Why I Stopped Building AI Agents and Started Using Claude Cowork

I'm sorry. This is just more efficient. This is what I'm going to do. I don't want to spend all this money on these other API calls constantly going out. And if I have multiple clients, I'm going to convince them to get their own subscription and just do it. Because two things. If I have 2,000 clients, they're all paying me money and then I'm managing all these API things. Sure, I'm making money. Sure, it's handling certain problems. But what's the job to be done? The thing is, I'm trying to create value. This is going to get updated naturally. I don't have to spend any money, any effort, any work other than my monthly subscription to get an entire team of people building a better agent for me.

Okay, so I wanted to try something a little different today. I wanted to use Claude co-work instead of my usual VS Code setup to essentially create a writing and scripting area, but not just like, oh, hey, let's write some scripts. I want it used on things that have been working well, my actual social media, things like that. And what I'm going to do is I created a folder on my desktop called writing and scripting. Now I already have a kind of architectured folder that does this kind of stuff, but I wanted to start fresh to kind of figure out what I wanted and then apply it to my main folder later. So I logged into my good old Claude desktop app and just clicked on co-work and then went ahead and selected my folder that I had there. Um, this is a brand new one.

Now, the way I want to kind of frame this is you want to break down your processes in your head first and then figure out how to tell that to Claude. So, it's not just, oh, let me have this copy and paste prompt. Anyone that's selling you copy and paste prompts, minus like templates and markdown files, it's not really going to help you. So, the main thing that I'm going to do here is actually kind of describe um what I'm doing. And usually I do voice to text, but I want to type it out in front of you.

>> [music] [music] >> Right? We don't want to tell it just simply task to do. We want to conceptualize the working space for it before we start working in that space. And so I'm just going to go ahead and send that to that. Um, now again, usually I work inside of VS Code when I'm doing this. I I just find it a lot easier to navigate the files um and create like Claude MDs like this. Like, hey, here's how it actually works. And this is what my actual um file system looks like. I've used a much longer process over months creating this uh because it's just based around my workflow, what I like to do, my fuzzy tasks. Um, but I wanted to start fresh and I wanted to use Claude Co-work to show that you don't necessarily need to use VS Code. It's just a different way of working.

So, obviously, it's going to ask me some questions. So, what's the main subject matter of your channel? Um, I am going to paste transcripts and everything. I want you to ask me questions like this after I send you the transcript. So, the reason I'm doing this is I know that if I told it here, it might be missing some context or it's the wrong order in which to give this information because you want to think about the order you're giving information to Claude. It needs to it's slowly creating a map, right? Every word you give Claude is a coordinate in a massive multi-dimensional vector space. So, you are giving it a map to find data. The data that you want it to create technically exists inside of it already. you're just creating the map and the instructions to go find it and generate it, right? And so you want to make sure the order in which you're giving it is going to change, right? If you take a left and then a right on a map versus a right and a left, you're going to end up somewhere else. And so you want to make sure you know how you want to describe this. So in my head, I know that I want to tell it after. I just want to focus on this kind of process. Um, perhaps we wait to build the skeleton till you get those. Um, right. And this is me messing up in the original prompt. I should have known that it needed more information on my channel and perhaps described something there or added it in. But because it's asking me this question, I can immediately see, oh, it's misaligned in exactly what I want to say. So, let me go ahead and switch it up a little bit. [snorts]

Now from there, how many pillars do you typically work with? Let's decide as we explore this space. So now I realize what I should have done as my first prompt is explore the transcripts and everything first and then create the kind of folder space. But now it knows that the end goal of exploring all this is a folder space um first by looking at the info and such. Um, and then for scripts and writing as I do, do you want a single idea folder or folder? Um, yeah, single idea folder. That one's fair. So now it should go read that and it should, if we answered that correctly, say, "Hey, here's what I'll do. Create a minimal receiving structure, dump transcripts, and start a clock." Cool. Happy with that. So what's going to end up happening is the folder that I created right here on my desktop. It should be immediately creating stuff in there. And as we can see, that's exactly what's going on, right? writing a file, creating the initial raw folder. So, let's see. Source material, we've got transcripts, voice brand, community concepts. Okay, cool. So, source material there. We've got the Claude code, which again, this is why I like working in VS Code because I don't have to open it up in another folder or file. But hey, same but different. Um, cool. It's already mapped that out. And again, the reason we want to map out the folder structure in the Claude MD is because every time Claude's going to have to go and look through all my folders to see what's in there. But if we tell it what is in there simply and quickly, it'll be able to roll through it. Um, we're going to edit this Claude MD as we go along. This is to save tokens or if I stop and move along. We don't want to just fill this Claude MD with a whole bunch of stuff. We want to constantly removing and putting back in, removing, putting back in. There's ways to do this with memory files and things like that, but I wanted to start really fresh to have control of everything. And I want to make my own skills along the way. And the reason I want to make my own skills is right now Claude has great skills that I can use, right? So if I if I come over and we go to customize and we go to plugins or skills, they're skills that you can automatically use. But skills are again just markdown files. They're processes. And if we look at brand and marketing and we go to the skills, they have a brand voice skill. I don't like this personally. The way I do my brand personality or how I would ask my questions would be a little different. If I was making this skill, I would write it a little different. It's not a bad skill. In fact, it's a great one, but one of the big things is the kind of way in which we're breaking down the pillars. I'm going to take a different approach, which you'll see towards the end of this video. This might end up being a really long video, so I apologize for that, but you know, that's how the cookie crumbles sometimes. So, yeah, skills are great. You should download skills, but it's even better to make your own skills, uh, if you can. So, let me go ahead and go back to our thing content here. So, it's made that. All right.

