Nate Herk: The EASY Way To Use AI Agents To Make Money & Build Automations

[SPEAKER_00] When you build something in Claude Code, it's incredibly easy. You don't have to be technical at all. I said right here to Claude, I want leads, go get them now. And it does it. And this ran in under 10 minutes. There are 50 leads in here, just like I asked for. And then what it does over here is it creates me a personalized email subject and email body. [SPEAKER_01] You start using AI within a year or two, it is doing $250,000 a month. What are the opportunities that you're even seeing? [SPEAKER_00] I think the advice that I would give myself would be to pursue a service-based business in AI. [SPEAKER_01] If someone pulls up the desktop app with Claude Code, what are the first few things that they should even be doing? There are a few things that you want to do right away. Before we get into the episode, I have to say that it's pretty surreal to me that I get to do this and record this show for you guys. And so many of you tune in and leave comments and likes on the content. [SPEAKER_01] The reason why I do this show and I record every single week is that I want to show that anything is possible. Any dream, any delusion that you have in your mind right now, you can bring it to reality. One of my goals at the moment is that I have this crazy delusional dream that this time next year will reach a million subscribers on this channel. And so if you click that subscribe button, the promise that I'll make to you is that we're going to improve this show every single week. [SPEAKER_01] And so I want to take you on that journey with me. So hit that subscribe button and let's get into the show. This morning, I was watching one of your videos, actually your course on Claude Code. Okay. And right in the beginning, you said something that stuck with me. So I wanted to give it back to you. You said, I don't have a technical background, but I've been able to do some pretty incredible things with Claude Code. [SPEAKER_01] I've got multiple businesses, content, education, events, consulting, and all of it is powered by a small team that just effective at using AI. And so, Nate, I think about the person that's listening to this at home that's curious about AI and Claude Code. But even when they hear that, they feel like it's too complicated for them. It goes over their head. This is for technical people like coders. [SPEAKER_01] If someone has clicked on this video with that feeling, before we even get into the detail and the weeds of it, what would you want them to know? [SPEAKER_00] I would want them to know that getting results out of AI at this point is just about communicating clearly what you want. And I think as humans, we're very good at knowing what we want and we kind of get in our own way. And I think it's really interesting you even hear these people, Boris Cherney, Sam Altman, these people that are making these foundational models are giving advice like give AI a task that you wouldn't expect it to be able to do and just see what happens. [SPEAKER_00] Because it can really do some incredible stuff. And like I said, it's really just about communicating clearly what you want and communicating clearly what good looks like. And I think those are just the two most important things. So, yeah, you don't have to be technical at all. [SPEAKER_01] Yeah. You know, I think it's such a big thing because I think a lot of what holds people back is like this intimidation factor. And it's interesting because you and I were having a conversation a few weeks ago, even in the planning of this episode, you said something that was interesting to me. You said there's a model that's so strong that the government took it away. You're talking about Claude Fable. But then you said on the other side, we have people that haven't even opened ChatGPT ever. [SPEAKER_01] And that gap is ginormous. I'm just curious because you come across so many people with your consulting, with what the educational things that you do with your content. It almost sounds like there's like two worlds when it comes to AI. And there's like the power user and then there's the person that's really not fully engaged with it yet. Can you talk about what you meant at the beginning, what you meant in that quote when you said that gap is ginormous? [SPEAKER_00] It really does feel like there's two different worlds. It really feels like me and all of my buddies in the space and everyone in my communities, it feels like we're living in this this this AI bubble. And we're trying to all keep up with this new cycle. And there's so many different things coming out every day. And we're seeing all these crazy things that people are doing with AI, like what happened with that model being so strong that it got taken away. But then on the other side, you know, whether it's because people aren't interested or whether it's because they don't have access to the right tools. [SPEAKER_00] Like, for example, some of my friends that I graduated college with at their jobs, they're maybe only allowed to use something like co-pilot. And so their experience of using AI is it doesn't really give me the right answers or it puts the wrong formulas in my Excel sheet. So, uh, AI is hype. It's I haven't felt the ROI myself. So how could I ever like be convinced of this ROI? And I think that's the, the moment people need to have is where they enter in one prompt and they get something back. That isn't just an output that they have to fix, but it's an output. [SPEAKER_00] That's like, wow, that would have taken me an hour to do. And I got this in two minutes. And once you have that moment yourself, rather than just like seeing a demo or seeing a LinkedIn post, I think that's when it's just like it opens your eyes up completely. And now you're in, then you've entered the AI bubble. [SPEAKER_01] Yeah. Yeah. You know, it's so good is that I think there's Nate, there's going to be people that listen to this conversation until the end. And they have that moment like after listening to this. I hope so. And so you said something which is interesting, which you talk about the people that use like AI in the workplace and they use co-pilot and it's like not really this great experience. And I think for that person, because I've been there, you then hear like terms like Claude code. [SPEAKER_01] And even if you see people and like the value and what people are able to build, it just sounds technical. And I think it's because it has code in the word. Yeah. Like it just sounds like, okay, you need to kind of be a coder, a software engineer to get this. Can you just make it clear for people, maybe those people that have used co-pilot or they use chat GPT here and there. [SPEAKER_01] What could be the impact in their lives if they're able to internalize some of the things that we show them in this conversation and actually get started using Claude code in some of the ways that you're going to demonstrate? What was going to be the impact, the significance of that? [SPEAKER_00] I think what you're going to feel immediately is that you now have a system that you aren't having to repeat yourself to. So if you are a frequent Claude or chat GPT user, you're probably feeling a lot of value already because you're generating docs quickly or you're generating presentations quickly. You're doing research quickly. But you probably are having this moment where you're thinking, I have to re-explain my priorities. I have to re-explain what my team looks like or what we're driving towards or like my brand guidelines. [SPEAKER_00] But when you build something in Claude code or codex, you're building a system that has memory. And I joke with my team now, like if you want to get an answer from me, you're better off reaching out to my AI operating system than me because it's going to respond faster and it's going to have a better memory than I do. So not only are you creating something that compounds, it feels less like a coworker. It feels more like a co-founder. [SPEAKER_00] You're also getting things that can, instead of having to do the copy and paste, oh, copy this from chat GPT into my email or copy this and put it over here. It just has the hands to actually take action as well. And it can then start to do things while you sleep and you wake up and you're already like, you know, halfway through the day. So I think overall what you're going to feel is if right now one unit of effort, let's call it, is getting you three units of output. [SPEAKER_00] Once you've adopted this sort of new way of working, one unit of effort is going to get you 10 units of output. And that's just going to become the new baseline. [SPEAKER_01] Like it just makes you way more productive. Way more productive. It's interesting that I got excited in the lead up to this conversation because specifically for like non-technical people, because a few years ago, like you were that person. Like you were the person that was starting to use some of these tools. You were non-technical. You're actually like an intern at Goldman Sachs. [SPEAKER_01] And then in just the last couple of years, you've gone from that to you're running multiple businesses to the point you even you're a successful creator as well. To the point where you even had an agency where you're making hundreds of thousands of dollars a month at one point. Can you just high level, just map it out for people? Because it sounds almost like unbelievable when you hear it, like just the speed of it. [SPEAKER_01] But how were you able to go from a couple of years ago working a pretty like an entry level job at Goldman to then being where you are now? Like how does that even happen? [SPEAKER_00] It has happened very quick. I've always been in the business of let's help small teams be empowered to do more rather than let's scale without ever hiring humans and let's fire a bunch of humans. That's always been the way I think about it. But ultimately before I brought on the team, you know, we were I was at some pretty significant milestones and was growing very fast. The channel was growing very fast. The communities. So was my freelancing business. And I think. [SPEAKER_00] One thing before I even talk about like the AI systems I used, it all was very cohesive. But then from there, I was able to build a bunch of systems because I identified repeatable and just like kind of boring processes in my own workflows. You're not asking your agents to do things. They're just taking action proactively. And when you get into that sort of world, it really opens up a new opportunity for you. I started the business in November. In December, I got really, really sick. Ended up going to the hospital. [SPEAKER_00] And during that month when I couldn't really do anything, my business, if you look at the revenue, if you look at the views, nothing really took a hit. Because it was still just me at that point. But I had the right things in place where I was able to just kind of not be the core constraint. And so even if I wasn't sitting down working every day, my agents were still letting me know what needed to be sent. Or reaching out to the clients and letting them know that we had this bottleneck or whatever it may be. [SPEAKER_00] And that's when I realized how much you can really do as one person. [SPEAKER_01] Yeah. You know, one of the things, and it's crazy for me, Nate, like sitting in this chair and getting to speak to people like yourself. Because, and we'll get into it in a few minutes. You'll share your screen