WEBVTT

00:00:00.000 --> 00:00:44.950
So Anthropic just released computer use, which is basically where you give Claude access to your browser and your keyboard. And a lot of people out there are using this for silly toy demos, like how to organize your file workspace or whatever the heck. In this video, I wanna show you eight ways you can actually apply computer usage to real economically valuable tasks. Like for instance, reaching out to tens of thousands of people automatically in a highly personalized way using my browser. So I'm gonna start with a brief demo, and then I'm gonna show you guys how you guys can set this up on your end. We're also gonna be bypassing Anthropix built in, uh, read block, which basically is where they don't allow you to use your browser for incoming rights. Uh, I'll show you guys how to do that, and then I'll also show you guys how to set it up in case you guys wanna know. Okay. First and foremost, what's going on here? Uh, I'm on LinkedIn, and what I'm doing is I'm sending connection requests, which are the way that you sort of do outreach on LinkedIn.

00:00:45.485 --> 00:01:00.685
And one thing to note is that LinkedIn inherently blocks the vast majority of browser automation because it's obviously in their interest to maintain a highly human sort of environment. Right? Most social media platforms will do this. Instagram does this. Uh, you know, x is gonna do this.

00:01:01.360 --> 00:01:17.065
Facebook and and so on and so forth. And so any sort of, like, automation of this task is pretty difficult to do in practice, and you'll receive some sort of either major platform block or, you know, you're just gonna run into some form of issue. Computer use basically negates that for you entirely.

00:01:17.305 --> 00:01:25.065
If you feed in the right prompt and then use the right browser analog, you can have whatever automation you want done basically constantly for you in the background.

00:01:25.305 --> 00:01:46.175
And I foresee a future where we all have a 100 Mac minis or whatever the heck is the the trending item. Pretty soon where stuff like this is occurring on autopilot all the time. How do you actually do this? Well, I'm just gonna open up my thread here and you'll see that what I've done is I've said, I want you to use computer use to control my computer and send connection request to people on this page with the min browser,

00:01:46.175 --> 00:01:47.695
which is full read write access.

00:01:47.935 --> 00:02:03.730
That's important because Anthropic automatically blocks the usage of Chrome, Safari, Firefox, and most other major browser tools for this. Then I say customize each connection request so it doesn't suck. Use a connect request like, hey, first name, saw casual version of company name, and I think we'd vibe quite a bit given our interests.

00:02:03.890 --> 00:02:09.490
So when I have computer use on, as you guys could see, the browser looks a little bit different. Right? We have sort of that, like, yellow outline.

00:02:09.945 --> 00:02:46.975
And it's actually, like, controlling my mouse. What it's doing is a variant of what it's already capable of doing with, like, Chrome DevTools, MCP, or whatever, where it takes a screenshot and then, you know, it, like, kinda controls it. But you can see that at any point in time, I can just jump in and make minor adjustments and so on and so forth. And, you know, as long as I'm not doing that while it's doing something super valuable or important, it sort of figures it out on its own. So that's use case number one. And I just showed you guys it on LinkedIn, but I wanna show you guys just how many other platforms you can do this on as well. Um, obviously, a big one here is Instagram. I know sending Instagram DMs automatically is obviously very high in demand and, you know, a lot of people wanna wanna do that. But the same thing applies to any other platforms. You could do this on x.com.

00:02:47.215 --> 00:03:07.960
Um, despite the fact that I don't think Facebook is a great place for this stuff, you can also automate Facebook entirely. Then obviously, you could do this on other platforms like, I don't know, TikTok and in a short form and and so on and so forth. The way that you do this to make a long story short is one, you install a browser that not a lot of people have heard about. A browser that is not on the block list. Anthropics current block list, I think, is

00:03:09.035 --> 00:03:11.115
Chrome is Chrome.

00:03:11.195 --> 00:03:16.155
I think it's also Safari, and then they also do, obviously, Edge and Firefox.

00:03:16.315 --> 00:03:45.125
Min is a simple, free, and straightforward browser that I'm not affiliated with at all. It was just the first one that I tried that wasn't part of this block list. And what I'm doing is I have Minh set up here. I then logged into my LinkedIn account, and then I just give it a list of all of the different, you know, leads I wanna actually, uh, you know, send requests to. And in my case, I did this pretty straightforwardly. I just pumped in the term AI automation into my LinkedIn sorted by people, and then I just have it going. Now I should note that, you know, most of these platforms have outreach limits, so you can actually send, like, five quadrillion pieces of outreach automatically.

