#146 How I Built An AI Marketing Team With Claude Code (full tutorial)

#146 How I Built An AI Marketing Team With Claude Code (full tutorial) Speaker 2: What if you could build AI-powered marketing systems without writing a single line of code? Well, that's exactly what Claude Code allows you to do. So in this episode, we're going to sit down with Craig Hewitt, who's going to walk us through his entire Claude Code for marketing setup live. He is the founder of a seven-figure SaaS, and he's grown his YouTube from 250 subscribers to over 13,000 by educating non-technical business owners on how to use tools like Claude Code We're marketing and he said something that stuck with me in this episode. He built the first iteration of some of these SAS tools that you're going to see sitting outside of his son's basketball practice on his phone. He's not technical and you don't have to be either. So let's dive right in and learn how to use Claude Code for marketing. All right, Craig Hewitt, what are we going to learn today? Speaker 1: We're going to learn about Claude Code for marketing. I think Claude Code is probably the most poorly understood AI product out there these days, outside of developers, and it is a powerhouse for marketing. So I'm going to walk through some of the things that I use internally, we've open sourced that anybody can use and kind of both specific things and kind of frameworks and abstractions that you can use in your business and in your products. Speaker 2: Yeah, and Claude Code is crazy powerful. And one thing that we're going to do is we're going to kind of bring it down to the ground level for the non-technical business owner. I know myself included, anytime I heard the word code, my brain just shuts off. But I promise you guys, this is going to be worth your time, especially if you're non-technical like me. So before we jump in, Craig's going to share his screen and kind of dive into some of his workflows in action. But before we do that, Craig, do you mind just giving us the rundown on why people should listen to you? Like, what's your background? Speaker 1: Yeah, so I'm the founder of Castos. We're a podcast hosting and analytics platform. We're a seven-figure SaaS business and been doing this for eight years. And in that time, like, kind of gone from conventional SaaS to like very AI forward as an owner and operator and marketer. My background is all in kind of sales and marketing. And middle of last year in the fall, kind of ramped back up my YouTube channel. All about AI now and did 100 videos in 100 days. Grew my channel from 250 subscribers to 13,000 as of today and had some really big wins like 100,000 plus view. Videos about AI. So I live and breathe AI. I advise and coach folks on how to use AI in their business. And so both like a practitioner doing it every day at Castos and also helping folks in situations like this to teach them how to use AI, specifically Claude Code, often just because in my opinion, it's the best AI platform and tool on the market these days. Speaker 2: And a hundred days of AI, I love that concept. That's actually how you came on my radar as I stumbled across that playlist, but no, that's so cool. And I think there's a lot you can learn from that, both as far as like learning about all the AI tools. If I'm going to have to publish a video about it every day, I'm certainly going to be in the weeds quite a bit, but also learning about YouTube and, you know, how to run a YouTube channel. So I think that was a great experiment. Speaker 1: Cool. Thank you. Yeah. Speaker 2: Okay. Speaker 1: For me to go ahead and share my screen then? Speaker 2: Yeah, let's dive in. Speaker 1: Yeah, cool. So I think we'll be a little general at the beginning to kind of orient folks on kind of tooling here. And Corey, just like most folks here, you think familiar with what Claude Code is or do we need to really like start from basics? Speaker 2: Yeah, we can kind of start from the basics as far as like, what is it? What does it do? How is it different from like your normal Claude? And then we can kind of get into the weeds from there. Speaker 1: Cool, cool. So at a high level, normal Claude, just like ChatGPT in a lot of ways, or Gemini, Grok is just a chat interface. And like, it's like mental ping pong to me, like I have an idea, it says a thing, I come back and we do a thing and we might like hone in on an idea like that, that's fine. And it's kind of dumb, right? As amazing as ChatGPT and Claude are, they're kind of dumb and that they can't do things for us. But within Claude Code, like first of all, Claude Code, you interact with it in what's called the terminal. And the terminal is that super scary place that developers go to write code and do scripts and a bunch of stuff that like, I don't understand, like I'm not a developer. So I think like, first of all, Yeah, so like don't don't be scared of this place. I'm using an application here called Ghosty, Ghosty with two T's. It's a terminal emulator. So it's kind of like it just makes the terminal kind of look and feel a little more human than it would otherwise. And in here I'm in this project and a project is just a place on my computer. So I have this folder called coding and I'm in this project called Castos Writer. And so this is the tool we use to create All of our content on our website, so SEO, landing pages, comparison pages. We use it to do like conversion rate optimization. And there's a lot of ways that you can use Claude Code. I'm going to pull up. I'm going to pull up a different program to show you kind of the structure because I think at the conceptual level, it's easier to see here before we get in and like actually. Actually start like using the tool. Speaker 2: Yeah, and I think that's a good preface too, because again, it's like you hear terminal and your brain shuts off, or at least mine. And then when you finally muster up the courage to get in there, you're like, well, how do I even arrange my folders? Like, how do I set things up? How do I kind of set the foundation so that I can start using what I think is quite honestly the most powerful software tool on the planet right now? So I think this is a great segue. Speaker 1: Yeah, so and in terms of like tooling, there are ways to do this all in one place. And that is in what's called like an IDE, an integrated development environment. And so folks might have heard of like VS code, or cursor or anti gravity from Google. Those are all cool, you can use those. And actually, you know what, for for today, I'm going to do that. So we can just see everything in one place. So let me just pull up And all of these kind of are and can be free. Antigravity, definitely free. VS Code owned by Microsoft, free. And these are just like applications that run on your computer where you can access all this stuff. So let me just open up this project. Speaker 2: And I'm using cursor as well and I've been using it for a few months completely free and it's a great tool. So yeah, very similar. Speaker 1: Yep. Yep. Um, okay. So this will be the last, the nice thing about this is the last time I'll stop sharing my screen and switching. Um, okay. So the nice thing about an IDE, uh, and so