and you'll like show these workflows. And it opens up my mind because I was like, I didn't even know that that was possible using AI. Like I didn't know that it could actually do this. [SPEAKER_01] And it's funny because we always hear about this reality where like, people used to talk about it even a year ago, 18 months ago. But it seemed like this distant reality where we'd have like AI employees that do tasks on our behalf. And so I think about that story and that image of like, unfortunately you being in hospital or like what happens all the time for us in work. There's some times where we can't work because of unexpected circumstances. [SPEAKER_01] But then it's like these systems and these workflows were able to take over for you. And you know what, Nate, here's where I want to go. Because we're going to specifically go into Claude Code. And it's interesting because I had a conversation with this creator named Olya a few weeks ago. And he showed, and it was like this stunning moment for him where he showed how he was able to build this website and like digital membership platform using Claude Code. [SPEAKER_01] And in the first week that it was live, he made $20,000 for it. And it was like, I built this in a weekend using Claude Code. Can you just share for people, before we go into the workflows and everything, that you have, what was kind of your first moment with Claude Code specifically, where you were like, it's kind of unbelievable to me that I can even build this. [SPEAKER_01] And it actually led to like this downstream effect and like impact for you, either in the form of like saving time or helping you even generate revenue in your business. Like what was that first almost spark moment with Claude Code for you? I see you smiling even. Okay, so what I found is that one of the easiest ways to use AI to be more productive and save time is to connect your favorite tools to the AI model that you're using. [SPEAKER_01] But sometimes what you'll find is that your favorite tool won't actually be available as a connector with inside Claude or ChatGPT. And that's when you'll want to use the Zapier MCP inside Claude or ChatGPT. For example, the other day I wanted to connect YouTube to Claude so that I could do some research on thumbnails and titles that are working in my niche. So I went to the connectors tab looking to find the YouTube connector, but I couldn't find it. And it was this really frustrating experience. [SPEAKER_01] But after asking Claude for help, I found out about the Zapier MCP and how it could help me connect YouTube to Claude. And all I had to do was go to Zapier MCP, search up Claude here, click on it, click add apps here, and then just search up YouTube. Click on it and then click select all tools, hit connect, and then hit connect again. I then typed in my API key. And then once YouTube was connected, I click on connect right here. [SPEAKER_01] And so you can follow this simple two-step process to connect Zapier to Claude. And then after that process is complete, I can come in here and tell Claude that it's connected to YouTube through the Zapier MCP. And I can ask it something like, please go find the thumbnail and title formats that are hot right now in the business podcast space. And it went ahead and did it for me. And so if you ever run into the same issue where you can't connect Claude or ChatGPT to your favorite tool, [SPEAKER_01] go to the link in the description and use Zapier much in the same way that I just did. And you will get your app connected. Okay, so a lot of you watching this episode might be struggling to get bigger customers for your business, or maybe even struggling to scale your business. And it might be because you don't have a way to prove that you're trustworthy. Even testimonials aren't enough in this new era of AI where anything can be faked. And so potential customers want to see proof that you're secure. And one of the easiest ways you can do that is by using Vanta. [SPEAKER_01] Vanta is the market-leading agentic trust platform that gets you compliant fast with in-demand frameworks like SOC2, ISO 27001, and HIPAA, and keeps you there. And so when you show these frameworks to potential clients, they immediately trust you more than other companies that don't have these frameworks. Take Hyperbound, a Y Combinator-backed startup. They used Vanta to get compliant proactively. [SPEAKER_01] They proved to enterprise buyers that they took security seriously. And they unlocked the Fortune 100 in less than a year and generated $1.5 million in under three months. And now the cool part is you can access the Vanta agent within other tools that you use on a daily basis like Claude, which is why Vanta is trusted by more than 16,000 companies like Ramp, Harvey, and Writer. [SPEAKER_01] And so if you want potential customers to trust you more, hit that link in the description to get $1,000 off at Vanta.com slash Callum. [SPEAKER_00] Yeah, because I remember it. I remember literally feeling like a kid on Christmas when I discovered what I could do. And at this point, I had been running my YouTube channel for a little over a year. Because I only really dove into Claude Code about six months ago, honestly. Maybe a little bit, seven months ago, in like January of this year. So I had kind of built my whole brand around N8N, which was an automation building platform. It's kind of like Make.com or Zapier. And that's kind of, I was like the N8N guy. [SPEAKER_00] And when I first picked up Claude Code, what I went to right away was, okay, let's see if I can use Claude Code to help me build an N8N workflow. And I'm pretty solid at N8N. I could build automations in N8N, you know, a pretty complicated system in maybe 30 minutes to an hour. That might take the average person a couple hours, right? And I started talking to Claude Code about, hey, could you connect to my N8N account? Yeah, of course I can. Just give me this credential. Cool, I gave it the credential. Can you see my workflows? Yeah, I can see all of them. [SPEAKER_00] I can see you've got these in production. I can see these have run this many times today. Okay, could you build me one? Yeah, what do you want me to build? And so I basically just sit there. I yap into my microphone for two minutes. And it says, okay, five minutes later, I come back and there's a workflow done. And I go into N8N. And this thing is like almost done. It wasn't perfect, but it was almost done. And all I had to do was I remember mapping a few variables and putting in some test data and hitting run. And it ran all the way through, gave me the output I was looking for. And I was like, that just took me 10 minutes. [SPEAKER_00] And I went off and I grabbed a snack, got more water, came back, and it was done. And that would have taken me probably 30 to 45 minutes of sit down, focused work. And it was just done. And all I did was talk for two minutes. Then from there, I just started pushing it as hard as I could to see what was possible. But that was the moment where I was like, this is going to change everything about how I work. [SPEAKER_01] Yeah. You know, I think about that visual of like everything you asked it, if it could do, it was just like, yeah, I can do it. Yeah. And it's even just that visual of you like talking into the mic and then like a few minutes later, it's like there's actually something built. It sounds almost like science fiction, you know. And one of the things that I wanted to ask you is because what you described sounds very simple. [SPEAKER_01] And I think because I also put myself into this category where like I've historically kind of avoided, procrastinated away from using clawed code. And much of the reason why was the terminal was like this terminal, like I'm non-technical, I'm not a coder. And the terminal just looks, I don't feel like you get more technical looking than the terminal. Yeah, it's ugly. So like, I guess my question to you, because the way that you describe it yapping into the mic, it doesn't sound like you even needed the terminal. [SPEAKER_01] Can you just clear that up of like for someone getting started, you even need a terminal to get started? How easy is it to get started? [SPEAKER_00] It's incredibly easy. And now they have a desktop app that's pretty solid. The desktop app looks almost the exact same. It looks almost identical as clawed in the web that you're probably used to. So when we pull it up later, you'll see that. You don't need to ever look at the terminal if you don't want to. But yeah, I mean, I wish like they have clawed cowork and clawed coworker's really good at automation, but clawed code is just the most powerful. But I do wish they didn't even call it clawed code, because I think that word, like you said earlier, [SPEAKER_00] does scare a lot of people away from opening it up. [SPEAKER_01] Yeah. You know, Nate, I just think about what you mentioned earlier with like the impact that this technology has had in your life. And so I just want to get into it for people. So can you share your screen and just show us high level some of these different workflows that you've been able to build using clawed code that have helped you save time or even generate money revenue for your business? Kind of give people that picture. [SPEAKER_00] So this is what we're looking at right now. This is the clawed code desktop app. And as you can see, you've got a chat down here. You have a few little options here that's like it's choosing a project to work in, essentially. But, you know, if you're over here in chat, this is really all it is. You have all your chats on the left. You can switch here between chat and coworker. You can switch between the different models. And you just talk to it. And that's the exact same way you should approach using clawed code. [SPEAKER_00] You know, I think a big part of working with AI is so mental. It's all about mindset. And I actually wrote an entire book. It's like, it's right up there. I wrote this book called Becoming AI Native. And the whole book is just about changing your mindset. It's about what do you reach for first? And how do you talk to the thing? And how do you have the thing prove to you that it is done and that it did good work? It's so much just about being a good, like, manager. [SPEAKER_00] If you're someone who has had people work for you and you set expectations and you tell them this is what good looks like, then all of those skills transfer really well over to just like managing agents. And I think that's a big kind of like thing for people to realize. [SPEAKER_01] From the people that you've interacted with and the conversations that you've had, what do you typically see as the biggest mindset shift? [SPEAKER_00] I think there are two big moments that I always think of. The first one is exactly kind of what I told you earlier in my story where just be curious. Just ask, is this something that's possible? Can you do this? And you'll realize that it's not a binary question. You know, it's not can AI do this yes or no? It's to what extent can AI do this? So that's kind of the first piece. The second piece that I think that is really cool, and I'll show this off later, is what [SPEAKER_00] normally happens when you use AI? You get some sort of output and then you're like, okay, I'm the human. I have to verify that this is good. You know, I have to check this. I have to test it. But you can also have AI test it for you. So that way you're not getting the first output. You're getting like the eighth output because the AI