Systematische Entwicklung von Codex-Skills für KI-Agenten

Today, I've got this proven six-step process for building Codex skills better than 99% of people, so let's not waste any time and just get straight into this one. All right, so these are the six steps that I'm going to go over. I'm going to explain each of these and tell you why they're so important, and then we're going to get into Codex and actually try some of the stuff out. And real quick, before we get started, I'm going to talk about what actually is a skill, so if you already know what a skill is, then just go ahead and skip past this part. But for those of you who need a refresher, here we go. So, I'm going to use my chocolate chip pancake analogy. Let's say we have this chef, and this chef makes the most amazing chocolate chip pancakes, and you want to know how to make those. This is essentially the output that you're looking for. The way that you would be able to copy this chef is by looking at the recipe that the chef used, or the recipe that the chef made. So, the chef gives you the recipe—this is you—and now you're able to pretty much make the exact same output, the really, you know, popular, famous chocolate chip pancakes because you followed the recipe. So, this recipe is basically the skill. This is essentially the skill.md file, which means a markdown file. It's just a simple language; it just means that inside of the skill file, there are like pound signs and asterisks to indicate like bullet points and headers and things like that, so it's—it's just natural language. But then the agent is basically able to take this skill file and just use it. So that if you say, "Hey, Mr. AI Agent, you know, make me those chocolate chip pancakes," it wouldn't have to be like, "Okay, well, let me just do some research on how to make them and I don't know exactly what kind of pancakes Nate wants, I don't know how big they should be." So, what I'll do is I'll just follow this skill, I'll follow the recipe, and now I get the same output that Nate's looking for, and I get it consistently the same every single time, because it's all documented for me right here. And these can be really simple, it can be a simple prompt, like, "Hey, you know, help me turn this email into something that's more professional or something that sounds like me," and maybe that's like my Nate email skill. But it could also be complicated processes like doing research on the market and analyzing, you know, 50 stocks and telling you which one to buy. So, really it's whenever you want to basically like codify some sort of process that you do so that you can delegate that process to an agent. And then you turn it into a skill, and now your agent can use those skills. All right, so now that that is out of the way, let's start with number one up here, where we have Reverse Engineer. So, the whole idea with reverse engineering your skills is basically that you want to start with an output. You want to start with what is the definition of done? What are you actually looking for? Because let's say you ask your agent here for chicken, for example, and you actually want, like, chicken parmesan on a bed of pasta, but because you just said chicken, the agent might interpret that a little bit differently and make you a chicken sandwich, and then next time you ask for chicken, it might make you, you know, chicken thighs. It doesn't actually know specifically what you want. So, if you start with an output and you have, you know, essentially this chicken parmesan and you say, "Okay, let's reverse engineer this food and see what went into it, how long we cooked it, how did we get here?" And that's how you build the recipe. So, for example, if you wanted to build a skill for the end-of-the-week report, you've got certain columns in your Excel sheet, you've got certain calculations that were made, it's way easier to give the agent that Excel sheet, to give it the final deliverable and say, "Hey, this is an output that is really good, and this is what I want to build a skill for," so that you understand how to take some raw input and turn it into this output that I have already told you that I like. And this is what we're looking for every time, and then it can basically walk you backwards through that process. "Okay, what data did you look at? Where did you get it from? How did you calculate it? Where did you format it?" You answer those questions, and then you have a version of a skill that already knows, sort of like the North Star that it's building towards. I think it's so much easier to run the process, get the output, and say, "Okay, let's turn that into a skill," rather than saying, "Hey, build me a skill for building a YouTube dashboard," and then you might have this vision and your agent might have a completely different vision, and then you're going to get frustrated when it delivers you a chicken sandwich and you actually wanted chicken parm. Now remember, all of these concepts I'm going to, you know, bring back together when we actually hop into Codex and I show you some stuff, but let's just keep moving down the list, because I think this foundation is important. So, number two, we have one specific job and one specific trigger. So, remember how I said that these skills are markdown files, skill.md. What happens in the skill.md is you've got a YAML description—it's called YAML front matter—and then you've got the actual skill instructions, which is the meat of the skill. So, right here is one of the skills that I'm going to keep coming back to in this video because it's—it's one of my favorite skills. And you can see up here this metadata, this is the YAML front matter. So, if I view the source, it looks like this. It's actually just separated by these three dashes. So, this would be the YAML front matter, and then everything below it would be the actual instructions of the skill. Now, this is markdown, and when I said markdown is very simple, it is. You can see here are two pound signs, and that just indicates a header. We can see that we've got these little dashes which are bullet points, and so when markdown is actually rendered, it looks more something like this, and you can see it's just formatted better. So, this is the way that I like to look at it. But if you click on "View Source" inside of the Codex app, you can see the raw markdown file. But anyways, the metadata up here is the YAML front matter, and that tells us what is