you can see the files on the left here and then the terminal down here. Is that right, Corey? You can see what I'm looking at. Yep, cool. So this kind of looks like a combination of the two screens I showed before, right? The terminal down here, this is the place where you're going to talk to the product. And then the file structure over here, kind of like, hey, what's in this folder? If you were imagining looking at Windows Explorer or Finder on a Mac, like what, you know, these are folders, inside folders, you have files and stuff like that. So just to orient, You have a couple of folders over here. By far, the most important one is this context folder. And you can see, like, I'll just pick up a couple of these, right? So brand. Who am I? Who is Cast Us? What is this all about? What kind of product do I have? Who do I serve? What do they want? What kind of things they want and need? What are they scared of? Who am I serving? What kind of tone do I use in my language when I write? Right. So so this is like who I am and how I want to represent myself. And we have a bunch of these. These are the most important documents in this whole project, because I use the term intentionality a lot in AI, because like when you get Subpar performance from an AI. 95% of the time, it's because you didn't give it the context that it needs to be successful. Especially now, like the models are so good. Models being like within ChatGPT, you're using 5.2 medium or whatever. In Claude, you're using Opus 4.5. That's the model. The models are so good. You don't have to like prompt engineer anything anymore. You just say like, hey, write me an email. If it doesn't know who you are or who your business is or what your customers want or what features you have, it can't write you the email you have in your head. This is the way that you seed this project with all of the things that it needs. So you all can see here like features, competitors, other examples of stuff I've written. We have a thing called a creator spotlight at Castus where we highlight one of our customers like this is what goes into this. Landing page templates, right? So this is where if you're like going from zero to Claude Code expert, This is where you're going to spend 70% of your time. And this is where you're going to spend 70% of your time because like I'll share in a little bit, like we've open sourced a version, like a skeleton of this project for folks to just go download and start using today. And you can, all of the other stuff that I'll talk about is a hundred percent in there. It's for free. But this, like, this is you, like you gotta, you gotta come in and like tell the machine who you are and who your, who your company is and who your brand is and what your product is and all that kind of stuff. So Corey, any questions or anything you want to touch on here with context seeding? Speaker 2: No, I mean, I think you hit the nail on the head. That was pretty much my exact LinkedIn post today where I think the opening line was like, I downloaded an AI agent expecting it to work magic. And what ended up happening is it was just another dumb chat bot. And my purpose of that post is like, well, it's because I didn't seed it with the proper context. So, I think you're spot on. It's like most of your time, most of your effort, it's like that Abraham Lincoln quote where it's like, he's got six hours to chop a tree, he'll spend the first four hours sharpening the ax. And the sharpening of the ax in a Claude Code world is giving it all the context that it needs to do its job. Because once it has the proper context, It's a Ferrari, like it's a racehorse. It'll just go to work and you just let it do its thing. But that's like 80% of the battle is giving it the proper context for the task. Yep. Speaker 1: Yep, absolutely. And so in the kind of sample project that we'll share, y'all can check it out. It's seomachine.io. It's like entirely free. You can check it out and we'll include a link in the description. You basically download a skeleton version of this app, but you will need to fill this out. It's basically all you need to do. And so go spend a few hours doing this. If you're using something like ChatGPT already that has infinite memory, it might actually do a good job. Be like, hey, I need to create these 10 documents. You could download the examples and load it in there and say, like, hey, make this for me. That's kind of my favorite thing to do with AI. It's like, hey, take this other thing over here that's an example and make my version of it. You can do that. So load the context, right? That's like by far the most important. From there, if you think about the analogy I love with this kind of stuff is like, okay, I, I just hired a really smart marketer that kind of doesn't know what to do yet, right? And so we've solved the problem of it knowing my business now because we have context. Now we need to give it SOPs basically. We need to give it standard operating procedures. And the term we use internally is like the definition of done or what good looks like. And that has to do with skills and commands. And so within this Claude folder here, you have A couple of things going on and I don't want to gloss over any of these because they're all really important, but they're all pretty different. So the first one is agents. And so y'all heard like the term like AI agents and you're like, what the heck does that like actually mean? I think it means a lot of different things to a lot of people. But to me and on my channel, the definition of an AI agent has, it has three things. It has access to a model. It has access to one or more tools and it has memory. If you think about this content analyzer agent, it has access to these scripts. I'm going to go look at the search intent of this term using this Python script. It has a lot of knowledge because it's going to reference all sorts of other context and information in this project. And then it's going to call a model. It's going to call OpenAI or it's going to call Anthropic, right? So these agents can run anytime. You can invoke them manually. You can call them down here in the terminal. You could call them. You can also set up in the definition of the agent for it to run on specific times, like when I commit this code to GitHub or when I run this other command, like down here in commands and skills, and we'll talk about those in a minute, or the machine, Claude Code, can just decide when it wants to run these on their own. So agents are These are pretty autonomous. They can be invoked. They can just run whenever. I like agents as a secondary thing in my productivity kind of suite, my tooling. And I want to pause here to just underscore something. I hope we're starting to see how much more powerful this is than just ChatGPT. In ChatGPT, you could do this. You could load up a GPT or a Claude project or a Gemini gem with all of this stuff, and you could even have a really fancy system prompt in there to do a single thing. Speaker 2: Right. Speaker 1: And then that's it. And then you're like, I got to go create another one when I want to create email. And then I got to go create another one when I want to analyze my AdWords. And I