agents or maybe like teams of agents have reviewed it, have discussed it, iterated on it. They've proven to you that it's good before they even give it to you. [SPEAKER_00] So now instead of AI getting you 70% of the way there, you're getting something that's like it really just needs a quick review and it's already good because you've set the standard of this is what good looks like. Don't stop until you've hit this metric. You know what I mean? [SPEAKER_01] I believe that I heard you say it. It's like AI prove your work. It was like this process. And we're going to show it because I thought it was so good. We're going to show it later. Of there's this actual process that like steps that you build into your process where the AI actually proves its work. And there's like things that you're doing, like tactical things that you're doing to add that in. 100%. But yeah, let's show some of these workflows. [SPEAKER_00] Cool. So yeah, there's four things that I wanted to show off today. For some of these, I have the original prompt and you guys will see it's just one prompt. And then some of the other ones, I just have like the final output to show you. Let me start off with this one, which I think is always pretty cool to show. So I said right here to Claude, I said, I need you to use my clay. So like, you know, the software use clay inside of my project to help me find 50 enriched and optimal leads for my business that I can reach out to today with personalized outreach [SPEAKER_00] messages and give this to me as a Google sheet. And so what's really interesting about this prompt is I didn't instruct it here. What are optimal leads? You know, I didn't have to tell it about my business because it can look through my project and it can figure out who is Nate trying to sell to and who would his, you know, ICP be. And then it goes ahead and I'm like, basically what happens is it starts to go through tools, right? Like it will call things, it will think, it will reason. [SPEAKER_00] And that's what you call the agentic loop is basically like thinking, taking action, inspecting, thinking, taking action, inspecting. And it just keeps looping through and you'll see it does a lot of things. And then all the way down at the bottom, it basically says, okay, cool. Here is your deliverable. So just super quick, Nate, what is clay? Clay is a software that has a bunch of B2B leads. It's got B2B leads. And then inside of clay, there's also like automations you can build. So it's been around for a long time. [SPEAKER_00] But what I really like, the way I like to think about cloud code or codex or whatever tool you're using is I don't really want to learn a new interface. You know, I don't want to learn a new tool. But all I had to do was give Claude my API key. So essentially my password to clay. And it went to clay and it figured out how to use the entire system. So all I say is I want leads, go get them now. And it does it. And I don't have to learn a new tool. So that was the value prop here for me with clay. [SPEAKER_01] You interact with cloud code. Cloud code figures out how to get what you want from whatever the tool is, clay. [SPEAKER_00] Exactly. Exactly. That's what I think makes it so cool is you can, a really good quote that I always quote is you can outsource the thinking, but you cannot outsource the understanding. So you have to understand why are you asking Claude to do this? And when it comes back with data, what does the data mean? And how do you use it? But you can outsource as much thinking as you want. Go figure out how this tool works and then explain it to me. Go research all of this stuff and then bring it back to me. [SPEAKER_00] Everything, everything that I learn, everything that I make a YouTube video about, I have Claude teach me. It does the research. It tests things. And then it comes back and says, hey, here's what you need to know. So let me show you this output that it gave me. And this ran in under 10 minutes. So now I have this Google sheet. You can see that there are 50 leads in here, just like I asked for. And all of these, if you look at the title, president, owner, company, owner, business owner, co-owner, co-owner, general manager, owner, president, business owner, company, [SPEAKER_00] like founder, president. These are all people that have decision-making authority. What I did is I have this demo project where I have a fake business in there and all of these people, you can see all of these leads fit into one specific category. And it didn't just find leads for me. It enriched them. It did research on them. You can see that we have like information about their reviews. We can see what they're getting on Google, on Google maps. We can see if they're, you know, advertising 24 seven, if they have online booking, we can see all this information. [SPEAKER_00] And then what it does over here is it creates me a personalized email subject and email body. And so now all I could basically just say, hey, can you plug this into clay? Cause clay can also do email campaigns for you and say, okay, cool. I've read through those. I approve them. Go send them off for me or go schedule all of them. And so now I've been able to just kind of build a system where whenever I want to send off more cold outreach, I can just ask for it. So that is one quick example that I think is, is really, really cool. [SPEAKER_00] You can see it comes through and it verifies all this and it does all the research. And then what you guys didn't see is in the process, it didn't stop working until it had 50 enriched leads. So it probably got like 50 initially. And then it realized, okay, I, I, I reran this and I checked through and some of these weren't good. So I scraped 10 more and it just kept doing that. What I call the verification loop until it knew that what it was going to give me was good. [SPEAKER_01] And to be clear, Nate, that verification loop, is that because of something that you've set up with your Claude code or is that just natively like out of the box? [SPEAKER_00] It just will do that. The harness Claude code is a harness wrapped around the model and the harness will do few checks. And I I'm assuming every single month when they push out updates, it will get better and better at verifying on its own. But that is something that I bake into my instructions and my skills and my prompts that it should basically never give me something that's a first pass. And real quick, before we move on to the next one, I wanted to show you guys this. [SPEAKER_00] So in this run, it was nine research agents that fetched all the company data and found the hooks and everything wrote all 50 emails. And then right here, you can see two adversarial verifiers fact-checked every claim against the recorded site facts, and they caught five real problems and all five are fixed. So if I didn't have this verification, I would have had to catch those problems and I might not have even caught those. And then maybe I'm sending out data that's wrong or I'm emailing the wrong people. So it found those and fixed those. [SPEAKER_00] And then it gave that to me, which is obviously huge. [SPEAKER_01] Yeah. You know, you know what I love, Nate, even as we just do the next workflow. And this is why I got when we spoke about this before, it's why I got excited. I think that the reason why I think there's a group of people that start using Claude code or like a codex or a program like this, and they get super excited. They, you know, they start yapping into it similar to what you did. And then the first output that they get is just not good or it's like mediocre at best. [SPEAKER_01] And it like punts them off and they're like, okay, this thing was just all hype. Like it was all a hype cycle. And so the idea, I feel like what you demonstrated like displays it so clearly. It's like, oh, you saw the first output, but you can actually build it into the process where you're not seeing the first output. You're seeing the one where it, you know, it checked itself and it iterated by itself before then actually giving you a good output. [SPEAKER_01] And I think that would just change the experience of so many people is being able to see that. [SPEAKER_00] Absolutely. Absolutely. Yeah. I talk about that a lot and I call it the bike method, just kind of the understanding that the first time you start building this stuff or building a skill or a workflow or whatever it is, you have to kind of think about it the same way you would teach a kid to ride a bike, you know, because I think a lot of people jump into an AI chat and they start talking. And then it's like, it's like they put a kid on a bike and then just kind of walked away. That's just not going to work. You know, you have to kind of be there. [SPEAKER_00] You have to watch, you have to guide, you have to say, hey, you know, you're leaning too far to the left. Why don't you kind of center out a little bit? You have to give it a little bit of feedback and iterate. And it's only that way that you can then start to, you know, then you put on the training wheels and you can step back and then you take off the training wheels, but they're still wearing a helmet. Like you're still being safe about it, but just the understanding that this is going to take a little time. But once you get there, now the kid can go off and ride the bike, you know, 20 miles an hour down the road and you won't even worry at a certain point. But I'm really glad you called that out because it's super important. [SPEAKER_00] Because that's the worst thing is when you see someone try it, they're not impressed and they give up. [SPEAKER_01] Yeah. The bike method. So good. [SPEAKER_00] Absolutely. So this second use case I wanted to show you guys, you can see right here what I said. I said, I need you to pull me a report on my YouTube analytics through 2026 so far. So I want a Q1 report, a Q2 and a Q3 up to today because we're only about two thirds through Q3. And what I did here is I gave it my motivation because if you just said this, that's also a pretty vague prompt. And this is still pretty vague, but the more specific you can be and the more context you can give it about why you're doing something, [SPEAKER_00] it's able to then use that context to give you something a little bit more tailored. So I said, my motivation here is just to see the trends, see what's been working and identify what type of content I should be focusing on making in Q4. And I asked for it as a Google sheet. It could have made this as an Excel. It could have made a PowerPoint. It could have made a video, but I wanted a Google sheet. So that's all I said. It goes through my project. I don't have to give it a new API key. I don't have to give it new data. It knows my audience. It knows me. It knows how to get to my YouTube analytics. [SPEAKER_00] And it just goes ahead. You can see. And it just reasons. It does the agentic loop. And then it comes all the way back and it gives me a Google sheet. And so similarly, as the clay output, we get a Google sheet that I can open up. I can go ahead and share with my team whatever I need to do. And this one is a lot more data. You can see already at the bottom there are six tabs. We've got the overview, which shows things like each quarter we can see the views, views per day, watch hours, view durations, subs gained. We can see all of these statistics here. We can see the monthly trends. [SPEAKER_00] So each month we can see these same statistics. And then when we really want to get into these next sections, I'm assuming these are all going to be very similar. But just think about how much time this would have taken me to do myself. To format the sheet, that's one thing. But to go through, I don't know, I would say at least 200 videos so far this year. Well, maybe not 200, but definitely in the hundreds, right? 