So, now what we're going to do is I'm going to go ahead and start pasting some transcripts. So, um, I already kind of took a transcript. These are actually from Instagram videos that I took transcripts from. So, these first set are from Instagram videos, two of them. Sorry, I let the recording run long. I I recorded it quickly with um uh OBS because I was lazy and just transcribed it quickly. So, it's like double recorded. Um then the next set is from one single YouTube video that I'm good right now. Gemini, um I don't need the timestamps on this one. So, we're just going to go talking talking and we'll hit control C. There we are. Pop that in there. Okay. Also, I have two rabbits in my lap right now. If you see me fidgeting around um the they got into a fight and I basically had to wash them and whatnot and so uh they're kind of warming up in my lap right now as I record this video. So, if you ever see me fidgeting, that's that's exactly why I've got two wet rabbits in my lap.

Now, this next transcript is a different type of video where instead of sharing my screen, I create animations. Now, Claude has memory on this. This is Claude. I've set the animations, but if you didn't have memory, this is what you would want to describe. And I'm just describing it to double check. There's actually a research paper out there. Um, I'll probably post it in post- editing right here where it says saying something twice actually increases the accuracy in context windows. Uh, so it's not a bad thing to have it in memory and then also remind it again of that memory. Um, so there's transcript files. I'm sharing my screen and create animations. Other videos I do or sorry sharing my screen. Oh my gosh, I lost my train.

Now, this next transcript is a different type of video where instead of sharing my screen, I create animations and do voice over. This and this one did well, too. [snorts] Um, so let's go find that one. We'll go to my channel. Oh my gosh, I broke 10K. I didn't even notice. Holy crap. Thank you, everyone. Super kind. Uh, that's I've been trying to get subscribers on YouTube for so long. I've been posting forever. Um, and it's probably up like I mean years ago I started posting and finally finally I think right about here I started getting more views and then kind of none and then more views again and then this one which is the one I'm going to copy started getting more interesting. So I'm happy that people are digging it. Um, that's why I need to start get more focused on how I'm creating these videos because uh there's obviously something there that works. So let's see. This one was a good video. Really liked it. Lots of other people also liked it. So, we're going to go ahead and throw that in here, too. Say, okay.

So, I have a bunch of comments and things as well. We will look at for the audience, but I just want you to understand my voice, the the things that work. These are my best videos and highest views. Um, I feel like that's fair. Pull the content from what works best. Uh, now granted, both of the my best videos I didn't script beforehand. So, I'm going to be kind of hesitant on how much I lock in scripting and writing because obviously I I've got one that did 1.4 million and I just recorded it that day. I had another one, my best YouTube video I recorded like this. So, I'm going to be careful how much I'm actually worrying about Claude or doing this normal content process because I think um this copype content production is actually why some people are struggling to get their views and stuff uh to get things out there um [snorts] within reason of course.