got to go create another one when I do conversion rate optimization. And that's where things get interesting is within the .claude folder, by the way, for like the non-developer folks. When you see a dot in front of a folder like this, it just means that like it only ever lives on your computer. This stuff doesn't get sent to GitHub. And GitHub, Git and GitHub are what's called like version control. And so like this project I share with our team so that when we go down here and like draft articles, Or we run research. I share all of these with my team so we can push them to a website. I can pull them down kind of like you would with a Google Doc. You can share and collaborate on these. But this folder here. This .claude folder does not get shared. It's only yours. And so if you're like, gosh, that kind of sucks. I don't want to do that. Well, cool. Like figure out a way to include this in your Git or do something else, have a central repository for all these skills. Because if you're a 10 person marketing team, you might really want everyone to be playing on the same playbook. And so figuring out a way to share these in a standardized way is pretty important. But but so skills and commands, truth be told, are pretty similar. So commands. You invoke manually and I'll show you what I mean here. So down here on the terminal, the way to start Claude Code is just say Claude. And if you don't have Claude Code installed, like just look it up and there's like literally a one line thing. And from here, now you're still in the terminal, but you see like Claude Code is running here and you can invoke a command here just by doing slash and then optimize. And so I'll just do a slash research. Um, how to start a podcast. Uh, I'll just do, uh, let's fast USB microphone. Unknown Speaker: Okay. Speaker 1: And so what just happened is it's reading the context here and then it's doing whatever I asked it to do here. So it's doing a bunch of... Speaker 2: Like I'm assuming, because right when you hit enter, right, it said read brand voice, read context, read like two other things, read target keywords, read SEO guidelines. So I assume that it reads those first every time because that was probably part of, like the research command probably tells Claude Code that, hey, before you go out and do your research, Read these four reference files first. Is that correct? Speaker 1: That's correct. Yep. Speaker 2: Okay. Yeah. Speaker 1: Yep. And I don't, just off the top of my head, I don't see where that is in this document, but yeah. Yep. So right, exactly. It's going to get the context of things it needs to do these things. It is running three different agents in parallel now. So it's doing keyword research, it's doing competitive research, and it's doing USB mic market research. And it's using web research for all, like it's going out to the web to research all of these. And it's asking me, like, do I want to access this website? And I'll just say, yep. I'll click down here and just say, yep. And I just wanted to access all these websites. And so it's going out and it's getting a ton of information here. And that's going to come back and give me an outline of all this. Again, like how much more powerful is this than just asking ChatGPT to do a thing? Like there are other tools that can do a similar amount of work. The only other one that I know of is called Manus, M-A-N-U-S, manus.im. Manus can do this level of work. It's all just in the browser. And in a lot of ways, I like Manus more for specific things. And by that, I mean like if I'm just starting something new, I might just go into Manus and kind of just like one shot it. Like if I'm not, I'm not going to set up all this context. I'm not going to like do all this stuff. I'm on my phone maybe cloud code doesn't work really great on your phone it works on your computer which is why it's so powerful because it has access to the whole kind of breadth of your computer. You know man this is just an app like chat GPT or whatever so. But it's a it's an agent to get a tool so it will go out and do all this it can spawn up to a hundred sub agents to go do all this research. And it can create scripts and write documents and in a lot of ways is really similar to Claude Code. Manus was acquired by Meta, like Facebook, a couple of months ago. So it's gone out and it's done all of this research here, right? All of this research, like, whoa, that's really great. Like, what do I do now? Like, this is all in the context of content marketing. So you're like, oh, cool. What do you do with this now, right? And it's done this and it's saved this And this file here, or it wants to ask, do I want to save this in this file? And I'm like, yeah, cool. Let's save this. And so if we come back over here to our little file tree, we can find research. It's creating the document now. And in just a minute, it'll pop over here. It's probably here. Speaker 2: Yep. Speaker 1: So it's here. This is the document it created. So cool. You could just stop here and be like, I'm going to hand this off to like the marketing person on my team or the writer. That's fine. Or you can invoke another command and say like write. This file. When you click up here, you're now referencing this file in this command. If you're active in this file up here, then it knows that when you say write, you want to write about this thing. Once this is done stewing as it says, you can just write based on this. While this is working, I want to pause and just say like, I'm not technical. I don't really know. I'm not an amazing marketer. I'm not an amazing AI person. Truth be told, I did the first version of this. Sitting outside my son's basketball practice, like I wasn't allowed into basketball practice. I was sitting out there like 45 minutes and I was like, this is really what I want. And I was just talking to Claude Code about how we could build this thing. So like, don't think that like, oh my gosh, like Craig, some kind of like freaking AI wizard and I couldn't ever possibly do this. Like literally you could just, and we'll do this in a minute. We'll like spin up a whole new project. You just talk to Claude Code about what you want. And it'll make it. And especially if you have like an example project like our SEO machine, like it is literally all of this stuff. We just updated it this week. I gave a big presentation yesterday on this topic. So we tuned everything up and published it. It's entirely free for you to go check out. If you're like, hey, this just isn't right for us, but like, this is a cool kind of starting point. You'd be like, hey, I want to make a project kind of like this, but not exactly because I'm a D to C consumer brand. And really what I want to optimize for is ads. But I really like the concept of like, I want to research, I want to write, I want to create ad scripts, I want to create maybe like video, you know, storyboards or something, be like, I want to do that, but in the kind of robustness and style of the stuff that we have here. Speaker 2: Right. Speaker 1: Cool. So Claude Code is really smart. And it's like, cool, I created this brief, next step is run best We're going to talk about the best USB mic. And so let's do that. So