00:03:45.205 --> 00:04:05.430
You'd need either multiple accounts to do this or, I don't know, some other sort of, like, black hat method. And I don't recommend doing that. You know? I don't recommend overly gaming these social media platforms because there's obviously downsides and, you know, a fair amount of risks. Um, but you can apply the exact same approach to whatever the heck you want. After that, you just feed the list. And then perhaps most importantly, you have some form of what we in the cold outreach business call an icebreaker,

00:04:05.430 --> 00:04:11.455
which is some sort of, like, kind of quasi customized template that allows AI the ability to fill in variables dynamically.

00:04:11.615 --> 00:05:24.570
And so in my case, what I did, if I go back to co work, is I said, use a connect request like, hey, first name, saw casual version of company name, and I think we vibe quite a bit given our interests. You can insert or replace whatever variables you want here. It'll work pretty reasonably so long as Claude obviously has access to information about that. My next major use case is going to be scraping social media for posts and then saving that to a local file, which I can then feed into, I don't know, Claude code or some other features thing in order to, like, generate content on my end. Now the issue with doing a lot of this stuff organically is obviously, one, it takes a fair amount of time. But then two, there's no, like, API based way to plug this into a workflow. You either have to have, a virtual assistant or somebody do it, or you have to do some sort of hacky workaround. What I'm doing with cloud computer use is I'm giving it a really simple and straightforward query, just saying scrape my LinkedIn feed for 10 posts about AI, and then save the trending ones to a file for me with computer use. And it's just going through right now, basically taking screenshots and then also saving the text to a file. After I'm done with that, I could use the file for whatever I want. So some really cool use cases for this. Obviously, there's, like, trending news scraper style systems, which is personally what I'm using this for. Uh, but there are variety of other ones as well. You could try, a parasite based system, which goes through somebody's social media platforms. Again, you know, on platforms that do not explicitly allow or explicitly prevent any sort of browser automation.

00:05:24.810 --> 00:05:28.570
And then get the text. You could feed that text into AI to have it like rewritten.

00:05:28.650 --> 00:06:45.880
You could, um, save photos and stuff like that by right clicking the pictures saying, you know, save as and so on and so forth. And then eventually just have like a big file setup that you can, you know, then plug in your own content calendar. Once you're done, you then get a list of trending AI posts, and then you can do whatever the heck you want with these posts. In my case, maybe I'm gonna use these trending posts to create some parasite content or whatever and use that to rank on LinkedIn. My next use case is going to be giving it a list of forms. So in my case, I'm just gonna use dentist, but you guys could do whatever the heck you want. Contact forms for agencies, uh, I don't know, big people to get in touch with, whatever. And then just have it actually reach out to them by filling out those forms. And this is valuable because a big chunk of the Internet right now is, like, gate walled by contact forms and so on and so forth. You can't find people's emails, but you can get a contact form. And so you can use or have computer use, um, actually control your browser, click buttons, and so on and so forth to one, mitigate the appearance of captchas and other things, while also two, um, filling it out in like a human natural way with, you know, your browser and stuff. So in this case, I said with computer use, fill these forms out. Loose info, gave it my name and so on and so forth. You know, initially, it tried using Chrome. The reason why is because it always tries just using the specific, um, thing that it's connected to. What I'm gonna do now is I'm actually just gonna go into the min browser again. It's going to open up some new tabs, I imagine, using the URLs that I gave. And again, it's literally using my my keyboard, which is why it's so cool.

00:06:46.280 --> 00:07:08.885
And then it's gonna try and, uh, you know, fill out their forms and stuff. So here's a simple walk in dental clinic. You know, it's like a new patient form, and it's just gonna go top to bottom and click on everything. Just started by closing the chat widget, and now it's gonna start with the what looks like first name, last name, and then work through each field. I should note that, you know, a couple of these input fields are pretty dynamic. Right? One of these is like a birth year or birth date picker. It's not very straightforward.

00:07:09.440 --> 00:07:31.715
What it's done is it's just like clicked on four things in like two seconds for me, and it's just working its way top to bottom. And when all is said and done, it then fills out the form and then submits it. The next one's gonna be managing ad platforms without API access. So think of things like Google Ads or Meta Ads or TikTok Ads. Like, these platforms all lock down their ads stuff really, really strongly, and then they also just make it a pain in the ass if you do have any sort of API access.