Now, again, I think I'm going to end up Let me actually open this file up in VS Code anyway. Um, just so for those of you who haven't used VS Code a bunch or are just getting into it can understand like why I use it over the cloud app. Um, so we'll go to my desktop. I think what did I call this is writing and scripting. Yeah. Um, so for me, I I us I have a second monitor, so I usually put this on another monitor, but now I'm able to just quickly pull up and I can bring Claude right in here and chat about it and be like, "Hey, I want to work on this or similar. I want to add this." And then I can add multiple instances of it to be able to write and talk, which I just I love that ability so much, which you can kind of do in here with chats and things like that, but it's just it feels a bit more clunky to do that. And I prefer to do it in VS Code, but to each their own. Some people want to focus on one task at a time. I'm insane and like to multitask and have like 37 different instances of Claude running um at a time, but you know, it's the way it is. And sometimes I like to look and see what's going on, what it's kind of creating. Um, I like that it's doing its to-do list in JSON format. Nice structure just to be a like, hey, this is what we want to do. Um, which is really good. Let's see what its thought process. Where is it at here? Um, so user given me three transcripts, two from Instagram. And for those of you who aren't reading its thought process, you should really, really start doing that more because it helps you inform yourself of what happens when you give it something. And then you can re-evaluate and be like, "Oh, it thought of this. Maybe I can say this differently in my next prompts or injections or whatever you're looking at. But they want me to understand their voice and content before building. Let me do an absorb here. Key observation subject matter AI specifically folder architecture concept using file system markdown instead of coded agent frameworks. He's teaching people how to use Claude code AI tools with folder based routing. That's fair enough. Authority. Yep. I did I was in the Marine Corps. I did teach professors. I have been in the AI space. I've been in it more than two years. So we might want to double check that one. Um, you don't need to build agents with Python frameworks. Yeah, I mean, yeah, you don't. That's or at least the concept of what people are trying to do with agents. I I do kind of hit on that from a marketing standpoint. You have got all these crazy frameworks out there like Crew AI. Um, not crazy, they're not that crazy, but like Crew AI or, uh, semantic kernel is a big one that Microsoft's put out that you can do in Python. um if it ever loads. It's just basically a way Let me see. Yeah, here. It's basically you have your codebase which is separated from your models. You can swap models out without it ruining your codebase. You can do hooks and plugins and things like that. Basically everything that you can do right here with a couple button clicks um or with a folder. Um and you know there's certain use cases for enterprise where you might want something fancy like like these or uh you know you've got lang chain of course as well. Uh, Lang Graph. Um, oh, you know what's a fun one? I don't know if they're still doing it, is DSPI. Um, yeah, these guys, these guys are interesting. They're it's a framework for modular software that actually iterates on the code itself. It's been a while since I've uh kind of messed around with it. So, we'll have to see how they've updated because it's all about optimizing prompts and weights and things like that. And uh I like their ideas. I'm I'm curious to see the feasibility of it being widely adopted, but that's, you know, I don't think that always matters. Something doesn't need to be widely adopted for it to be useful, but wide adoption makes it easier to have impact, which is important there. Anyway, beside the point there, let's go back to our boy here. Um, so it's going through and it's creating some series of files. So, um, voice character direct and confident with slightly confrontational edge. I guess that's fair. I might switch this up. I have my own voice file, but breaks down technical content simply, weaves in historical examples like Jacquard loom. One thing about having an AI talk about your voice without you writing about your voice is it's going to write about your voice in a way that an AI would write about your voice. Even the best models still struggle with this in my opinion. I mean, Opus 4.6 on it, right? Claude has been one of the top models for handling voice. So has Clling. I mean, I've seen Grok do a good job. ChatGPT once in a while can as well. At the end of the day, if you want your voice characteristics, you probably should write or at least use voice to text out a bit. So, we'll mess around with that here in a bit. Content types. Um, yep, that's spot on. That's exactly what I want to do. More screen sharing. Um, and then also animations. Um, oh, this is pulling it from I got to say that the other ones uh performed better as well because this used to be my best performing video and in my memory I said, "Hey, this is my best performing video. Can we look at it?" So, I might need to double check that. But again, this is why it's important to read what it's thinking. You want to look every single one of these words is a coordinate in its vector space. It's directing where it needs to go. So, you need to make sure it is nuanced align with everything you're thinking. If it's off a little bit, that little bit grows the more you use it. So, you need to nip that in the bud as soon as you see it saying something you might not like. Um, let's see. Uh, all three transcripts are saved. Here's what I'm picking out. Your voice is a clear pattern. You teach through layers. You start with something uh people think they understand and peel it back and show them what's going on. I do like that. That's fair. Instagram clips do it fast and punchy obviously because short attention spans. Tutorial does it by walking through a live folder structure. Sometimes I do, sometimes I don't. Animation crosses 200 years of computing history. Absolutely. Same instinct every time. Go deeper. Connect it to something real. Make the listener feel like they just gained X-ray vision. I like that last part. All right, that was pretty good. I'm not going to lie, that was good on the AI. Uh, you direct and lead with your authority, but you earn it. Marine Corps's crypto work. Yep, yep, yep. Two content modes that are working. Practical, tactical. Oh my god, I love that rhyme. Narrative and conceptual. One thread pulled through history. Yeah. Okay, that's fair. I do think these are two really solid content modes. So, this is this is this is useful, right? Because this isn't locking me into a concrete way of producing a single video where every video has to be structured this way. It's more of like, hey, this is the rough concept of the videos we want to sit in. This is what people seem to like. Provide value in a tactical way or give a narrative of something. And honestly, I think both of these modes could probably be combined together as well. Uh, what I'm noticing about this animation videos work because it's familiar act. Yeah, that's fair. Take clear positions. Absolutely. I just did that five seconds ago. Um, that's sure it works because it's generally useful. I'm ready for your comments and brand voice. Okay.

First though, we're going to address before we give it anything else. Let's address these main things. So, um, I think I think you got pretty close close on the voice. Um, but the main goal here, the main goal is to create a way of thinking around my voice. I know my voice and can handle it naturally most of the time. I want to c to create files that catch you up to being aligned with my voice and more importantly thought process. But that being said, I think you are darn close. So the goal here is to not lock ourselves in concrete with exact structures or format but certainly have a good set of guiding themes, principles or pillars etc. Now for the comments these are going to be huge. You will see some of my comments as well but it shows both why people like my videos and more importantly why what they need. And now this part is going to be harder for a lot of you to do and why it's so important to share your work, show your work. Speaking of that, great book, not sponsored at all. Uh if just buy this, just buy this book and read it. It takes like you could read in an afternoon. This is one of the most fundamental books that I have ever read, watched, whatever you want to call it. Um it's it's just all about showing your work. You should show the whole process. What I'm doing here is showing my work, right? This is how you get out there. It's how you learn. It's how you can teach others. It's that whole deal. By showing the work, you then get things that you might even if it's one or two comments, right? You get an idea of what people need and want.