write best USB microphone. And it's going to write the article based on and with all this context, based on the research it just did, which is exactly how every good writer in the world would do it. Three years, we've done a lot of content marketing at Castus. We would pay thousands of dollars a month for writers. And what they would do is they would have like five different tabs open in Chrome and they would read all the articles on the topic. They'd create an outline. They'd pick and choose the best parts of each one and they'd smush it together into an article. It's essentially what we've done in like two clicks of a button. And then you're going to get like a 95% optimized article here, basically for free. What it costs like 10 cents probably in tokens or you're on a subscription. It doesn't even cost that like Claude Code cost. Like a paid Claude plan could cost 20 bucks a month. Speaker 2: Yep. And I'm on the, I saw you were on max, like I'm on pro. So I pay a hundred a month, you pay 200 a month, but like the everyday person who's just getting started with this thing. I mean, you can be very dangerous for the $20 a month plan, right? Yeah. Yeah. Speaker 1: Yep. Yep. And so this'll take, this'll take a while, right? And this always makes me feel good that it like, doesn't just spit something out instantly. Like, like some of their AI tools, like, it's, it's, yeah, it's doing its work. It's doing all this kind of stuff. I'm going to, I'm going to pause here and just flip over to share one, share something in a different screen, because I think it's, it could be easier to like, to kind of get what I'm talking about. Unknown Speaker: So to do. Speaker 1: Okay, so talking about skills. Skills are things that you can invoke and you say slash research, slash write, slash analyze. A friend of mine, a guy named Corey Haynes, really, really, really good marketer, has put together a list of marketing skills. Again, it's entirely free. It's in this, we talked about GitHub, this place, and it has all these skills. You can go in and you can check them out. For free. So he has, I want to set up a split test. I want to check my analytics. I want to look at like competitives and alternative pages. I want to do like a conversion rate optimization and you can go in and you can see these. And you can open them and see kind of how Corey thinks about doing a job. And a skill is just a job that you want the AI to do. So in this case, the job is I'm going to do conversion rate optimization on a page, a homepage, landing page or whatever. And you have all of these skills. Y'all can be like, I want all of them. They're free. It's nothing to just install all of them or I don't really want to do any of this, but I want to see how a really good marketer and a pretty savvy AI person thinks about it. You just go look at these skills and see like, hey, if I want to create my own version of this, how is this as an example set up? How is it organized? How does it define what good is? How does it give a parameters, things to do, things not to do? All of these marketing skills are bundled inside our SEO machine, you know, to kind of sample project. And so you get these all ready to go. And so you can do all of these kind of activities within your Claude Code project. Speaker 2: And just to kind of touch on skills a little deeper, so I agree. I mean, I think skills are some of the most valuable weapons you can arm your AI with, for lack of a better example, right? And what I tell people, I'm like, think of a skill as an SOP. It's just a process that you do in your business. And so by creating a skill file for your AI, when you ask it to go do that process, it's just going to do it the same way every time now, based on whatever you laid out in that skill. And what's really cool about Claude and Claude Code is that it actually has a skill builder skill built in. So you could go into Claude. You don't even need to be in Claude Code. You could just go into the regular Claude app and say, hey, I need you to help me create a skill for how I draft emails. And it'll kind of interview you or walk you through the context that it needs in order to generate And we're here to help you build a solid skill and then it'll create it and it'll even install it to itself. So now every time you ask it to write an email, it's gonna do it based on the guidelines that you laid out in the skill building process. Speaker 1: So skills are very important. Skills are important. And I think that you think of skills as like, yeah, an SOP is a really good one. And if we think about what we did to do the research action, they're both an SOP and they're a command. You can invoke them manually with the slash, slash research or slash write. They also could be called automatically. So you could say, hey, go do CRO on castos.com. It's obviously very smart. Yep, it will know to use that skill. And so the last thing I would say, and then we'll just create a new project from scratch here in just a minute, is SEO Machine, go check it out. We'll include this link or just Google for SEO Machine. It should pop up. And then just grab the link from this page. And just tell Claude Code that you want to use this and it's super smart. It's like you would say like clone this or it'll know what to do. It's like, hey, I want to use this as a marketing project on my machine. Make this happen. On my computer and look like we just updated all this stuff like yesterday or the day before yesterday. So an SEO machine is pretty cool. So like back to context, as opposed to all this stuff existing in a bubble, what we do is we get data via like custom integrations we created from three places. And I'm trying to figure out where we write them here. But the three places are your Google Analytics, and you'd have to connect to this, your Google Search Console and a third-party data provider that's kind of like Ahrefs, if folks are familiar with it, called Data for SEO. Here it is, Data Sources. And so we're getting actual traffic metrics from your website, getting actual clicks and impressions from Google Search Console, and we're getting actual comparisons of How you're doing versus everyone else in data for SEO. So this is not just like, hey, I want to go write an article about something. It's like, hey, what should I write about? Like we have a skill for prioritize. And it does all of this analysis that again, like a really senior content marketer would do. you, like basically for free. And we look at like search intent and keyword analyzer. So like super, super robust kind of platform here. Download it. Check it out. Plug the data sources in. You'll have a lot of fun, if nothing else. It's pretty cool. Speaker 2: And of course, we'll link that in the description too. And again, it's free. So anybody that wants to grab that, like Craig said, something I like to do is just when I find a skill pack that I want to implement, I just copy the GitHub link, literally paste it into Claude and say, hey, install these. And it'll just go to the link and pull the files and install them in 30 seconds. Speaker 1: Yeah. Speaker 2: Yeah. Speaker 1: Yeah. Cool. So, um, I thought that we could like create a new project, uh, from