00:07:32.355 --> 00:07:36.835
Um, the reason why is obviously they just don't want, like, a fleet of agents that control everything completely automatically.

00:07:37.320 --> 00:08:02.255
But what's wild is with computer use, you can do that. And you can make a fair amount of money in the interim by automating all of that stuff before, you know, they catch up. So this is one of the companies that I run, and we do ads for a fair number of dentists across the Canada area. We typically generate ROASs anywhere between eight to 10 x, so very strong stuff. And let's just say I wanna automate the process of, I don't know, like, turning off ads that aren't very good. So in my case, maybe I want to click through all new patient video ads in the current ads manager page,

00:08:02.255 --> 00:08:05.695
and then find the cost per lead. I want the inside view, so I actually want it, like, inside,

00:08:06.200 --> 00:08:49.110
not just on this little dashboard. And then I just wanted to turn off the lowest performers. Obviously, I could standardize this in, like, a bajillion different ways. Hopefully, that's clear. I don't just have to do it, like, super loosely, like, find the best ads. I could have strict SOPs where it's, find the top three performing ads, find the lowest five performing ads, and so on and so forth. But just like I use my own computer and my mouse to do this, I could do this with computer use controlling everything. This reminds me of like those Tesla self driving ads, but like my hands up here. So I don't know. It's finding the new leads ad, you know, it's seeing that there's no budget, and then maybe it's turning it on or it's turning it off. Hopefully, you guys get the point. You can automate the large majority of more or less any valuable knowledge work like this. I think, honestly, the thing that I'm coming to realize is the more important thing than actually just, like, having a list of things to automate.

00:08:50.115 --> 00:08:53.555
It's just knowing what to automate and then having a preexisting,

00:08:53.555 --> 00:11:26.180
like, SOP for that. Um, so believe it or not, like, the big businesses are going to have the most success with this stuff, uh, not necessarily, you know, the the small ones. Anyway, But here's it going through, clicking on my ad sets and and stuff like that, finding the highest performers, then enabling them or disabling them. Another cool use case is uploading and then managing things for YouTube. Very similar idea, so I'm not gonna show it here. But what you can do is, in my case, you know, I record a lot of videos via OBS, that's what you guys are watching right now. Um, previously, in order for me to automate the process of uploading them to YouTube, I had to use the YouTube API, or I had to use some sort of third party service that connected with YouTube like Ophonic. And the issue behind uploading things automatically to social media platforms via API is, like, the the current consensus is they will block you or significantly restrict your reach if you do it automatically. Again, because they don't want automatically generated content. They want, like, real user generated content, people to actually click buttons. Well, now you can just have computer used to it for you. You can automate the process, then you can also get all of the upside of, like, having organic reach and brand. The next use case is gonna be pretty straightforward. I'm just going to have it compile a bunch of invoices to me locally on my computer. Obviously, I could do this manually, and there are probably some API based ways I could do this as well. But this uses the exact same interface that I do. Very similar to, like, a humanoid robot. Why do we make humanoid robots nowadays? It's because despite the fact that other types of robots are probably better, our natural environment is geared towards, like, humanoids. Right? So humanoid robots, obviously, despite not being perfect at any one thing, can do everything, which is valuable. In my case, I just have it grabbing the first Apify billing invoice because I pay them for the automation of a few other things that I do. And then it's just gonna save that to an invoices folder. 49 US dollars. Good lord. What's funny is I think it was right across the street from their office when I was in Prague. That's wild. I didn't realize they were headquartered down there. Anyway, it's scrolling down now. And then sorry. I didn't wanna show you guys my address or anything, but it went through. I downloaded it. Now, um, you know, it's in my downloads folder right over here. You can also have it automate some desktop apps as well. Now this is sort of like the more boring use cases that most other people have shown on YouTube videos, but still pretty interesting. You can have it automate, like, Premiere Pro, for instance, to identify, um, low waveform points on a graph and then actually cut right at, like, the lowest point. You could have it do a combination of, like, terminal based commands, and then I'm just thinking about video editing because I do a lot of video editing right now. GUI based commands to both shorten it, uh, you know, have Premiere Pro generate captions for you, slap the captions on the video automatically, and then do a couple of other things. Uh, you know, really, the sky is the limit there. And, unfortunately, there's just so many different products that we could use there. I don't feel like I'd do a substantive positive service by demoing just a couple of mine. Okay. And then finally, QA testing. Because we are using our mouse and our keyboard here, we are gonna be interacting with the screen with Plot in a very similar way to in which a human would really realistically interact with the screen.