the name of the skill, and when do you explicitly use it? And there are other fields that can be populated in here, like an argument hint, and basically this x-article skill is used when I ask my agent to turn an YouTube video into a long-form x-post. And so sometimes it'll feed in this argument, which is my YouTube video URL that it will take, it will download, it will transcribe it, it will take screenshots of it, all that sort of, you know, all the stuff. And then once the agent has basically said, "Okay, cool, Nate wants to turn this YouTube video into an x-article. So, okay, cool, I found this skill that I need to use." Now that I know this is the skill I'm going to invoke, let me read the entire description of the skill so now I understand exactly what to actually do. And you can see this skill is pretty long. There's a lot of instructions in here, and this wasn't like me on one try turning this into a skill. This is probably a skill that's been iterated on 25 or more times, and we'll talk more about that in a bit. But the whole idea is that you want a skill to do one very specific job. You don't want to have a skill that's like, "Run the marketing team." You want to have a skill that breaks that marketing team, all of those processes, into individual steps. So, in my book, right up here, becoming AI native, I talk about this and I call it the function breakdown. I call it the tree. Basically, your job is a tree, and you've got different trunks with different, like, bullet points in your job description, and then each of those trunks have branches, and each of those branches have ultimately leaves. And you want to turn all of those little leaves, or those little tasks, into skills, because you can chain skills together over time. But the point being, if you have a skill with one very specific job, you can give it one very specific trigger, which means that your agents are going to be able to execute them more consistently, invoke them automatically. It just helps you separate out what your agents are doing in a much better way than if you have a skill that's like 30 pages long, and it's supposed to do so many different things. So, think about breaking down processes at the task level and turning each of those tasks into skills. All right, moving on to number three, we have thinking about the freedom level. So, what I mean by this is when you're automating things, you basically have to figure out, is this a deterministic automation or a non-deterministic automation? Basically, is this something that's predictable and we know exactly what comes in and we know exactly what happens, and we know exactly what comes out? Or is it more of a non-deterministic AI agent, more of a non-deterministic black box where we know sort of what the input's going to look like, we sort of know what's going to happen inside the process, but we do kind of know like what the output should be, we still know the definition of good and the definition of done? You can think about it like, can I automate this using rules? Hard rules, hard logic. Is X greater than 10? Is X equal to 500? Or is it something that needs a lot of judgement? Analyze this, turn this into an email, turn this into a spreadsheet. And it's really important to think about, where does your skill live on that spectrum? Because if you have a very deterministic skill, like, for example, it just has to take an Excel sheet and it just has to populate the cells into the CRM or something, and that's basically just data processing, transferring. That's a very deterministic process. And if you write the skill in a very non-deterministic way, there's a lot more room for error. There's a lot more room for AI to interpret something wrong and do something wrong. And that's the case where you want the skill to be like, "Step one do this, step two do this," exactly, exactly, exactly, do this. It's very specific. But if you've got something non-deterministic, like, for example, my x-article skill, we don't want to be so specific. We want the AI to be able to use its judgement and to be able to use research and thinking to generate a unique output, because every YouTube video is different, which means every x-article is going to be different, which means all the screenshots that it needs to take are going to be different. I couldn't say something like, "Screenshot the video at the 1-minute mark, and then the 2-minute mark, and then the 5-minute mark, and put those into the article at line 400, line 600, line 700," because then the articles would come out feeling generic and they probably wouldn't even make sense, and it just wouldn't be a very good output. So, think about the freedom level of the process, and when you're doing this manually, do you follow the same set of instructions every time? Or are you constantly using judgement, and are you constantly doing different things within the process to ultimately get to that definition of good. By the way, guys, I've got this completely free SOP for you about getting your first AI automation client. It's going to go over the exact steps that has been proven for hundreds of our AIS+ members to get their first paid gigs. It goes over the one-sentence service pitch that can get you started today, why your first client should cost you money, the 5-minute video that answers, "Can this person actually deliver?" before you've actually received any money, what to do when you have zero case studies. There's so many good things in here that are going to help you out, even if you already do have clients. I would recommend grabbing this because, like I said, it's yours completely free. So, if you want to grab this, there's a link for it down in the description. Let's get back to the video. All right, moving on to number four, we have Verification, and this is probably the most important piece of building skills. So, the whole idea of building a skill is that now you can trust that your agent's going to give you an output more consistently, that meets your standard. But what you want to do in there is you don't want to be constantly being handed the agent's V1. You want to be handed an output that's already good to go, that you can give it a quick skim, approve it all, and then shoot it off. And the cool thing about agent's building stuff to ultimately achieve some sort of North Star is that you can also have agents check its own work, verify its own work, or, you know, other agents verify