here because I think folks might be like, gosh, this seems like, like a lot and I totally get it. Um, so we're back in our IDE, anti-gravity. And so it could be cursor, it could be VS code. It could just be the terminal on your computer, which is how I use it a lot of times, but I want to, um, Okay, so we'll work from here. I just want to figure the best way. So this is all done, right? It has like a comparison table here. It has the full blog post. This was running while we were talking. So it has this full blog post. It's like 3,000 words. It's optimized based on the research we did. So this is cool. So I don't want to create anything here. So what I'm going to do is quit out of here right now. And I can see I'm in this folder on my computer. So I'm going to do a little bit of like the little bit of terminal stuff I know, which is like CD change directory. And you do space dot dot means go back one folder. So like in the hierarchy, like go up one folder. And this is where like all my projects live. And so like I have all these projects on my machine. So I'm going to make a new one. So MKDIR, make directory. Corey, what do we want to make a Claude Code project about? Speaker 2: Let's make one on, so, I mean, we're on the SEO topic. Let's make one for, uh, or would that be kind of duplicating the one you already have? Would you rather do, let's actually do a project maybe for coming up with YouTube, uh, topic ideas. Yeah. Speaker 1: Cool. Speaker 2: Let's say what's trending in the AI space. Speaker 1: Yep. Um, let's be real original and call it YouTube machine. Okay, we'll call it YouTube Magic. Okay, cool. So we made a directory, made a folder on our computer called YouTube Magic. And then let's say CD, change directory again into that YouTube Magic folder. Magic, cool. And then we just say Claude to start Claude here. We see Claude is running here and what I'm going to do now is make sure my Super Whisper is working. Cool. I use a tool called Super Whisper on my computer. Again, in terms of tools, Claude Code and Manus are amazing. By far the most powerful AI tool is called Super Whisper or Whisper Flow or there's a couple out there. I put a touch a key on my computer. It starts recording. I speak. It transcribes and it pasted in wherever my cursor is. And so we'll say I'll do mine. For me, it's option space. I want to create a Claude Code project that will search YouTube and find viral video ideas that I can implement into my Channel. So I need an area for context for me to tell you about my channel. I need a way for you to create, uh, to gather data on existing YouTube channels and things that videos are doing there. So you'll need to go find winning videos. We probably should have a concept of like a winning formula because we'll probably want to repeat several different formula types throughout time. And there's probably some other stuff that we'll need. But the goal here is I want to use this YouTube magic project to find winning video ideas so I can grow my channel. Speaker 2: Now, Craig, let me ask you something, too. So the project that we're looking at here currently on the left-hand side of the screen, that's your writing project, right? So that is for anything that you're doing on the SEO side or content writing. But this new project that we're creating is specific to YouTube. So one of the things I was going to ask is, so like this YouTube project, I know you already mentioned it, but obviously we're going to have to give it some context in order to give us the best details. So what is some of the context you would give Claude Code for the YouTube project. Would it be like, hey, here's my channel. Here's the videos I've made thus far. Here's what my average viewer looks like. Here's their demographics. That type of thing, just more specific context that it would only need for that project? Yeah. Speaker 1: So I think in terms of how we're using this right now, That's a bit why I don't use an IDE because if y'all saw when I was sharing Ghosty before the terminal, I got six different projects going. At a time because I have like I have one that I call like my second brain. Like it's just it's everything right. I have one that's like Castos co-founder. It's my co-founder. That's just like gut checks me on what I'm doing. I have the writer one. I have like the actual product and products that we're building. And so the IDE to me is not a great way to multitask like that. I know that that's kind of what you're asking, but like there are Man, I hate to say it because we're talking about Claude Code, but Codex actually does this really well now. Speaker 2: Really? Speaker 1: Yeah, like I don't, I tweeted this out last night, like I didn't want to like Codex, but I used it yesterday and it was like, it just came out yesterday. The new version of it, it really does this well where you can have, let me just, let me just share that real quick because I think it's just like, it's appropriate to To have folks see like the the kinds of ways you can work. That's almost as important as like what we're what we're talking about Okay, so so this is codex This codex is a chat GPT product and I This is just, it's a different style, but it kind of gets at the same thing. And so this Polyboard is a project I have that is, it's like a Trello board for AI agents. So like I have OpenClaw on my machine and I was like, I need a way to see what all my eight agents are doing. And Polyboard is like just this open source thing I want to do. So like they can report and I can give them feedback and stuff like that. So I have, in this case, two projects going because I just started using this, but you could add a new project. But in here, like you can just say, like, go do a thing here and then you come over here and say, go do another thing. You can have five or six projects running and all of them running at one time. So like compare that to how you would use a standard IDE. Like, it doesn't matter, right? Like cursor, VS code, anti-gravity, all are kind of like not meant to do this, right? They're meant for you to work in one place so you can have the files up here, the terminal down here or multiple terminal windows, you could do that, right? And the file structure over here. Well, like truth be told, I don't use this and I don't use this hardly ever. For marketing, I use I visualize the documents a lot more than anything else. But even then, I just don't use it that much. So that's why most of the time from a tooling perspective, I just use, especially if it's like my thought partner, I just want a place to chat that has all of the context. So I'll use Terminal. If I'm developing products, increasingly, I'll use something like Codex or for Claude Code, there's a product called Conductor. Conductor.Build, exactly the same thing, uses what's called get work trees. So it creates a separate instance of the application, so you could really multitask on a project. And the primary way you use it is just chatting. And I think that's the thing. So because like what you're like, what you're getting here is like really true, which is like we're about to create this or we've created this project. But then like we're stuck in this folder