150 videos maybe. [SPEAKER_00] To go through and pull each statistic from all of those videos, find the averages, format the sheet, throw them in here. That would take me probably, I'd say, at least three or four hours of manual work. And this was one prompt, went to the bathroom, came back, and it was done in about 10 minutes. And this is like useful, real data straight from my YouTube dashboard. [SPEAKER_01] Yeah. I need to do this. You know, I'm curious, Nate. Nate, once you get an output like this, like you get a report, like this YouTube report, and obviously you're someone that creates content prolifically. What is it that you do next? Like how did you then go and use this? Did you even, and in those next steps, did you even use AI or was it more so like now you're just analyzing the data yourself? [SPEAKER_00] Yeah. Well, for the most part, this is good to keep context in my system. So now it always knows, okay, we just pulled this report. If I ever need to look, I can go here. But if you go back to what I said about you can outsource the thinking but not the understanding. Normally at this point, I like to look at the data myself. I like to try to form a story out of it. But I still use Cloud Code as a thought partner here. I ask it, oh, why do you think this happened? You know, can you do any research on Q3 and see why did we have the spike? [SPEAKER_00] Or why did these videos do better? And it can pull in other data. The other really, really helpful thing here, which I didn't show off in this specific example, was that it can go grab all the comments. And it can look through 100,000 comments in, you know, five minutes. And it can give you common themes. So I do have automations that are constantly every week saying, hey, this week you had 300 comments. And here are the three big pain points. These are probably things you should make videos on. Same thing with my school community. [SPEAKER_00] Every week it goes through and it looks through all of the tech support questions, all of the business related questions. Here are things you should probably make videos on. So like ideation is not really an issue anymore because now it can get primary data from my sources, but can also look at, you know, competitors or it can constantly be scraping X for news. And I just have all of this aggregation of data being brought to me. But then I'm the one who ultimately wants to be able to make the decision from that data. [SPEAKER_01] Yeah. You know, it's so good. But I keep going back to that quote that you mentioned earlier, that you can outsource the thinking, but you can't outsource the understanding. And on that point, I think the reason, part of the reason, and I'm even just thinking about the person that's listening to this at home because it's what I'm thinking, seeing it. Part of the reason that you're able to get so much value from Claude Code, and it feels from what you're showing, like it's very seamless within your processes, [SPEAKER_01] is that Claude, your Claude Code has access to a lot of the tools and like logins and credentials that it needs to pull this information. Can you kind of just take us behind the scenes for a second? I know you mentioned it with like the API key with Clay or like, I'm even curious here, like how it even has access to your YouTube analytics. Can you just describe high level? [SPEAKER_01] Is this something that you almost did in the beginning of using Claude Code was like setting up these accesses? And like how easy was that to do? Because I can see that being a stumbling block for a lot of people. It's like their Claude Code just doesn't have the access to get into some of these accounts and credentials that would then give them this information so seamlessly. [SPEAKER_00] A hundred percent. Yeah. And I'll go ahead and share this tab real quick. So the way I use Claude Code or Codex, any of my AI tools, I call it an AIOS. So it's a term that's going around right now. It basically just stands for AI operating system. So what I did is I put myself basically through this challenge where everything I do on my computer, instead of opening up Chrome, instead of opening up these different apps, let me just do it only through Claude Code. [SPEAKER_00] And the reason I wanted to bring this up is because there's four main pieces. I call this the four C's of an AIOS. The first C is for context. This basically means, can your AI system have all of the context about you, your business, your priorities, and your goals? And once you get all that business context in there, that's where you get to the point where you're not repeating yourself. That's kind of like stage one. Then you have connections, which means, can your AI touch things that you need to touch every day? [SPEAKER_00] Your email, your calendar, your Slack, your school, your YouTube. And over time, you just build up these connections. And then these last C's are basically about capabilities and cadence, meaning capabilities, you build automations, you build skills. And then cadence means you turn those on so that they're not only triggered by you, but they're also triggered by action so that when you're sleeping, when you're on vacation, when you're at lunch, things are still going on. And so once you set up all four of these things, you're constantly adding more. Every day I add more context. [SPEAKER_00] Every day I add more connections and capabilities. But you've got a really, really good foundation that you can just keep building on top of. And then you just move so fast. [SPEAKER_01] Yeah. You're really good, Nate. [SPEAKER_00] The way you just explained that, you're very clear. Thank you. I appreciate that. That's what I've tried to do. Yeah, you've mastered it. So this next one, I've got two pieces of this that I'm going to share. So right here, I said, hey, so I'm about to jump on Callum Johnson's podcast. And I gave it the brief for the episode. So I gave it just, you know, this Google Doc. And I said, I want you to create me a quick 10 to 15 second animation, thanking him for the opportunity. And tailor this towards what we're going to be talking about today. [SPEAKER_00] And hoping that the audience will get value out of it. I told it to do this with Hyperframes, which is just basically a, it's a little open source project that anyone can use. That helps you create motion graphics. So I said it should be modern, clean, liquid glass, motion graphics that tell a story, not just text. And so it loaded up my Hyperframes project. It looked at any of the Hyperframes skills that I've built in the past. And then it went ahead and created this 15 second video, which I can just sort of play right in here. [SPEAKER_00] So kind of like what we had talked about, what agents can do for you, analytics report, website, Calum Lee clone. The first output is a starting point. And Callum, thanks for having me. Hope you guys get a ton of value. So that obviously wasn't a crazy output, right? But if anyone in this audience has ever animated stuff, like a 15 second clip that you're animating with motion graphics and keyframes and everything could take you an hour to animate. [SPEAKER_00] And a lot of the things that I'm doing in some of my video editing and some of our courses, it's a full pipeline now where I just drop in a raw file and I say, hey, can you like, you know, cut out the mistakes, throw in some motion graphics and then, you know, let me know when it's done. And it can do a lot of that. I also, just because that wasn't a super impressive one, I want to show you this example of a YouTube video that I actually uploaded. And in this example, let me just pull this up and then I'll share it. I basically recorded a raw file. [SPEAKER_00] It was like 15 minutes. And I threw it in this pipeline and said, edit this for YouTube. And then it created this for me, which was, it ended up being about a 10 minute video. But you see here, I've got these motion graphics. I'm going to just go ahead and mute this so I can play it while we're talking. But it creates these motion graphics where it kind of like created this open loop. Oops, there we go. So now we have this open loop where we see what's going on. I get into like these flow charts. So I'm explaining these different processes. [SPEAKER_00] And then what it does is it kind of keeps the spatial awareness, if you know what I mean? Like when it goes to the second use case, this was like use case one, it puts it back here. And it knows to create this open loop where now the audience is, okay, oh, I want to see what number two, three, and four are. And it consistently does this throughout all of them. So if I go to the end of number two, you can see as it transitions from two to number three, it does the exact same thing where it puts it back there. And then we jump into number three. [SPEAKER_00] And as I'm explaining these different automations, it's animating it as like a flow chart. You know, all of this stuff would have taken me so long to animate, you know, myself. But if you see all of these things that are going on now, this was all because I had built out a pipeline for me to literally be able to drop in a raw file and it cuts it up. It puts all the stuff in here and it does the whole verification loop. So like something like this could take up to six, 10 hours. So sometimes I'll run this before I go to bed, but I come back in the morning and I've got [SPEAKER_00] a video now that I'm ready to post on YouTube. [SPEAKER_01] Yeah, that's crazy. So in that example, it's actually like editing the video. So even when it goes from like the animation back to showing your camera, that's the, that's crawled code. That's the AI making that decision like of shift from like the animation moving here to like then show his camera on like a split screen, like those editing decisions, which would be made by an editor. The AI was handling that? Yeah. [SPEAKER_00] Yeah. It transcribes it. It tries to contextualize what I'm talking about. And then it and a few other agents debate on what would be good animations to show during, you know, those 10 seconds or those 20 seconds. And then it creates the hyperframe animation. It renders it in a browser. And then what I have it do is I have it go through and screenshot like every three seconds. And then it basically watches back the whole video and it says, oh, you know, this element was out of bounds or this feels a bit, you know, off brand and it will go through. [SPEAKER_00] So that one that you just saw, actually, this might be kind of cool to show. If I share this real quick again, the output that you saw there was essentially it created this in this folder. So what you see is it created an intro and a V1 and a V2 and it took photos and it came down. And what I ended up showing you just there was the V8. So that was the eight version of the actual animation that it made. And so like this was the full project that I was it was creating in. [SPEAKER_00] It created assets, compositions, the script. It had an index. It had QA, which means it went through when it screenshotted everything. So this was a big project. It created a ton of files from it. But the ultimate render that we got was version eight, version eight