different agent's work. So every single skill that I build works in some sort of verification loop, meaning the agent who builds the thing delivers this output, and then we either have the same agent or different sub-agents come in and verify that output and then provide feedback. And this verification loop can go on multiple times. Sometimes it's just, "Hey, here's the V2, and it's approved. Here's the V7." I've even had agents verify over and over and give me, like, a V15 or 16. And that basically just ensures that your time isn't being wasted, verifying and sending feedback when you can have agents do that for you. Now when it comes to verification, there is a difference between objective checks and subjective checks. So, let me explain what I mean by that. An objective check is something that can literally be proven, sort of like the non-deterministic versus deterministic thing. Objective means, okay, I wanted you to do research, and then the verification was to pull in 500 sources, cut that down to the best 250, and then every single fact that's in the article, I want that to be double-checked by a second agent. And that's objective. We can literally prove 500 sources cut down to 250, every single fact double-checked by a different agent. That's objective, that is a rule. Does X equal Y? Is A greater than B? But then you have some things that are subjective, like verifying the video, right? If it's—it's creating a video for you or if it's creating a website for you. A lot of these subjective checks are more like, oh, you know, does everything look good? Is everything in bounds? Is it loading in a certain way? When you don't have an actual hard metric to align it to, you have to sort of be a little bit more subjective on your verification. And what you're thinking about doing is turning that agent into an LLM as a judge, which basically means, okay, so if AI is going to be applying its approval, and it has to use judgement to approve, then how can you tell the agent exactly what you're looking for? What does good typically look like or feel like, even though there's not a hard rule? So, for example, with the x-article skill, I've got some objective checks in there, like there are 10 screenshots from the video. But then there's other subjective checks like, how do you make it flow in a way that makes sense? How do you make it sound like Nate? How do you make sure that the images that are put into the article, the screenshots, are cropped? There's other subjective checks that we have to let it do things where it like opens up the app, and it scrolls through, and it reads through, and it screenshots, and it verifies. But a lot of those checks, like I said, are more subjective. And so that's where iterating on these skills takes a lot of time, because you're constantly running the skill, analyzing the output, giving feedback, running the skill again, analyzing the output, giving feedback, and you're basically just like training the skill, where every single time you use it, it gets better. But the thing about verification is there's always a way. You might be thinking to yourself, "Uh, well, I don't really know how I would have agents verify that process." Agents can do anything. They can use your computer, they can use your browser, they can look at things, they can listen to things, they can do anything. So just think about it like this, if you assigned that task to a human, what would you do to basically give it the stamp of approval? And then just explain that to the agent, and that's your verification. That doesn't mean it's going to always come out 100%, but it's going to be more like 95% rather than giving you something that started off at like 75 or 80. Okay, and moving on to number five, we have Walk It Down. And when I say walk it down, I mean walking it down the model list. So, whether you're using Claude Code or Codex, we all know that different models have different strengths, and we all know that models that have the strongest strengths are the most expensive. So, for example, Astra is more expensive than Sol, but more capable than Sol. Sol is more expensive than Terra, but more capable than Terra. So, the idea is you build a skill and you're getting this good output, and maybe you were testing that with Astra. Okay, well, let's run that same skill on Sol. Are we getting the same result? If we're getting the same result and it's cheaper, okay, let's test it on Terra now. Are we getting the same result? Is it cheaper? Okay, let's test it on Luna. There's no reason to be running a skill with Astra if you could run it with Luna and get the exact same output. But it is important to test it because not every skill fits that. I've had some skills where I run them on Haiku or Luna because they're just super simple. For example, when I give my agent a YouTube video and I say, "Hey, I need a YouTube video description and timestamps," that I can do with Luna, and doing it with Astra is overkill. But for this x-article thing, I typically like to use at least Sol or Astra because it just has a better understanding of screenshotting things and putting it in the right spot and cropping it down and even blurring things out. It's so much better with the browser use with Sol and Astra. And then, once you've landed on a model that you like, you can take that one step further and you can walk it down the effort level. So, start it off on maybe High, and if it's not good enough, then move it up a little bit. And if it is good, then move it down to Medium. And just keep finding, basically, the simplest model or the lightest and cheapest model that still executes at the level of quality that you're looking for. And then moving on here to number six, we have the Bike Method, which is I've kind of already alluded to it multiple times throughout this video so far. But really, it's just the idea to me that your skill is never done. Every single time you run the skill, you're going to improve it. You're going to say, "Hey, here's what I really liked, here's what I didn't like." Update the skill so that next time it's better. Almost every single time I run a skill, I give it feedback and I tell it to update. And if you think about this, like you're teaching a kid to ride a bike, that's what a skill really should feel like. You start off very cautious, you're watching everything, you're guiding them, you—you've got your hand on this the the steering wheel, steering wheel, I