that's the wrong folder. So anyways, we're going to hit go here. And it's like, okay, cool. It has kind of the gist of what we want here. Actually, I'm going to hit stop. Oh, it entered plan mode by itself. Cool. So that's smart. So if it hadn't entered plan mode, I would hit shift tab. And you can hit shift tab for it to accept edits when just like it's just going to go do whatever it wants, which I usually do. And then plan mode, when you have a prompt like this and you're like, hey, I got this really big thing I'm going to ask you to do. I want to enter plan mode so we can kind of like square on what I'm talking about and that you understand it and that we're on the same page. And so like if you've canceled your like progress here, I had just hit escape. You would just hit up and it pulls the previous prompt back and you're in plan mode and I would just hit enter to kind of restart this even though it was entering a plan mode just for demonstration purposes. And so entering plan mode, it's like, cool, let me get a lay of the land here. Let me figure out what's going on. It's going to read a bunch of files. From there, it might go out into the internet and do some more research. If you're referencing another project on your machine, like Okay, cool. So it's asking me already, like, how do you want to gather this data? Because this is what we do at Outlier. I know this is what you want to do. So we built this into a tool. It's called Outlier, outlier.so. So use the YouTube API. It's free. What makes a winning video to you? Cool. Let's see, view count. And here I'll just hit spacebar to check this. All these things. Recency, no, it can be a little older. So I'm picking three of the four here. And if you wanted, you could put something else down here. And I'll just, you can type in your own answer. It's like, hey, are you sure this is what you want to tell me? And you say yes. And from here, plan mode often will like go out into the internet. It'll ask you if it can access other projects on your machine. The really common one is like we have the Castos and the Outlier application that we developed from like a code perspective, and we have marketing and kind of ops projects for each of them. And I want them to talk to each other to understand. They're not in the same place from like a folder perspective, but they can and should reference each other to understand back to context, like what's going on. Cool. And so it's done a lot of thinking and it has this kind of like project structure, right? So it has a folder for context. It has formulas. So this is where like winning formulas that we want to be able to repeat are. It's going to do research. It's going to have this script to go get the data from the YouTube API. And it'll have a place to put environment variables, which are just like Kind of like passwords for computer systems to talk to each other. So it has the plan and like the plan is good and we're going to say like, cool, just go make it. And so Claude Code about two weeks ago introduced this thing where like it clears the context and auto accepts like the implementation of the plan. Context being an important thing here and context management is always an important thing in a powerful AI tool like this because this is a pretty simple thing. It'll go work for a couple of minutes, but it could fill up the context window. And if that happens, it has to like compact the context. And context window just means like the number of tokens that it can send from your computer to the model, to the AI. And over time, like that just grows. And so at some point it's gotta be like, okay, I gotta like trim this back down, give me the CliffsNotes version of it, and then I can kind of start again. But clearing the context at the end, and like if you had gone back and forth We've been working a lot with it up here to get to a point. Your context window could be like half full already. And that would be a bummer. So this is pretty smart because like, cool, we have landed on the plan and we're just going to go make the plan now. And so I'll just say, yeah, and it'll go re-familiarize itself with everything because it gave itself all the context and the plan up here. And so it wants to start doing things on my computer here. So it's going to go create these subfolders. And I'm just going to give it accept edits and it'll just start doing stuff. And so, you know, this is like a pretty quick and dirty example, but you really don't need to give it a lot more instruction than that for it to start creating the structure and the file system and the functionality that like what we have here in the Castus Writer tool is, I mean, Really sophisticated, but you get like 70% of the way there in like 10 minutes. Speaker 2: Right. So it's creating an entirely separate folders. We would have to technically go up and open that new folder to start working in this new YouTube magic project. Is that correct? Speaker 1: Yeah. And to be honest, like I don't know how to do that. I don't know how to do that without like messing up what we're doing here. But like what I would, yeah, what you could do is like create this. It's done. Close all of this out and then open up. This project, the YouTube magic project, you can see down here, like this is the folder. So that's what I would do is like, well, what I would ideally do is like start in the folder where you want this thing to live eventually. Speaker 2: But create everything inside this folder. Speaker 1: Yeah, but but that's a pretty cool. It's a pretty cool thing that like Claude Code can reach vertically and horizontally from where you started. So we started it in this Castos Writer folder and now it's working in this other folder because we told it to. So yeah, and like, oh, I guess we changed this, but even if we had started in one folder, you can tell it, hey, go look and work somewhere else. So some of my friends just have one Claude Code, kind of one place from which they run Claude Code. It's like a, I'm a software developer. I'm also a parent, if you will. And then they have like the repository for the application, like the software. They have like the marketing assistant. They have the salesperson. They have the ops. Project and that kind of parents has context to everything. And so you could do that. Speaker 2: I can see that being maybe even like a cleaner way to keep everything together. Uh, especially if you're just, you know, if you're going to have like a few projects, like if you're not going to be a power user, like you, somebody who's just like, Hey, I want to have a couple of little, uh, specialized quad code projects for each of these functions in my business. I don't plan to do much more outside of that. Then that can be a really good structure for that. I feel like. Speaker 1: Yeah, I think one thing to be kind of a little bit aware of there is there is definitely a concept of like context bloat. So kind of the other side of it, like we talked about, like, if I just go in and say, like, hey, write me an email for Corey's business. I'm like, who's Corey? What's his business? Who's the customer? Doesn't know any of that stuff. If you just throw the whole kitchen sink at it, Especially