master. And, you know, that was obviously a lot of iteration had to go into this whole skill in this whole pipeline. But the cool thing is every single time I use it, it gets better because after I did this [SPEAKER_00] one, I was able to say, I really liked what you did here and I really didn't like what you did here. So update the skills so that next time you do more of this and you do less of that. It's just about constant feedback on the good side and the bad side. And then you literally get in this cool place where every time you use it, it gets better. [SPEAKER_01] Yeah. So I understand that you've like optimized this and even proved it and you even just have this verification like process, which we're going to show people in a second. But for you, I guess how autonomous did it feel to make that video? [SPEAKER_00] Now I'm able to do something called a goal prompt. So a goal prompt exists in Cloud Code and Hermes Agent and Codex. So pretty much every AI harness now is working in goals. But basically when you set a goal, you tell it the job and you say, this is what I want. And essentially it won't stop until it hits that goal. Like I'm setting the stakes. I'm giving it something to work towards. But you do have a good point. [SPEAKER_00] If you don't have, earlier we talked about the connections. If you don't have some of those connections and permissions set up right, then it probably will get roadblocked and say, hey, I need you to allow this. But now I literally say, hey, slash goal, edit this video. I go to bed and in the morning it's done. This is just, it's native to the tool. Cool. So right here I just go slash goal. And it says set a goal, keep working until the condition is met. And it's as simple as that. It'll tell you the goal is active. It'll tell you how long it's run. [SPEAKER_00] So, you know, sometimes I'll run a goal and it will take 10 minutes. Sometimes I've ran goals that have taken six days. And it's literally just ran for six days straight building things. If we want to hop into the next one and I can show you guys real quick the Calendly clone that I built. That was a big slash goal prompt that ran for, it was four and a half days, I believe. So I can, I can switch over to that real quick and show what that looks like. [SPEAKER_01] And so I've been so curious to like see where you're at in that process, Nate. [SPEAKER_00] Yeah, let's, let's do it. Let's do it. This is what we've got now. It's called snag time. And what I think is really interesting is, like I said, I don't come from a technical background. I've never built software products. But what I did is I gave a slash goal prompt here. And I basically said, I want to make a clone of Calendly that my team can use. It's going to be free. So what I need you to do is research Calendly, figure out all of the best features. [SPEAKER_00] And then plan out the build, build it, test it, and give it to me when it's done. So the point I'm trying to make there is, let's say you didn't even know what Calendly was. You could still instruct it to do that because it went ahead and found out everything Calendly can do so that it could essentially clone it. So now I have this little app, you know, it obviously has some authentication. I can go ahead and come in here and log in real quick. [SPEAKER_00] Let me just go ahead and get my, the, the example login that I had here. So now I log in here and you can see that I have this interface. Like it's not designed super beautifully. It's just a proof of concept. But what I have here is an overview. I can see upcoming bookings. I can see how many I had this month, how many hours I can create new event types. So this is a test one. This is a paid scoping session. So someone would actually have to pay. [SPEAKER_00] I connected Stripe. We have a free strategy call. We can set our availability. So this is all basically the exact same way that all of those calendar booking, you know, platforms work. I connected my Google calendar, connected a sandbox of Stripe. So this isn't real Stripe yet, but I guess let's just go through a quick flow. So if I copy, let's do this free strategy call. And I come into here to this tab. [SPEAKER_00] This is what the user would essentially see. They would get this ability to book in a strategy call with Up at AI. We could obviously add a description here if we wanted. We can choose the time. And if I go through here, you can see that it's actually syncing to my live calendar. So if we're looking at Friday, we only have 2.30 to 4.30. And this is the calendar that's looking through. Friday, we're, you know, we're basically booked until 2.30 and beyond. So this is synced to my live calendar. [SPEAKER_00] If I added a event right here, 3 o'clock, and we go back into this form and I give it a refresh, this should now be showing that that slot has been taken. So this is real live syncing. I could go ahead and make an event right here for 4.30 and, you know, put in some information. [SPEAKER_01] It basically is Calendly. [SPEAKER_00] Exactly, yeah. It's missing some of the automation features that Calendly has. But for the bare minimum of being able to send people links to book in your calendar, you really, if you want to, like, don't have to be paying for that subscription. So confirming the booking here, what this is going to do is it sends a notification to you if you're the account owner. And then it also sends an email to the person who just booked in. [SPEAKER_00] If I go back to the calendar, you can see that this just popped up, which wasn't there before, which is our strategy call with Nate. This is the email that I just put in. So that essentially works. And this was exactly when we hopped on that pre-call. This was what I said had been cooking for about four and a half, five days. And then it finished, and I came in, and I was able to see that it tested everything. So if it didn't test everything, there would have been certainly a ton of bugs in this environment. There would have been things that were off. [SPEAKER_00] There would have been bugs with the way that the UI clicked the buttons. You know, you can even come in here, and you can set up your workspace. You can put a logo. There would have been bugs. That's just what happens. And there probably are still tons of bugs. But what happened is I had at least 50 agents create fake accounts, create fake bookings, click through the UI as an admin, change the logo. They came in here and stress tested every button and almost every edge case they could think of so that I didn't have to come in here, [SPEAKER_00] or I didn't have to have my beta users experience those bugs and then send me feedback, you know? So that's the way I think about all these automation, software, building websites. Just have a bunch of agents do the work that you would normally have to do anyways, because they can access a computer. They can click around. They can take screenshots. They can, you know, submit things way quicker than you can. It's really, really cool when you just have that mindset shift, right? Because all of us in the world are possible. [SPEAKER_00] Like, all of us could possibly say, can you go spin up 50 agents to test the app? There's no one on this earth that couldn't instruct an AI to do that. It's just how many people would think to do that. [SPEAKER_01] Yeah. You know, I'm so curious, Nate, with like where you think this goes, especially given your background. Do you like, I'm thinking about the ability of companies, but even more so individuals to basically be able to do what you've just done, which is like build like custom software. And I'm so curious, like if you were talking to the Nate Huck that was like just coming out of college and was like, you know, [SPEAKER_01] you kind of had this like love and this hunger for like business and building things. I'm so curious what you would be doing. Like, what are the opportunities that you're even seeing right now? Because this thing of like building custom software feels pretty crazy to me. [SPEAKER_00] What I would, the advice that I would still give to myself coming out of college, if I wanted to, you know, turn this sort of AI knowledge into a business, I think the advice that I would give myself would be to pursue a service-based business in AI. Helping small businesses upscale. So essentially the agency that we were running, the consulting firm that we were running, I still think that going service-based, there's a lot of value in that because, I mean, there's a huge gap that needs to be filled. [SPEAKER_00] And people are looking for experts to help them fill that gap and, you know, start to automate roles and businesses or process in their business, things like that. But I think what a lot of people are doing is we all don't really know exactly where it's going. I think that humans are ultimately really bad at making predictions, especially in a space that moves so fast. [SPEAKER_00] And so what I'm trying to do and me and my community are like, what we can do is we can stay close to it and we can all learn together so that when we need to pivot, we can pivot fast. And when you work with companies, you get to hear what are their concerns and what are their roadblocks and with different industries, what are they interested in? What are their pain points? And that can help you figure out a lane to go down because right now there's this huge, I guess, kind of buzzword, AI consultant, and it's working really well. [SPEAKER_00] AI consultants are getting good business and they're in high demand. But ultimately what's going to happen is these roles that have an AI prefix, the prefix is going to get dropped and it's just going to be consultants. And if consultants don't use AI, then they're not going to be a very good consultant. You know, it's the same thing like I think about when the internet came around. We had like a bunch of internet marketers or digital marketers, but that's just marketing now. You know, if you don't have an internet presence in some form, then you really don't have marketing. So it's just going to become the new normal. [SPEAKER_00] And that's why I think if you were, for example, educating businesses, going in there, you know, I know some people that not anymore, but they were printing like $10,000 a month plus just teaching businesses how to use ChatGPT, you know, in the early days of ChatGPT. And now they're just teaching businesses how to use Cloud Cowork or something like that. And they're just helping people upskill because it is kind of an intimidating topic and change management is very tough. [SPEAKER_00] But what I've learned through educating people and businesses is I've recognized a bunch of patterns and I've been able to get a better sense of what I think, like where the true value sits. And if I ever wanted to go off and build like, you know, some software products and stuff like that, you need to be able to recognize, you know, pain. You need to be able to recognize how does your service or product solve that pain? [SPEAKER_00] And for what's very specific type of person is feeling that pain and would want you to solve for them? [SPEAKER_01] Yeah. If someone's watching this and they want to like succeed and just win in kind of like this new world that we're going into where AI is at the heart of a lot of things. What for you is actually the correct and very doable, like first step? Like they're at a point now where they're non-technical. They've never even really used Cloud Code. But they're kind of