00:11:26.580 --> 00:11:44.515
A lot of the ways that people have been doing some sort of QA testing so far has been using, like, Chrome DevTools MCP or browser automation. And that's okay because, you know, it still, like, does things on the computer, but it doesn't do so in a way that a human being might. For instance, it, like, doesn't click specific parts of the page. Right? What it'll do is it'll, like, run JavaScript that simulates a click event.

00:11:45.010 --> 00:11:48.770
So what I've done here is I've said, hey, I want you to QA test at nixrife.com.

00:11:48.770 --> 00:14:10.745
Go through my entire sign up flow. Try to break it. Screenshot every step. The goal is some form of real user simulation. I just want you to actually, like, do it the way that I would do it. And you could see that it's, uh, it's gonna try and break it in, a million different ways. And the reason why this is valuable is because sometimes there are, like, buttons on a page or features in an app that aren't immediately accessible to something like a browser automation version. Personally, I think this is what, like, a lot of big teams have probably been doing for at least a little while. Uh, I would almost guarantee you the Anthropic team, in particular, has probably been using it to, like, stress test their products. But it's gonna start clicking various buttons. It's gonna try all the different validation approaches and so on and so forth. And this isn't a product. It's just my own personal website. But hopefully, you guys can see how you'd able to scale up an approach like this to make it as as substantial as you want. Alright. So how do you actually get this set up? It's actually really straightforward. First of all, this is in research preview mode. So some of the features that I just showed you may be invalidated if you're watching this video, like, two or three months from now. But realistically, all you really need to do is just, like, download and then install the Claude code desktop app. So you can do that by just going to Claude desktop app. K. Give that button a quick click. You'll go to the downloads page, and then we have a couple of different options here depending on if you guys are on, like, Mac OS or Windows. But in my case, I'm on Mac OS, so I'm just gonna click download on Claude desktop. That's gonna open up this page in the top right hand corner. You can see it done this a couple of times because I wanted to reinstall it. Then, um, all you do is you drag the Claude app over to applications. I'm not gonna do that because it's gonna reinstall it for me and, you know, already have it installed. And then just type Claude to actually open up the app. Once you're done with that, you'll have access to three different tabs, chat, co work, and code. Right now, it's available for co work and code, not inherently chat. But in order to do this, just go to co work. And then now you can basically say, hey. Do x, y, and z. It may not be enabled for you guys right off the bat. If that's the case, head over to settings first, and then underneath general, scroll all the way down to where it says browser use and computer use. Enable browser use because that'll allow it to use the the Claude Chrome extension, which allows it to do some browser stuff. And then also enable computer use, which allows it to take screenshots and control your keyboard and mouse in apps that you allow. Then you could do any of the demos, you know, that I say. And, you know, just in order to maximize the probability that it actually uses the computer use feature, I have so far just been appending or prepending my request with the term computer use. Otherwise, sometimes it'll just try doing API stuff. The reason why is probably because it's very token intensive, and I'm I'm sure they'll figure out ways to make it less token intensive over time. But right now, this is the simplest manner. Okay. And then the last thing is you need some sort of browser app that does both read and write access. The one that I'm using right now is called minbrowser.org.

00:14:10.970 --> 00:14:22.810
If this does, for whatever reason, get patched or whatever, you can also just look up other browsers that are safe, hopefully open source, that'd be ideal. The way I did that is I just went into Cloud Code and I said, like, find me 20 different browsers

00:14:23.175 --> 00:14:50.690
And, you know, one of those ended up working. Okay. So that's it. Hopefully, you guys appreciated this look and how to actually make money with Cloud Computer Use. It's not just fancy toy demos here, but, you know, if you apply it to a couple of interesting workarounds, you can make some cool things happen, and you can also take advantage of browser fingerprinting and stuff like Aside from that, if you guys enjoyed this video, please subscribe to the channel. Something like 70% of people aren't, and I'm trying to grow as quickly as I can right now. So you'd doing me a big solid. Check out my four hour Claude code course for more on how to use Claude for economically valuable ends, and I will catch all of y'all on the next video.