Now, this video I didn't get as much. So, I'm actually not going to use the comments. I mean, I got a bunch of comments obviously. Like, uh I think what we we did, yeah, 46 comments. That's pretty good for some of my videos, but it wasn't as you know, the the feedback isn't exactly what I wanted. Here in this video, we get a lot more phys specific example. Hey, I was looking for this or honestly, I'll share this with this. Right? These are really good. So, I'm just going to rip through all of this and and there's also a lot more comments on this video. Now, I just take that and I'm going to paste it right in here. Bunch of them. Uh, these are comments from the folder architecture video. I think we should take our time with each one and go through them. I will also give you a bunch of intros from my school community. I usually wouldn't give it this much. Usually, I'd scrape it, put it into an Excel sheet or something like that, but right now, we're just we're trying to move quickly. I think it'll handle the idea. I I'm not trying, you know, I want to spend individual time with each person and comment. I'm just trying to get a an overall vibe of them. I mean, the more details and data I could create, the better. Um, but we're rolling through it here. All right. I believe in you, Claude. Look at all of it. Go get them. Whatever I do, do not hit show more on that is what I should do.

Okay, Claude thinks it's a gold mine. That's not surprising. Massive amount of data user wants to save these comments and analyze them. Oh, shoot. I didn't mean to save it. Oh well. You know what? Let's see how it handles it. I'm very curious. I think it'll be able to go through it. That was a lot of comments without me organizing them, cleaning it up at all, throwing it in Excel. I usually like to do that for these models, but 4.6 has been pretty monstrous at handling most of these tasks. So, we'll go ahead and let them. But, um, appropriate folders and give the analysis. So, oh, it's Yeah, okay. It's running their agents to go handle that. And this is another reason, right? like [snorts] in order for me to do the work and create an agentic system that this is doing off of a single subscription uh with zero API call or not zero obviously there's API calls going in the background but like I don't have to pay for an API subscription this is just a straight up one once a month payment it's the same amount every month I'm sorry this is just more efficient this is what I'm going to do I don't want to spend all this money on these other API calls constantly going out and if I have multiple clients I'm going to convince them to get their own subscription and just do it. Because two things, if I have 2,000 clients, they're all paying me money and then I'm managing all these API things. Sure, I'm making money. Sure, it's handling certain problems, but what's the job to be done? The thing is, I'm trying to create value. The clients are going to pay me to either create a more robust system that works for Claude or B, we're going to look at something that we absolutely have to create that needs this kind of next level stuff because at the end of the day, this is going to get updated naturally. I don't have to spend any money, any effort, any work other than my monthly subscription to get an entire team of people building a better agent for me. It doesn't make sense for me to compete with a multi-billion dollar company unless I know for a fact I have one niche thing that I'm going to out compete them in. And don't get me wrong, this isn't me. If you're building some sort of very specific little agentic structure that you think could compete with whatever Anthropic is doing, go for it. But at the end of the day, I'm not trying to compete with Anthropic. I'm trying to get business deliverables. I'm trying to help people work in their companies. I'm trying to use this technology to do something else, not just make a better version of the technology, right? Um, and so for me, they're handling all of that absurdity so much easier, so much cheaper. Right? If I was to do this with a Claude bot or something like this, this would cost me probably I'd already probably be at $100 in token usage based on the agents that are going in. Maybe not that much. it'd be like $5 by now, but still $5 in token usage and we're barely we've barely started. It's been a minute and a half. Other part is it's cleaner. I don't have to think about it. And people are like, "Oh, but what how do you use it from your phone?" You can use dispatch inside the Claude app, which I'll make a whole video on this. But you can literally just connect it to your phone right now. I can go log into my app, go say, "Hey, go do this thing." I can even have it control the browser and stuff as well. So, I I'm I called this in my Claude bot video that Anthropic would drop an update that made um Claude bot basically useless or not useless. It's still very useful and powerful, but like just so much more secure and simple by, you know, having it inside their app because all Claude bot isn't an AI. It's it's just an orchestration layer. Orchestration layers can be added to every other AI. And the companies that have the AI can do it better, faster, and cheaper than everyone else who's creating the wrapper and a third party. So why not build systems and companies that assume that these apps, these processes are going to get better? Like that's what they were said three years ago and that's what I've assumed and that's what I've done for the last three years. And my processes just keep getting better. I don't have to feel like I'm running or chasing or trying to beat the competitor of AI. I'm just getting better every year with what I'm doing. So, yeah, I don't know. Short little rant there, but um let's see. Let's go ahead and look at what's going on here. Um we don't get to see your thought process there. I need to So, there are ways to I didn't run it in this section to see what these agents are doing and actually um kind of dive in and look at their thinking even beyond just this basic kind of setup here. Um, but let's see. Save YouTube comments. Nice. It set it to YouTube architecture video comments. Oh, I love that. Good time collected. Today is in fact the 20th. Good, good, good. Um, and then there it is. Oh, and it's Marcy. Dude, [laughter] you know how long it would have taken me to build an agent that does this well, right? And yeah, maybe not that long, but it still would have been annoying to go to N8N and then have it scrape my YouTube comments and then have it pull that in and then have it turn it. I could just copy and paste it into a single really good solid agent like Claude or, you know, Codeex and it's going to do all of this for me, right? Like there's people probably charging $60 for a month for a tool that does just this, just what this agent did. Why waste your money on that when you can have all of the tools ever inside of one area and then build them as you need it? Again, I people who are saying they need all these crazy agents. No, you you need them, sure, but you don't need to be building them. You don't need to be wasting your time. You need smart agents. Again, if you're working for Anthropic or if you're working for a startup that's really ahead, maybe it makes sense. Maybe. But for me, for people who are in business, for the vast majority of the populations who are going to get value out of this and actually run companies, because remember