a bunch of different types of context, code and marketing and sales and financials, it might not do great at specific things. I don't know the answer. I've not done what we're talking about there of having a parent project, but, I would just be aware if you really start doing a lot in there, you might have to drill down the level to do specific things. Speaker 2: Yeah, and that's a good point too. And so this is a question that I've selfishly been looking for an answer for. So let's say, you know, I want to have different projects for different platforms, right? So if I'm posting on YouTube and then X and LinkedIn and all these places, I assume your advice would be to have like one project per platform, right? As opposed to just like a general writing project that might encompass say X and LinkedIn, probably more effective to have just an X project that has just context around X. Same for LinkedIn, same for YouTube, same for each individual platform. I mean, am I looking at that correctly? Speaker 1: I have this project and so I want to open it to just share how I think about it. Speaker 2: Yeah, I'd love to see that. Speaker 1: Yeah, so I just started on this and this is like a lot of this stuff is We've been doing this for days and weeks old. So like, I think this is all moving so fast. One of the biggest challenges we all kind of have is like how to stay up to date with this stuff without going crazy. Speaker 2: Yeah, that's my biggest like flaw. Speaker 1: Yeah. Speaker 2: So behind the eight ball every day and then I'm like, wait, 99% of people have never even heard of Claude. Speaker 1: So yeah, yeah, yeah, yeah, yeah. So my advice there and like what I do is I have literally a handful of people that I follow. So I have like a playlist on I'm on YouTube and a group or whatever it's called on Twitter. And I just kind of look at those folks once a day. And if there's something I need to be aware of, I kind of get it from them. So if someone wants to make me that person, that would be cool. But no, Corey, to answer your question, I create a fair amount of content. YouTube is my primary channel these days. And so what I want to be able to do is say, take my YouTube content, and slice and dice it a bunch of different ways. And so what it would do is, so we have like, again, context, and this is not like super robust right now, but like, it has, This is the context of what I want the emails to look like. And this is like my style guide, right? And then I have output formats that I want it to go in. And this is just where it would go. And then I have prompts that I'll send to the LLMs for each type of content. So a LinkedIn article, LinkedIn post, YouTube short script, sub stack note, which is like a tweet. And a Substack post. And then what this project does is it just goes and gets the transcript from YouTube. And this is just some Python that Claude wrote for me. And then it repurposes it by calling these prompts with the transcript that it got. And so ideally, and like I just, I'm like, this is like 90% done. Ideally, what would happen in this project is I would set it up on my computer. We're going to talk a little bit about how to set up content marketing on a cron. And cron is just like a way to make things happen on a schedule. Yep. It would go get the transcript. So set it to happen every day at 10 o'clock in the morning. It would get the transcript. It would repurpose it and send the transcript to all of these prompts, And then I would just wire these up to be published or give them to an assistant or give them to my open claw or whatever. But that's how I'm thinking about doing what you're talking about. So I don't have separate projects, but I have separate prompts. And because like that's tone and format and length and pacing and style and stuff like that. So that's how I'm thinking about this. And this is like, again, 90% done and we just have to like, finish doing the rest of this, but that's how I'm thinking about it. Speaker 2: And I mean, this is essentially a repurposing project, right? It's not like it's, I guess it's pulling your YouTube transcript and that's kind of the input that it's starting with, but it's not like, This project is skewed more towards one platform or the other. It's more of like a general repurposing project. So my question, you mentioned OpenClaw a couple of times. I've been deep in the experimentation with that. And I'm sure you probably have your opinions on that tool as well. So another question would be, So, obviously, Claude Code is an agent harness. OpenClaw is an agent harness. Is there any, like, drawback from, let's say, instead of having a Claude Code project that's like your repurposing project, is there any drawback from having, let's say, like an OpenClaw agent where they're just your repurposing agent? Same context, same everything, but it's like OpenClaw is the one doing the execution versus Claude Code. Is there any real difference there? Speaker 1: Yeah, I don't know. I don't think anybody knows. I'll give context on what OpenClaw is for folks and that helps me think through my answer here a bit. So Claude Code, if I think about it as an evolution, so the first time you used ChatGPT, you're like, whoa, it talks to me and it remembers stuff and I can ask it questions. You know, my wife, we were at lunch yesterday. She's like, it's like having my best friend that I just like chat with all the time. I'm like, yeah, like, it's crazy. But that's like, that's where she is, right? And so I think a lot of folks here are like, kind of cool. I'm past that or I want to get past that. I don't want to use Claude Code, which is like, I can type stuff. I can do commands. I can maybe put it on a cron, which is like pretty advanced for Claude Code. But really, if you're not sitting there, it's not doing stuff. The next level is an AI agent. And like we said, Claude Code and Manus, I put very much in a similar place, except for Manus runs in the cloud, are agentic tools, but they're not agents where they're just going to go do stuff on their own. Speaker 2: Right. Speaker 1: Janet, which is my open claw, shout out to the good place. Speaker 2: Just does stuff. Speaker 1: Like she pinged me this morning. She's like, hey, I made a website last night. Like give me the DNS so I can publish this. I was like, whoa, okay. That's cool. Like, so Janet has a website now. We got her a domain this morning and set it up. I've seen some insane stuff. Like, you know, a guy gives his like an email account and a credit card and it gets a phone number and starts calling people. It's where it gets a little weird. So I guess my answer is like back to the SOP thing. It is like if you want to be really specific with what you're getting out of your AI tool, I probably wouldn't just give it to an open claw that you gave kind of shitty, sorry, it's kind of bad. And today I'm going to talk to you about how you can use context and guidance too. Like if you just say, Hey Janet, go write me a tweet, like it's probably going to be pretty good. But if you said like, Janet, go read my last hundred tweets, create a style guide, create the do's and don'ts, blah, blah, blah, blah, blah, and create an SOP and a