like, okay, cool. [SPEAKER_01] I like the idea of like in the future, potentially being able to run this service-based business. What for you is the right doable first step? [SPEAKER_00] You know, I felt like when I got started actually working with businesses, one of my big jobs that I didn't expect was to manage overwhelm. So I think that if you're in a position where you're going through learning how to use this stuff and learning how to use it effectively, take advantage of being in that headspace of not knowing what you don't know. [SPEAKER_00] And understanding these are the questions I have. These are the concerns I have. Let me write all these down and I'm going to learn these and answer them. Because now I can sort of like hopefully proactively address concerns and issues and areas of overwhelm that most businesses and most people who are starting to learn, starting to enter this world will probably have. And I think that you, from there, what I would do is I would start to try to teach people you trust, you know? Try to see if your parents want to learn. [SPEAKER_00] Try to see if your best friends want to learn. And just try to teach them. Because if you can teach them confidently, then you'll have a lot more confidence going into teaching someone you don't know. But I think it's just important that you are getting out of the world of practice, practice, practice. And at a certain point, you are putting yourself out there and doing it in a way where it protects your brand reputation, right? Like doing free work until you feel like you can really deliver value, I think is something that is really important. [SPEAKER_00] You know, prioritize the experience over the cash. Because I think right now, everyone is kind of starting at the same starting line. [SPEAKER_01] A lot of the times the best person to learn from is only like one step ahead, right? So like we put all of this value on like the expert or the person that's been doing the thing for like decades or years and years. But the person that's like the most relatable and the easiest to learn from is probably actually only like one or two steps ahead of where you are. I want us to kick off this process for people of, you know, they see some of these use cases [SPEAKER_01] and some of the opportunities that we've even spoken about that you're seeing right now. They get excited. They open Claude Code. Like I want to just close the gap for them. And I even want to like read out something you said where you spoke about it earlier. You use this thing called the bike method with Claude Code. And you actually say like, I call it the bike method. That's your framework. And then you say that's how you build with Claude Code. [SPEAKER_01] Can we just start right at the beginning of like someone pulls up the desktop app with Claude Code for the first time? What is it that they need to understand? Or even what are the first few things that they should even be doing that's going to make that experience, that first experience a good one? [SPEAKER_00] Claude Code. So Claude Code works in your local environment, meaning it works out of a local folder and it can touch things locally. And locally just means, you know, in your file explorer or in your finder, depending on, you know, what OS you're on. So it can move around things in your downloads folder. It can rename them. It can organize them. And so what happens when you create a new Claude Code chat, you have to choose the project you're working in. [SPEAKER_00] So you guys have seen throughout this one, I've been in this one called Herc 2, which is what I consider my AIOS. I'm just going to say hi real quick because this basically what happens when you shoot off a prompt inside of this project. Right away, what it will do is it will read through something called a Claude.md if it exists. The first thing you should do is have Claude help you create a Claude.md. This is essentially a system prompt. So if I, for example, let me just actually show this. [SPEAKER_00] If I go to my Herc 2 and I open up my Claude.md, I know there's a lot of stuff in here. It's a little overwhelming, but this is my Claude.md. It says, you are Nate Herc's AI operating system. Your job is to help him spend less time on operations, people management, and admin so he can focus on learning AI tools and making YouTube videos. That is the number one priority. Everything else supports it. And what you'll notice here is my Claude.md, it sets up the context, but then it just routes. It says, hey, if you want business strategy or OTAs, you look in his wiki. If you need corporate structure, you look here. [SPEAKER_00] If you need his voice and style, you look here. So immediately when I say hi to Herc 2, it basically says, hey, what do you want help with? I know exactly where to look, exactly where to touch, exactly what to do. So you kind of want to set something up like this. And this is a file. You don't have to know all that right away. It's going to change every day. It's going to change because you're going to find out new things or make new files. So the first thing you should do in here is you want to create the Claude.md file. And you can just do that by asking. [SPEAKER_00] So if I say, hey, Claude, this is a new project called Hercules Advisory. This is for kind of just a demonstration. Just pretend that you are, or not pretend, Hercules Advisory is a consulting firm. So what I want to do is I want you to help me sort of like just build some automations and stuff in here. All you need to know right now is that what's in this project are some logos and brand guidelines. And that my name is Nate and I run Hercules Advisory. [SPEAKER_00] So it's going to create a very, very simple Claude.md file for us. And that's kind of going to be the foundation where everything important will go in there. So for example, what we could work in there is every time you're creating a deliverable, make sure you verify it two times before you give it back to Nate. And so if you work that into your Claude.md, like under like a rules section or something, it will always do something like that. So that's kind of the first thing to do. [SPEAKER_01] It sounds like you set up this Claude.md file. You set up this context and it actually makes, you take a bit of time to do this in the beginning and it makes everything like way faster and smoother nearer the end. Like it saves you a lot of time in the long run because it's operating off of this instruction manual, essentially. [SPEAKER_00] Yes. And I just realized I didn't even ask it to make it. So I just had to tell it real quick. But what it did is it also, it added that to the memory you can see. It created a markdown file for advisory for our brand. And it also created a memory file for it to look at. But now it's going to create the Claude.md. [SPEAKER_01] But one thing, Nate, because you said this before, and I think this mindset, it's been so helpful to me. But I also just think it'd be really helpful for people that are listening. And you actually said this when we spoke earlier. You said the way that you think about AI and using Claude Code is as if it's like an AI employee. And so you said all it really takes is you're managing an AI the same way you would manage a human, which is you give them the job. You tell them very clearly what you're expecting. [SPEAKER_01] What does good look like? And it reminds me of like when I first started working. And you're like sitting next to your manager for like the first few weeks, like in the office. And they probably would have given you like a job description. They'd share with you like certain onboarding documents. So you kind of knew what you were doing and like the expectations of the company. And just like context on the projects that you're working on. And it feels like a very similar process and mindset. [SPEAKER_01] Like once I started to see it in that way, it kind of clicked to me how I needed to also interact with the AI. Is almost coaching it as if, you know, I'm the manager and it's the new employee and I'm like onboarding it. [SPEAKER_00] Exactly. I think that is the best way to think about it. You're onboarding it. You're teaching it about you. The most important thing when you're working with these systems is how do you get what's in your brain into the AI's brain? You know, because once it knows what you know, it saves that and it won't forget it. But I think that you're exactly right. And you can see here what it did. It created this Cloud.md file. [SPEAKER_00] It says Hercules Advisory is a consulting firm run by Nate. It provides strategy operations, blah, blah, blah. And then it basically marks what the files are. So these are the only four files we put in there. And it knows what they are now. Guidelines, logos. We have this as brand rules. We have our colors. We have typography. It has my voice. It has other things. And I could just go in here right now and edit this. So like if I came in here and this, when I say .md, that just means markdown. Markdown is this format where you can have bullet points and you can have headings. [SPEAKER_00] And then it kind of like renders like this. You can have tables. So it just means markdown. Basically natural language though. And I could come in here myself and just add in a section right here called rules. And I could say, whenever you are creating an output, make sure you are doing a verification pass before you give it to Nate. So if you're building anything visually, take screenshots and show Nate the screenshots to prove that you have already verified. So that's just one method of verification. Screenshots is one type. [SPEAKER_00] Fact checking is one type. Clicking around is one type. There's lots of types. But I just wanted to put that quick example in there so you guys could see that. And you can see how when we kind of like prompt it to do these next things, hopefully it should be able to read that and acknowledge that. So this is kind of what I would now start to describe as like the bike method. What I want to do here is I want to ask it for a report. I want to ask it for just like a market census report on consulting for SMBs. [SPEAKER_00] And I want to make sure that it's like branded, but I'm not going to tell it that stuff. And I just want to see how it comes through. And we're probably, once again, this is the assumption we're putting the kid on the bike. We're going to have to guide it a little bit. And every time we give it a little bit more feedback, it gets better and better at riding this particular bike. And what we're building here is called a skill. Essentially, a skill is a repeatable recipe. So I have like a LinkedIn writing skill. And now every time I ask for a LinkedIn post, it knows to follow like this format and this hook structure. [SPEAKER_00] Essentially, you know, the same way if you're making chocolate chip pancakes, you don't just want to guess every time. If you made them really, really good one time, you'd probably save that recipe so you could make them really, really good every time. So that's what I'm going to do real quick with a very simple prompt. I need you to create a skill called market analysis. And I just want you to do a quick research report for me, create an HTML document of kind of like the consulting market for SMBs right now. [SPEAKER_00] So what you'll notice here is I didn't ask this thing to use my brand guidelines or verify itself. So we'll have to see if it actually does that. And it might be able to read the Cloud.md because we