companies, the only companies that exist aren't just AI companies. The people who are are doing well with refrigeration technology are not refrigeration companies. I mean, refrigeration companies do very well. They make multi-billion dollars, but like Coca-Cola. Coca-Cola exists because refrigeration exists, right? Uh, I mean that's not why they exist, but it's they're the reason they're able to get their stuff out there and successful and give you an ice cold Coke is because of that technology. How does this technology help you do something that was either a not done before or b at a scale that was never possible before? Everyone's trying to chase the hype of, oh, here's the best AI agent that I made and stop selling agents. Sell a company. Do a company that uses agents better than anyone else. And then guess what? If you build it around people who are already building better agents, you're just going to get better over time while they're wasting their time just focusing on something that's probably going to get replaced by the next Claude update anyway. Again, I'm not saying to be reliant on the big frontier companies. I'm saying understand that there's five or six companies, even open- source ones, that are going to do work that you don't need to do. And if you're in one of those companies, then it makes sense. Cool. Go for it. You want to be an agentic engineer using semantic frameworks or crew AI or Python libraries that you're custom building, go for it. That's fun. I've done that. I love it. But I realized where my value was and where the value is going to be for the next decade. And it's not in that in my eyes. Again, for most people, if you're if you're one of those engineers, go do it. Cool. My on me. But I guarantee you most people are not. And that's okay. You don't need to be one of those engineers. You need to be someone who does what you do well. Instructionalize your opinions, which is exactly what we're doing here. Um, all right. Let's see what's going on. Gosh, I keep getting these rants, right? First batch of school intros saved. YouTube comments are saved. 150 plus people. I have it all in contents. Here's what I'm seeing. Um, let's see. So, I've actually read through all of these, so this will be a good test to see if it actually found things. Um, data science, full stack devs, Ricardo. Yeah, Ricardo was cool because he he was coming at this from a software engineering angle and was really kind because he basically was like, "Hey, I really like the structure. It's not hyped." Uh, let's see. F senior backend. Yep. Axel, the Bay Area. Yeah, I've start noticed we're getting some Bay Area SWEs in here, which makes me feel good. Um, Don Cameron, for those of you who don't know what SWE means, software engineer, just Yeah, simple as that. Um, Don Cameron, 10 years in AI models, found you because they were already building agents the hard way and your folder thesis made them stop and rethink. They want depth. They want your research paper. They understand why it works, not just how. I think that's pretty agree, right? It's not that I'm saying agents suck or that I'm creating some super structured framework. It's that, hey, just structure your stuff by folders, make it work, then worry about the complexities after. Um, technical adjacent. Yep. We do have a lot of technical adjacent people. Um, Dean, yep, yep, yep. Breton, Jacqueline, absolutely. Product manager. Wonder if do they have um I didn't see the the other AI's. Wait. Oh yeah, Jesse. I I'm doing my work with um KPMG and I want to try to do some work with Deloitte here soon, but we'll see how that goes. Uh Jacqueline, they need frameworks translated into their language. Repeatable systems. Yep, absolutely. Hopefully they're getting a lot out of my work and stuff I've done there. Non-technical people who want to build appliance technician, roof tilers. Yeah. Yeah, we gotten some cool people from Australia. I used to uh I I spent a lot of time up in Northern Territory in Darwin. Um and I spent a bunch of time in uh Katherine, a little bit south of Darwin as well when I was in the Marine Corps. Uh crazy place there, man. I almost got a felony for disrespecting the queen. Um that's a crazy story. I'll probably save that for another time. Um let's see. How do I structure this for my workflow? Yep. How do I make this persistent or stop re-explaining? Absolutely. What's the difference between skills MCPs? I need to start answering this one more. I think this might be the first video that I do after this one is diving deeper into this because there's so many concepts out there. It's like they're all the same thing. You look at the background, all of them are just natural language with some really decent routing. Um, and I'm simplifying it. There is important differences between it all. But there is a lot of confusion there and I need to hit that. There [laughter] is a lot of confusion. Good job, Claude. Um, cool. Obsidian integration. Yes, I gotta show I'm I'm I I I use Obsidian to create mind maps sometimes. Um and so I really need to dive into that. Uh how do we use this teams? Yeah, going into Git, how you would share these files. I really think eventually it's going to be as simple as just logging into co-work and you all would have a shared context folder and they're going to handle a lot of that processing and work for you. Um but again, it's always nice to have your own control and build it. Uh your entire point is nonsense about tool calling. Yeah, that sounds right. 2023 era mega prompting. [laughter] Oh, these it's the objections I have from the agent crowd or people is is hilarious because they they're totally missing what I've been doing in my point. Uh which is fine. You know, maybe I can explain it better, but I can always Yeah, I do think it is worth addressing. I fill in that. See, architecture. All right, let's see. This is this is what I'm curious about. What What architecture did Claude choose for me? Or what pillars, right? Is this gonna align with the pillars I had in my head? And again, you don't want to take this stuff at face value. Everyone say, "Throw it in the cloud and it's going to make you rich." Stop that. If this aligns with your personal opinions, your conviction, then that's okay. If it doesn't, try tweaking it. That's okay. Sometimes it might know better. But the goal isn't for it to tell you how to do things. It's the goal is for it to organize how you're already doing it and make it better. Uh, so let's see. Folder system itself, three layer routing, Claude MD, workspace design, naming conventions, contextes, your core thesis can go to 90%. Yep, yep, yep, yep. That's I mean that's just fair, right? I I I want to be careful with how much it sticks on three layer routing because I do five layers in my research paper. Um the whole idea behind layer routing is nothing to do with crazy layers. It's the idea of how do I break down a task into its parts? Which parts are higher level in the abstraction? Which parts are lower? Right? If I'm building a watch, the process of putting the gears in are going to be more specific than what color I choose for the watch. And there's different layers in which it affects everything where each layer can be important to the other, but it's how detailed each layer gets. That's all it is. Sometimes you might need more than that, but I think three layers is usually good for most workflows. Um, goes with 90% of people throughout the or absolutely foundation concepts