skill around it. Like it's going to be a lot better. And I think at that point, whether Janet runs that or whether a Claude Code thing on some kind of cron runs that I don't think is really that much different. Speaker 2: Yeah, and again, I think I have my answer. I guess I was making an assumption and I should have vocalized this, but assuming context is the same in Claude Code versus OpenClaw, which is going to perform better on that type of task? But I mean, like you said, it's probably going to be roughly the same, especially if Claude Code is using Opus 4.5 and OpenClaw is using Opus 4.5. It's pretty much going to give you the same output. Again, the gaming context is the same. I guess what I'm trying to wrap my head around is like, Again, I have one open-claw agent right now. Her name's Claire. She's kind of just like my general assistant. I'm starting to realize I'm giving her all these skills. I'm giving her writing skills and outreach skills and all these different things. I'm thinking about it from the perspective of if she was just my employee, I don't really want an employee that's a jack-of-all-trades. I want an employee who's a specialist. So I'm kind of taking a step back. I'm like, all right, maybe I'm just going to make Claire my assistant. She's just like your typical admin assistant, email, scheduling, that type of stuff. I'm thinking I'm going to just spin up a new agent and maybe that new agent is my YouTube employee. Like his whole job is to go kind of like what you showed us with the YouTube magic project. He goes out and he researches titles and he writes intros for episodes and he does guest research and all the stuff related to YouTube and he's just the master of that domain. So that's kind of how I'm starting to think about these like agents or agentic tools. Yeah. Speaker 1: I think that's a very solid way of thinking about it. I don't know the answer. My personal approach right now is I have Janet, which is kind of at the head of things, and then I have six or seven different sub-agents under her. So like one for Castus, one for Outlier, one for all my content, one for like my health and fitness, one for our personal finances. And Janet can call and invoke and spawn those agents anytime she wants if she feels like she needs it. We're thinking about getting an OpenClaw setup or an array of OpenClaw for our employees at Castos. And in that situation, we would have what you're talking about, where like Kelly has an agent, Jimmy has an agent, you know, Sergey has an agent. They would share some context over here, like talking about the parent kind of thing. But we would all be horizontal. And in that respect, and maybe just my last thing on OpenClaw, It is not safe. It is not the thing you install on your main computer. It is not to be used on the same machine where you have access to Stripe and your email and your personal stuff. I have it on an old MacBook sitting here on my desk, and if we do this with Kastos, we'll put it in a DigitalOcean droplet behind Cloudflare and all sorts of crazy security, and it will all be dockerized into a container per person. We talk about like taking the rails off, like that's what we've done with Claude Code a little bit, right? Take GPT or, you know, Claude, and we put it on your computer and give it access to a bunch of tools. OpenClaude does the same thing, but even more so, really intentionally, like it can literally do anything. And you don't want that to happen without knowing what it can and can't do in a place where it can do harm to your business. Speaker 2: And we've seen, I mean, I'm sure you've seen the horror stories on X over the last couple of weeks. I mean, people just setting it up with no guardrails and just letting it go to town. And that can only end poorly. In fact, mine's set up virtually right now and my business partner is the tech brain behind what we do. You know, done all the things that need to be done to secure it properly. So I do feel comfortable in that regard, but I am literally, as soon as we finish recording here in a couple of minutes, I'm going straight to the Apple store here locally and picking up a Mac mini to start, you know, hosting mine locally and start building out an actual team of agents that, that are, you know, I can give a higher level of access to because it's posted locally and more locked down. So yeah, something I'm going to be doing future videos on and diving a lot deeper into. Speaker 1: Yeah, I'd love to come back and chat about it because I think it's... My interpretation of where we are with it is that this is probably not the answer. This is not the thing we'll be using in five years. But this archetype, this model is a bad word. This type of thing is the next evolution, very like ChatGPT to Claude Code or Codex or whatever to an autonomous agent or agents. OpenClaw is absolutely the thing. I don't think any of us expected OpenClaw to be the implementation of it, and it probably won't be the ultimate answer. I think that someone, and that is probably Anthropic, will come up with a version of this for you to use. Speaker 2: For sure. Yeah. And that's something that we've been saying internally is like, look, we're staying agent harness agnostic and just trying to master the underlying principles and just trying to get a really good grasp on this technology because you're right. It's not going to be open claw in five years. I can almost guarantee it. So whenever that next one does come along, it's like, hey, we're already prepared. We know how this technology works. We understand the principles. And then it's just plug and play on whatever the, again, the agent harness du jour ends up being. Speaker 1: Yeah, and the principles are all the same. Yeah. Speaker 2: Perfect. Well, Craig, this has been fantastic. I've learned quite a bit myself and I'm sure the audience has too. Speaker 1: Awesome. Speaker 2: Before we go, two things. One, go ahead and drop the URL for the SEO skills or the marketing skills pack you mentioned earlier. People can just go there and download it. What was that URL again? Speaker 1: Oh, gracious. I would just search for- Sorry, so yeah, the Claude project is SEO Machine, so seomachine.io. And then if you want to search for just the marketing skills, I would just search for marketing skills, Corey Haynes, so spelled like your first name, and then Haynes, H-A-I-N-E-S. Yeah, it's great. And so you could use either or both of them, the marketing skills, kind of just the skills. The SEO machine is the whole kit and caboodle, including the skills. And then you just fill in the context. So that would get you like zero to rocking and rolling pretty quickly. Speaker 2: Perfect. And then where can people go follow you or find out more about you? Where do you want to send them? Speaker 1: Yes. I think Twitter for social is the best place. So I'm the Craig Hewitt. And then YouTube, I'm youtube.com slash at the Craig Hewitt on there. And I talk all about this stuff all the time. Speaker 2: Awesome. Well, Craig, thank you so much for your time. We'll definitely have you back in the future for the folks in the audience. Hope you enjoyed it. And we will be back next week as always. Thanks, guys.