just created it in this chat. Sometimes if you're working in the same thread, it doesn't fully refresh. So it might not. But the point I'm trying to make here is it's really important to watch what they do, because this is where people tend to just like shoot off a prompt and switch windows. [SPEAKER_00] But when you're watching the kid ride the bike, you know how to steer it. We can see right here, I'll do this in two parts. I'm going to build a skill first and then I'll run it. So what it's doing, if we click in, we can see that it's creating a skill. So it's creating the skill and now it's creating different agents. And you can basically just watch what it's doing in natural language. It's running commands. And this is where we're able to course correct it if we notice that it's doing something that's just like really wrong. [SPEAKER_01] Yeah. I actually really like that visual of like the bike method. And it's funny, I haven't thought about this in years, but I think about my dad actually teaching me how to ride a bike. And it wasn't, you know, like they kind of get you on the bike, but without the stabilizers. I don't know if you guys call it the same thing here. [SPEAKER_00] We call it the training wheels. That's what we call it. The training wheels. [SPEAKER_01] And it's like, you're not on the training wheels anymore. And he pushes you off and he's like, just pedal, pedal, pedal. No. And I feel like what you described, what most people do is the equivalent of like, if you push the kid and then immediately just go went and did something and like, just let them go down the hill. But like, what most parents do is that they'll actually stay like right there with their child in case they look like they're about to like fall so that they can actually catch it. [SPEAKER_01] And that's actually, the bike method actually very nicely crystallizes kind of like the approach and the method that you're displaying. [SPEAKER_00] Totally. Yeah. Well, I'm glad it resonates. I'm glad it resonates because yeah, I mean, that is way too common. You would never, you would never put your kid on the bike and then go in the kitchen and take a nap or I don't know, might not nap in the kitchen, but you get the point. So I'm glad that resonates. So you can see it literally did all these searches, right? It searched for SMB consulting market. It searched for management consulting industry trends, pain points, fractional executive CFO studies and stats. It said research is done. [SPEAKER_00] Now I'm going to build the report, embedding the logo. So the HTML is self-contained. So it looks like it's going to understand that it should kind of build this with our brand guidelines in mind. And now it's creating the actual HTML. So this might take a few minutes. You can see it's taken two minutes and 30 seconds so far, but there's a lot of value in just reading what it's doing. Because I think you've heard this term black box. We don't exactly know what the AI is doing. Claude Code is a harness wrapped around the AI model. [SPEAKER_00] So Fable 5 is the model which powers it. But the harness, the analogy is basically like Fable 5 is the engine of the car, but the actual car is the harness. And you can switch out the engine within the car. So I think that there's so much value. The way that I learned Claude Code, I didn't watch YouTube videos. I just got in here, I asked questions and I just watched. And I remember doing that for a couple of days straight until I felt like I really understood how it thought. [SPEAKER_01] I'm so curious because I actually think that's really good advice. Take us behind the scenes of that process for a second of you hit the pro, like you click enter and then you just watch. What exactly was it that you were even looking out for? Like what would make you jump in? Like kind of just almost describe like what did you find was the most valuable way of actually like learning from just watching Claude Code work? [SPEAKER_00] Yeah, well, when I first got in and they've cleaned up some of this language, but it used to just say like bash, glob, grep. Like it would say all these random words that sound like a different language, but those are just like they're, you know, terminal commands. They're searching for things or they're executing commands. And you can see that this research report just popped up and it looks pretty branded, which is solid. And so we'll dig into this in a sec. But essentially what I was doing is I was trying to understand what it was doing and why. You can see it's taking screenshots right now. [SPEAKER_00] It's proving to me that this is good. You see, these are screenshots that put back in the chat. But what happened was I would watch every line and I would basically like highlight something and paste it back in and say, what did you do here? Why did you do this bash? Or what is this glob? Why did you do a glob? And I just wanted to understand, you know, and this was, I think, just me being genuinely curious about how it works and trying to figure out why it does things. [SPEAKER_00] But I felt like I was able to, once I understood a little bit better what it did, I felt like I was able to reverse engineer that into understanding how to instruct it better. So now this is done. It has two deliverables for us. The first one is a skill called market analysis. And this skill can be invoked with a slash command, meaning I could come down here and say [SPEAKER_00] market analysis or Claude will automatically run the skill if it thinks that it's applicable to my prompt. So if I said, hey, can you just like do some research on this news announcement I just saw? Like how is that going to affect my business? It would probably run the skill because it knows what type of deliverable I'm looking for. And then it also created this actual HTML report. Um, and you can see, I showed you guys like how it took all these screenshots. It was basically verifying that it looked branded and that it was all good. [SPEAKER_00] But what you'll notice is it didn't do any other type of verification. So for example, it shows me these sources. It pulled 16 sources. If I clicked on these links, it would take us to the actual source. But one other thing that I would probably want to do here is I would want to make sure that all of these claims don't conflict with each other and that they are verified sources. But for a first pass, it's branded. We have the dates. We have, you know, like executive summary, market size and growth. And it created this as an HTML that you could once again, like send to the team. [SPEAKER_00] But this is the part where, you know, if you actually read through this, you'd probably have some feedback. You probably wouldn't like how certain things were structured. Or like I said, maybe you'd want to make sure that all of the claims are actually checked. So then all I would do is I'd be like, hey, this looks good for a first pass. Um, I think that I would like a little bit more information about the market size and growth. That is really important to me. And every time I ask for a report like this, I want you to really focus on that element. [SPEAKER_00] And when you, you did a good job with the screenshot verification, but I want you to also always have a source verification. So just doing a double check with a different agent that every single fact that you pulled is verified and it's correct. And it's not like outdated because if you're pulling, you know, sources that were two years ago, that's going to be not relevant for right now. So go ahead and make those changes and then update the skill. And that's really all it is. You know, it's, it's telling the kid, hey, you know, you were, you weren't pedaling fast enough and you were leaning too much to the right. So let's try again. [SPEAKER_00] And next time, just remember that. And that's kind of the whole loop. [SPEAKER_01] There's something very subtle that you did, but I actually think it's really important, which is that instruction of like, you gave it the feedback and then you said moving forward or every time that you do this from this point, do X or update the skill. And I think it's one of the things that I've learned recently, which is if you don't do that step, it's like the next time you're generating a report, it's the same thing. And that's when you actually start to lose time. [SPEAKER_01] You know, like the time is money kind of thing. Like you start to, it starts to just cost you time in actually repeating it. So actually going back and being like, update the initial skill. And that's when you actually start getting to this place where it like gives you these outputs that consistently like hit the mark. You're like, okay, this is perfect. Like you'll, you'll start iteratively getting to that point where it's like, this is perfect. [SPEAKER_00] Totally. And I think that one thing that kind of just turns people away is that I think humans will default to what's comfortable. And let's say you were given this task to write this report and create like a little HTML. You already know how to do it. And so the idea of doing this, even though, you know, okay, this is probably going to take me 30 minutes, but you know exactly how to do it. So you're just going to do it. [SPEAKER_00] But then you also think about if I built this automation, there's a short-term cost that you have to bear. Whenever you're building things, typically you lose a little bit of productivity during the build. But the idea is that the long-term gains from it are exponential. You know, it's worth that short-term dip to experience that, the value that it actually creates in the long run. But I think that that dip is what causes a lot of people to procrastinate building the [SPEAKER_00] automation or procrastinate learning because they don't want to feel that, you know, the two hours it took that day to learn and to build when they could have just spent 30 minutes just doing it the way that they always do it. [SPEAKER_01] Yeah. It's like one of the first lessons that I learned in business and just building stuff is like, it's like sometimes you have to go one step backwards to actually go steps forward. And you're not actually, the ironic thing is you're not actually going backwards, but it feels like you're going backwards because you're not getting the thing right away. But like, it actually takes more time in the upfront setting this system up, learning how to use Claude code. [SPEAKER_01] But down the line, like, I think about what you mentioned right in the beginning of this conversation. Seven months ago, you weren't even using Claude code. And it's like, look at what you're doing now with it. Like some of the workflows that you've demonstrated. But you, to your point, you had to go one step backwards. You had to take the time to learn it and teach it and build the OS and all this stuff so that you could be where you are now. [SPEAKER_00] 100%. Yeah. And I just wanted to call out what's happening right now. You can see that I told it to expand the market size section. And if we go up, you can see it's significantly larger now. It's got more data in there. It went ahead and it took screenshots. You can see that it made sure that that new section rendered cleanly. But what's happening right now is it says one running task. And this is right here. It says, now the source verification. [SPEAKER_00] I'm spawning a separate agent to fact check every figure while I update the skill. And what's cool about that is if I click into this running task, we