that make the architecture make sense. Tokens, context. Yep. up your hello world lives here. Um foundations. Yeah, that's fair, right? Computing foundations. The it's not just about learning AI foundations. It's about learning why computing exists, right? We've been doing the same process for computing if this then that up down boolean logic for literally 200 plus years. Technically 2000 if you really dive into the the older ancient Greek things, but I have videos on that if you ever want to see them. Uh applied workflows. How do I do this thing for How do I do how do I do this for my thing? Yeah, that's exactly what this video seems to be right now. So, content creation, school intros, play cool ecosystem, skills, MCP servers, Claude, right? This conversion between skills and agent. Yeah, so this is this refers back to the previous point where people were kind of confused of uh where is it? This the difference here. So, point four is into this pillar ecosystem. um proposed folder. Okay, so here here's important. Really pay attention to this. This isn't just for organizing for Claude. This is organizing for you. You want to look at this and be like, "Okay, if I'm writing and looking at my transcripts or reading it is this is how I would want to look at it. If I wanted to come in here with Claude and start looking at the community transcripts and affecting my voice and brand, how would I do it? What order would I want? Subfolders. How many subfolders are you going to want to sit there and read through certain things? Right? This is important. What is your thought process in how you would not just build this stuff, but evolve it? When would you want to come in and be a human and edit this markdown file or edit this one or add something? Maybe in the pillars, you want a place or a folder where you have testing pillars, new ideas, right? You want to think about all of that. Uh for right now, um I like all of that. Love the design. Also think the what do what did you call these? Um what they eight things of what they want should be referenced or added in some way to a content pillar or each pillar individually. As in four could go wait. See statement for difference. Oh my gosh, I can't type anything. Difference between MCP skills, etc. should go in pillar 4, right? So this statement for I think falls under ecosystem, right? What are you pulling in? Um pillar four and so on. So that's going to immediately just a little tweak along its thinking because I didn't see it mention that, right? That it was going to throw those in there. But I think these are more useful than like, hey, this is how you should write. It should say these are what people are saying and then we'll creatively write because you want to do two things. You want to direct Claude to be talking a specific way, but you also want to direct it too much because then all of your content will sound that way because that Yeah, that kind of just it locks you in too much. And I've noticed that when you have it create scripts and things like this, again, you you're not trying to take the human out, you're trying to scale the human. You're trying to scale yourself. And a lot of people mistake automating with scaling. Automation does not mean you will scale. Just because you automate your DMs doesn't mean you're going to guarantee more people coming in. Especially in this day and age where people are just hating their bots. I went through and DM'd a thousand people myself. Now behind the scenes, you're like, "Oh, but that doesn't scale." It's the only things that scale. I got a thousand people to join because they're like, "Oh my gosh, this is so great. I'm talking to a real person." Like it's the I sent videos through an Instagram. Right? people are they're not thinking about the job to be done when they're creating these automations and processes and things like that. Um, so let's see. Uh, good call. Let me tap those recurring tasks and the pillars. Then build everything out. Cool. Here's how the eight audience the eight audience needs to land. That's it. Sounds weird, but I'll take it. Um, pillar one architecture. How do I structure this from my How do I Yeah, see this is this is gold right here. This is what we're looking for, boy. Right. We want we want to take these concepts and create a mental mapping of where it should be generating content, not how it should be generating content. Because now it's going to read this when when we go later and I ask Claude, hey, think about the pillars. Look at what these pillars are. I want to create a series of videos, one focusing on each pillar or area, right? We could maybe combine pillar one and three for this video or something like that. we're always able to reference the concepts, not the structure, which is going to allow me to be creative in what I want to write about while simultaneously giving me the structure to plan future videos and things like that. Um, so let's see. Went through created yeah, let's let's go ahead and open it. Let's see. Pillar 2 foundations concepts. And see, again, this is why I like VS Code, and I'm absolutely gonna work in this file later inside of VS Code. Um, because well, that's just going to make life 10 times easier. So we come in, I can look into the pillars and see now I can just zip these up. Let me actually get rid of these instances and read through it. Also, Obsidian would be great to work inside of here. So maybe you have Obsidian and then Claude or you have this in there. I don't necessarily need I actually like my plugin that colors markdown better than sitting inside of Obsidian. Don't hate me for it. I know you can do that in there and there's a lot of ways to clean up Obsidian. I I'm a very just practical like to be at the base level person but to each their own. Um folder self is system. This is why 90% of the audience got through the door. Yeah, that's fair. I think that's actually so usually I'm careful with these types of statements because it'll regurgitate them a lot. But I think this seems like such a simple edit. It's not. I promise. Instead of saying this is the core thesis, say Jake wants you to understand but not regurgitate that this is the core thesis that that simple thing is going to stop it from regurgitating. Oh, this is the thesis that 90% people. It's going to say that all the time if it said that it still might a little bit here or there, but that's the aspect. And again, look how short this is. This isn't a lot. We're not adding a massive markdown file. you can do huge markdown files, right? So, like we when we're looking at our transcripts, they're going to be large or when we're, you know, looking maybe Claude's going to get a the main MD is going to get a little larger. Um, this one actually, so case in point, this is nice because it gives an explanation of the voice, but then references my brand voice uh MD when when it's there. Um, I'm actually going to have it create one. I'm going to give it one with my constraints. So, I have a folder right here. Let's go to my script lab again. This is my crazy absurd one. Right here, I have my voice. I think this one's too long, too wordy. It's not structured well. This is why I'm redoing a lot of this. But one thing I do love about this is my hard constraints, right? No M dashes, antithesis pattern. Stay away from it. Formal informal. There's also a skill called uh humanizer. What is it? Humanizer skill. Uh Yeah, the humanizer skill. It kind of does the same thing. So, if we read

