasped how important that's going to be

for humanity in general. The second thing, I need to throw a little bit of shade before we wind this episode up, Josh. This tool was meant to be what Microsoft co-pilot was going to be. We've got to talk about Microsoft. But instead, they spent one and a half years shipping absolute BS. Open AI dropped the ball completely when they should have been keeping tabs on Microsoft. And Atthropic ate their lunch, or rather an A&A. model ate their lunch in 10 days a week and a half from scratch with four humans. That is like a tiny fraction of a percent of Microsoft's entire workforce. Just insane. Yeah. So there's two points on the fad and the, um, is this a fad or not? You can kind of put it through the utility test,

which is, I mean, the difference between fads and these real transformative tools over time is how you're able to use this over over the course of your day, over the daily work test. And these aren't really novel tasks. They're tasks that people actually need done repeatedly. So is it a fat or not? No. I mean, this is what people need every day. And as this iterates, it's going to get better. On the Microsoft point, Microsoft has 220,000 employees. And they had a three-year head start. And what Anthropic did in 10 days with four people fully assisted by AI is a testament to just how powerful it is and how unexpected, I think it was when you get a killer use case like this. If you remember, chat, GPT was kind of an accident. It wasn't really supposed to be the entire

business of Open AI. A similar thing is happening with Anthropics. and with, I guess, companion agentic protocols, because they certainly didn't expect this to be as big as it was because they didn't have a large team on it. In fact, after this went viral, they posted a bunch of job listings to join this new kind of experimental team where they test these new things because you never really know what that killer use case is going to look like.

And in the case of Anthropic, they've hit a home run. And it makes you wonder what Microsoft is doing. I mean, they have such a tremendous workforce. They have more money than they know what to do with. and all they're doing is building data centers, but have nothing to really show for it. And finally, the last two points I want to make is, okay, on the point of work, you kind of mentioned it, Josh, but I want to kind of dig into like whether this is going to replace people's jobs entirely. The answer is kind of like yes and no. It's going to replace a bunch of knowledge work that people do. I'm showing this on the tweet that we have up here right now where he describes he interacted with Claude Co-work and two hours later it had done his entire week's worth of work.

so he didn't know what to do with himself when he logged on every day. He didn't know what to tell his manager because all the work was done and he was just kind of like shipping it to him, which was crazy. But it's reminiscent of a trend that I think we're going to continually see in 2026, which is something I'm calling recursive models or recursive language models. Basically, AI models are smart enough now to understand what humans actually want and to go and build it themselves. So previously, they had the tools to do it, but they were dumb.

They kind of couldn't put context together. Now they have it, and we're seeing things like Cloud Code build tools like this, which I think are going to be worth billions of dollars, Josh. If it was its own startup, I think we're going to see one of the quickest ascents to a billion dollars in revenue. It might even meet Cloud Code. I very strongly agree. Okay, so how do you try this? For Cloud Code, you need to get the Pro Plan, which is a $20 month membership for Anthropic and Cloud. For Cloud Co-Work, which is what we demo today, that is their max plan.

That is $200 a month, which is a little more pricey. It's worth experimenting at least with CloudCode. I think as a person who is kind of working, if you are in the workforce today, it helps to flex the muscle to reach for an agent versus reaching for the things that you used to, like Google or a manual work process. And one of the things that I'm personally working on

is figuring out how to retrain that muscle to know when to reach for the AI, to know when to reach for the agent. And having Cloud Code on my desktop, on my phone, it's been very useful in kind of massaging that muscle to, oh, instead of going to Google or watching a tutorial, I can actually just ask my agent to do it. And it does it really well.

So that's a fun practice. I think maybe for the prompts today, what cool stuff if you have used these programs have you done? What have you automated that we should try? Because one of the most difficult things is figuring out where to apply this. What part of your life can you actually improve? Extra points for creativity. For anyone, like let us know what you've done if you have access to this plan

or to this product in the comments below. And please also like, like, subscribe and follow us because like we're putting videos like this more and more. Josh and I are inspired by Claude Coax. We want to stop building apps, demos for you guys to show on the show. If that's something you're interested in seeing, please let us know because your feedback is very important for us. And Josh, I don't know what the time is, but we are trying to keep this tight under 25 minutes because it's on the resolution. We're close. We might have done it. So if there's nothing else to add from you, I think that's a that's a wrap. Thank you so much for watching. For the people who are regular listeners,

the roundup is coming early next week. There's just so much going on this week that we wanted to cover. We're going to move the roundup to early next week. There's so much much to be included in that, along with things like Tesla economy, there's humans coming down from Earth. There's a lot of stuff happening. So stay tuned for that. We will see you guys in the next one.