can see right here what's going on. We can see that this agent is verifying report facts and sources. And we can click into this. And this is essentially agent one on the left that we're talking to. Agent one sent off this prompt to agent two. This is called a subagent. So it said, hey, you are a fact checker for this report. [SPEAKER_00] Your job is to verify each claim, you know, confirmed, outdated, unsupported. And here are all the claims you need to verify. So this is a main agent that just created out of thin air a subagent. And now the subagent is working for this main agent. [SPEAKER_01] Yeah. Okay. You know what? Because I think this is so valuable. And even when we were talking before this, Nate, I was like, I've actually never heard someone describe this process in the way that you have. So congrats on that, first of all. [SPEAKER_00] Thank you. [SPEAKER_01] But, you know, you talk about the checks. And there's three things that I've heard in terms of how you're actually getting the AI to check its own work. So the first one was like the screenshotting method. And you showed us how you directly inputted that into the ClaudeMD file, right? So that was the first one. The second one was it's spinning up like subagents to check certain components. [SPEAKER_01] The third one that I actually don't think that you've mentioned, but I know that you said it to me was like browser use. Like you give it access to use browser use so that it can actually check its work. Can you just quickly for those three things kind of just share what's the use case for each one? Like, is there a specific use case where you use the screenshot method versus the subagents [SPEAKER_01] versus the browser use? Can you just explain that? And then also very high level, I guess, for the browser use, what that even is and yeah, [SPEAKER_00] what that even meant. So the screenshot is whenever I need something visual. So editing videos, I have it screenshot the frames, making sure that these HTML reports or PDF reports are formatted correctly. You know, don't have any elements that are off the screen or out of bounds. When I'm building websites, I make sure that the text is visible, right? Like there's if there's white text against a light background, it probably won't look good. But when the AI is coding that in hex code, it probably doesn't even think about that. [SPEAKER_00] But when it screenshots it, it realizes, oh, the text isn't visible. Let me just add like a little overlay behind. So that's whenever I need something visual. I have it fact check and verify anything that's going to be going out. You can even see here. It said that 18 of the 21 original claims were confirmed. One of them was outdated. Two had problems. So if we wouldn't have worked in that verification, it would have been putting out some false information, which is why it's so important that we're doing the whole bike method. [SPEAKER_00] And then on the browser use or computer use, that just literally means that Cloud Code will open up a Chrome or whatever you want to use and it will click. And that's a combination of vision and AI reasoning because it has to look at the screen and figure out where to click. And it can also type. So in the Calendly thing, the final phase was that it had a bunch of agents open up the actual app that we built and it tested the UI. You know, it tried to book an event. It tried to change the logo. [SPEAKER_00] It tried to do those things because it literally had 50 different browsers open doing it. And then they would all come back and say, hey, there was a bug when this happened. This, you know, this was laggy. And then my agent would fix it all and then send off those 50 agents again. So we just got that constant verification loop. And now you can see here, the output is the section was expanded, the source verification fixed things, and then the skill is updated. So if I actually open up the skill real quick, you can see that this is our market analysis report. [SPEAKER_00] It has the steps which are to scope, do research, write the report, brand the HTML, write in a calm voice. You verify the source. This is mandatory. And then you do a visual verification. So now I could give you this skill, which is just a markdown file. You could give this skill to your cloud and you could say, create me a market analysis report. And it would already have all of that verification and all of this information baked in because it would just read the skill and just execute. And then, of course, every time you run the skill, you can give feedback and say, update the skill. [SPEAKER_00] And it just will constantly do that. [SPEAKER_01] Yeah. Nate, I think it's so good. And one of the things that it really clicked when I heard you say this, you mentioned it in a video. It really clicked just the value of that verification process that you have is you mentioned that take away all the verification. On a first pass, clawed code would probably give you like 60% of what you want. I think what you said, you said on a first pass, you're maybe getting somewhere around 60% of the way there. [SPEAKER_01] And then after you give it feedback and these like you add on also these verification checks, that's what actually is going to get you to the 100%. And so when people use clawed code or like one of these programs like Codex even for the first time and they get the poor or the mediocre output or the AI slop as people call it, right? It's because you're looking at the 60%. Like it didn't, you haven't elevated it to that point. [SPEAKER_01] So yeah, it's just such a, this is like, it's mind blowing in a lot of ways, like what it's even capable of right now. [SPEAKER_00] 100%, yeah. And I think it's also all about aligning expectations, you know? If you can be more specific about what does good look like, then it can actually work harder to get you to that good. But, you know, something like, hey, create me a market analysis is very vague. It's very open-ended. So it's going to get creative. It's going to give you what it thinks you want. But if you tell it exactly what you want and what you consider good and what's important to you, what your motivation is, then it will give you something on the first pass that's a lot better. [SPEAKER_01] Yeah. You know what, Nate? Something that I said earlier in this conversation, and I truly mean it. And if anything, as we've spoken more and more, I've just believed it more and more the case, believed that it's more and more the case. Which is, there's going to be someone that listens to this conversation. And it's not just that they implement some of the stuff that you've shown. They play around with Claude Code. It's going to make them way more productive in their life. [SPEAKER_01] And then there's going to be some people that take that even further. And some of the opportunities and some of the things that they learn how to do with Claude Code actually allows them to start a business, which creates like an additional income for them. Like, I think that's going to be the case. Just to warn you, in the next three months, six months, 12 months of people that watch this conversation that we've just had. But Nate, one of the things that I'm so cognizant of is until you start taking like action and you [SPEAKER_01] start that first step and you start building momentum, it's like that's never going to happen. Like that potential reality will never be here. And so my question to you, someone's watched this conversation up until this point and they've been inspired and they've thought of all these use cases and things that they want to do. And let's just say that going into the weekend, maybe they have like a free Saturday afternoon or maybe it's like the Sunday evening right before they get to work. [SPEAKER_01] They just have some time, like a few hours to spend with this. What are you telling them that they should immediately go and do in that time, that right first step so that three months, six months, 12 months from now, they're in that point, that new reality that I described earlier. What are you telling them to go and do on their free Saturday afternoon or Sunday evening? [SPEAKER_00] I think the first thing you need to do is when you open up this tool, do what's called a grill me session. So there's actually a skill which I can give to you, Callum, if you want to like link it for your audience, but it's called a grill me skill. And so essentially what that does is it relentlessly interviews you until it understands everything about the topic you want it to understand. [SPEAKER_00] So what I would do is I would open up a fresh claw. There's probably nothing in there. And I would put in the grill me skill and say, I want you to essentially be like my personal life coach. I want you to know everything about me, my background, what my motivations are, what's important to me and where I want to go so that you can just help me get there and you can help me be more productive. And this thing will, it will be relentless. [SPEAKER_00] It might ask you 50 questions and just take that time, take a couple hours and just brain dump everything about you and about your motivations and your goals. And I think once that interview is done, your system is going to feel, it might even feel like you're talking to someone that you've known for your whole life already, because it has all that context and it's going to save all that context. And then every time in the future, you open up a chat, you want to build something, you want to see if an idea is good. [SPEAKER_00] It's going to have all that context and it's going to help you just move a lot faster. Because I think that's the most frustrating thing is when you have to repeat yourself or when you have to drag things over and you feel like your systems aren't syncing together, you know? But now I have my AI operating system that I use every day and I can use it in Cloud Code, Codex, Hermes Agent, OpenClaw, GrockBot, any other new AI tool that I ever have ever or ever use. I just plug it into this directory and it knows everything about me. [SPEAKER_00] And that's all of my IP. That's like the most valuable thing I have right now is all those files and folders. [SPEAKER_01] Yeah. So first of all, on the grill me skill point, it'd be great to link that. And we will link that in the description. But secondly, and even more importantly, Nate, just thank you, man. Like, I just think that it's funny. I go into every conversation that we have on the show with like, I just want to over deliver, especially for like that non-technical person, that person that kind of like you and I starting out of like, we're not the coder. We don't work on Silicon Valley. [SPEAKER_01] We're not a software engineer, but like we're just hearing about AI all the time. And it's like, okay, what could that mean for me? And so I just always want to over deliver for that human being. And we did that in this conversation and you were just like such a huge part of it, obviously. So thank you, Nate. I appreciate it. [SPEAKER_00] Awesome. Yeah. Well, I really appreciate that. That means a lot to me. I wanted to come in here and make sure that we could provide everyone with a bunch of value. So I really hope we did. And yeah, I had a lot of fun on this session. So thank you so much for having me. [SPEAKER_01] So if you enjoyed this conversation and you want to hear even more stories like this, then just click here. And also my team is going to put some more videos that you can watch here. Thank you.