It, um, remove signs of AI writing from text, blah, blah, blah. Here's all the tools that it's allowed. Identify AI patterns, things like that.

Now, the way it's identifying those patterns is um, essentially just what I'm doing here. Here's what these patterns are. Basically, it says here's the patterns that are detected. Significance. This one is quite nice. I kind of want to turn it into my own use though because there's certain things in here that I actually do like to use. Like I actually love quotes. I use quotes all the time in my natural writing. Um the artifacts that's nice, but look, I'm sorry if you're not paying attention and reading your scripts to see if there's still artifacts in it. That's a you problem. I I I think you should spend a little bit more time or if you can't tell Claude to not do that, there's a problem. So, I think that's a little extra and it's a lot of tokens that I don't necessarily need to use. So, I'd rather kind of work it in my own way. So, I'm going to clean that up later and throw it in. But, for right now, great structure. We can add this later. I don't want to mess up my current chat and muddy it too much uh because of how much was in there.

So, let's see. 10 files, five folders, cloud MD updated to phase two. We've got source materials with transcripts and files and we can add more into it later. Uh, we've got pillar files, we've got ideas, uh, which there's nothing in there. Uh, we'll get that in there, which is great. And then my brand file.

So, let's let's do one more thing because I think this video is starting to get a little long now. So, I want to I want to finish up. I'm actually going to do this right now. So, I'm going to go in and I'm going to take this and I'm going to go honestly I think you get my voice and you get the idea that the brand voice document shouldn't give too hard of constraints but rather give directions or ideas that allow Claude to be aligned with me not stuck in concrete. Um however here is some important constraints that I have used and like please make me either ah here we go actually this is what I'm going to do to change so this is this is gold if you are watching along this this far and are able to keep up with my sanity what I noticed here this is a giant oops this is a huge voice doc it's too big this can be broken down this is too much context and tokens at the wrong time I could give it to a separate instance of claude after especially all the constraints and so what I'm actually going to do is something a little cheeky here. I think I think we should break down voice into two or three MDs instead of one. As in we can have one that is for my tones, styles, ideas, one that is for long form versus short form editing of the ideas, and then one that is constraints as in cleans up or is injected in at any time. to speak on what not to do in your writing as in you as an AI happen to have patterns I don't like. So we're going to here are those conraints and we're going to say uh what do you think? What do you think? it'll it'll figure it out.

So, what we're doing here, the reason I wrote it in this way is I'm saying, "Hey, I actually like how you understand my voice. I'm happy with that. However, I don't want you to describe my voice. I want you to describe a markdown file that lets you or first future versions of you, right, that might not be this chat, understand why or where and be aligned with where my voice is." because telling it how to write creates more patterns that end up sounding AI anyway across all your videos. So you want to you want just enough chaos in there to be able to have enough. And then the idea of breaking it down into multiple markdown files here is the idea of like okay perhaps my one instance right I open cloud I say read all three we're going to write something and then it knows about all it but maybe I've already written something and I just needed to edit it and make sure there's some pillar talk in there um and I don't need to worry about the hard constraints as much or maybe I have two agents building kind of pillared script ideas and they make some really great sets and there's a lot of tokens and I've had it build like 15 and Then I make another one that's not, you know, poisoned by the other chat and is clean and small tokens. And I have it go in and say, "Hey, can you go read these files that we made scripts and just use my constraints and edit it?" And let's see if it understood that. It should have from my prompt there. But let's find out. Constraint list is sharp. Yep. Yep. Yep. Breaks your voice. I like three. Voice and tone. How you sound, how you teach, how you move through it. Directional, not rigid. Hopefully gives me alignment without a straight jacket. Damn straight. Claude, you get me, brother. Uh, former format patterns, long form versus short form. How an IG script differs from a YouTube tutorial differs from an animation voiceover. Yes, yes, yes, yes, yes. Structural guidance for each format. And I don't think it needs to be that much. I think it only needs to be a paragraph per one. So hopefully it doesn't go overboard there. Constraints, the never do this list, 12 rules plus anything else we add. This one gets read every time. That's also why it's nice because updating constraints gives me a constraint file for all of my writings everywhere and I could use it outside of this folder as well. This is just going to be a good constraint file. Um, I'd add a 13th about list heavy structure. AI loves to organize everything into number lists and bullet points when a paragraph would sound more human. [snorts] Yeah, honestly I I think that's fair. And I do never use that. I I was thinking about sometimes in writing it's nice to organize that way, but I think that's fair. Um, yeah. Are you Are you just doing this already? Did you just Did you say I'm going to do it? Okay. Yeah. Do you want me to do it? I'm sure. I guess you've decided that you think you should do it. I'm okay that. Oh, I didn't mean to access your file. All right. Whatever. Um, either way, I'm going to go look at it in VS Code. So, let's see. Brand voice constraints. Simple and clean. A little long, but that's okay. We'll take a look at it. Uh, not too long. A little bit little bit shorter than the other one. This is I think this is a decent thing, right? About 46 lines to 100 lines. Once you get over a 100 lines for a markdown file, you're probably you're probably missing breakdowns. You could do like you could realistically probably break something apart even more into thought process because your whole goal with these markdown files or these writings is to break apart thinking itself. when do you access to compartmentalize in software? It's called separation of concerns. I'm just applying it to uh I guess neuroscience, not really neuroscience here. There's probably some sort of better way to look at this. I'll have to read some more neuroscience books to give you a better kind of um connection to psychology and whatnot, but not psychology. Psychology and neuroscience are slightly separate. I understand that. I'll get hate in the comments if I don't separate that. Maybe I won't. You guys are nice. I'm just going on a rant later. Anyway, um voice and tone formats and patterns and constraints. Okay, workspace is operational. We can drop in. We can move it there.

Let's let's do one final thing. So, ideas are empty. So, we can go through. We're going to go ahead and test this here soon. Let's test it on a fresh one, actually. So, let's look in Hey, I want to make a new idea for a YouTube video. Let's see. Let's see if it Let's see if it figures itself out here. Oops, I accidentally pasted that in, but that's okay. Um, so it went through, it read the context, it read the band voice. It immediately knows where they all are because the CloudMD says so. Yes, Cloud MD says so. Good. Um, so we have our folder structure there. So, it shouldn't waste tokens doing blobs and trying to figure out whatever's there. It's just going straight to read them. It did have to do a glob here because it was checking to see if we had any ideas made already. Um, I might actually edit that because I don't necessarily want it to read all of my ideas all the time. Uh, at least in that way because I want to be careful how it's reading too much. Um, it looks like it's reading a bunch. So, I would have I mean that's just because I gave a really terrible description. I would say and compartmentalize, hey, I want to make an idea. Use these pillars real quick. But ideas folders empty. So, what direction are you learning? Um, format tutorial. Let's do a tutorial style with a focus on eco system and foundations. Um, I want to make it a little more technical but have the nontechnical group be able to understand it. So you see right here this is hundreds of functions that are being called right now. Essentially this is telling it it's immediately reading my ecosystem one. It's understanding all of the concepts and foundation. It's looking at tutorial style right our formats and patterns. It's understanding the audience which is technical right. So if we come uh to did we not make an audience one? Oh. Oh yeah. Okay. What these people actually need right? So, what this pillar covers, skills, things like that, it's all embedded into here. And we actually have the comments. Oh, it added the comments to it. I didn't even think about that. That's actually brilliant having the comments of what people said from a questionwise. Oh, that's fire. I didn't even think about it. Good job. High five, Claude. High five. That was fire. Um, and see, now it's actually going to be able to go through that. And now I get to just work with my ideas, right, my writing, my concepts in a much faster way, much cleaner. And I could do the same thing inside the app. I could ask it that same exact thing and it's going to do the same process. Um, but that is built and hopefully you'll all see a video of something built from this and it answers a lot of questions. But yeah, cool. That is that's the video. I think this was a long one. I don't even know how long I've been talking now, but hopefully someone's still here at the end. Either way, my name's Jake Van Clee. If you're not part of my community, please go join it. It's free. We've got plenty of things there. If you have any questions, comments, concerns, drop them in the comments. That always helps me out. As you can see, I actually am using the crap out of them to make new stuff. So, if you want to affect and change what I'm making, comment below and let's see what we can do. Other than that, happy learning, everyone.