meant like handlebars, and you're—you're right there. And then what happens is you give feedback, you say like, "Okay, that was good, but you know, you were leaning a bit too far to the left, make sure your weight's more in the center." And then you keep going. Eventually, you take off the training wheels because the skill is getting better and you're getting more trust in the skill, and then you give more feedback again, and then eventually you take off the elbow pads, right? And then you keep giving feedback. But that doesn't mean that you just take off the kid's helmet and let them bike on the highway. You're not going to do something reckless like that. You're also not just going to shoot them off down the road and then go inside and take a nap. You're still going to be in in some way, watching or having certain guard rails in place in order to make you feel comfortable and in order for you to have more trust in the skill itself. And I truly believe in the idea that there's no such thing as a finished product unless your skill is super, super deterministic and it runs perfect every time and you don't have to worry about it. But when the skills are more on the judgement side of the spectrum, you know, as we talked about up here, the more your skills live over here, the less I believe that you can actually have a finished skill, because like I said, every time things might happen a little bit differently, and every time your process might change or new models might drop, and you're just constantly going to be giving feedback every time you use it. Okay, so let's hop into Codex here and we can take a look at how this actually works. Now, when you are inside of your project, and if you're inside of your AI OS, which is usually what I'm building in, the Codex skills will live in a folder called the .agents, and then inside of the .agents, there's a skill—or sorry, not a skill—there's another folder called .skills. And if you're doing this in Cloud Code, the skills live in a .cloud and then in a folder called skills. Now, these can be transferable. I basically just tell Codex or Cloud, "Hey, see all the skills in there and bring them over here, too," so it just I have duplicates of them. But that's where they actually live inside of your project, unless you have them at a global level, which just means they're living locally across all of your Codex projects, rather than just inside of your AI OS. So that's where they live, and you can always just say, "Hey, you know, Codex, I have this skill, I don't remember where it is, could you show me the file path?" And it will find it, and then you can move it if you want. But anyways, that's where they live. And you can see I actually just ran this x-article skill on a video, so it says your x-article draft is saved, I'll click into it, and I actually did obviously already publish this one, but I didn't even have to change anything. I basically just clicked in here, I read it all, and I published it. And you can see that as it's going through, it already formatted everything, and it used the video of course. It took screenshots of things and it like put them in the right spot to line up with what was actually being said in the article, and it even added this little, um, I don't know what you want to call it like a spotlight. It added that. I didn't actually do that, but it will basically crop things out and it will add things like that into my actual article because I've told it in the skill to do things like that. And you can see that for this specific video, this ran on GPT-6 Astra. So what I'm going to do real quick is I'm going to walk this down. I'm going to give this same prompt to Sol and Terra and tell it to make the x-article and it's going to use the skill and we'll see how those outputs compare to Astra's output. Because this is a pretty complex skill, it has to transcribe the video, it has to open up the article, it has to make the thumbnail, it has to format everything, and it has to screenshot everything and put it in and then actually like drag it around. So there's a lot of browser use, and there is a ton of judgement inside of this, like I said this is not a deterministic skill. So you can see right now I have Sol, Terra, and Luna all running. It's been about 10 minutes, so I'll just check with you guys when this is done. But I just wanted to show you how all of them because of the skill, because of the verification are viewing images. You can see how many times Luna has viewed these images here, because it's taking screenshots and it's trying to understand, okay, which of these should be worked into the article and where and why. Same thing down here with Sol, obviously this one is a super, super visual-heavy, sort of skill. But because I've worked that in there, all of them are viewing images no matter what. And so it'll be really interesting to see how they decide to place them and, you know, what images they actually choose to put in there, because X limits you to a certain amount of images per article. So, I'm going to let these run, we'll compare the results, and we'll kind of talk about how walking it down has changed my perspective on this specific skill. All right, so you can see now that we have Luna and Terra have finished up. Sol is still working, so we'll start reviewing these two, and then we'll hop over to Sol. Hopefully it'll be done by the time we're looking through these. So, Luna took 28 minutes and 33 seconds. Let's open up the draft and see what we are working with. So we have, um, the thumbnail in there. We have "I designed a one-person million-dollar business with Claude." Um, the title is the same as like the opening line, which I don't love. There's also this weird spacing was right there. I don't know if it tried to like center it, but I do notice that. Anyways, we have sort of like the TL;DR, we have these different headers here. I'm yet to see an image. Here comes the first screenshot. So this is talking about, um, the three different types of ideas to compare. And in the screenshot we do see all three, so that matches up pretty well. Um, we see Agent Report Card as well as down here. So, really what I'm checking for right now is that it all sounds like me. So this is kind of trained on like my LinkedIn style guide and my YouTube style guide and like the way that I speak, but I'm also making sure that the images that are put into the article, the screenshots, are cropped. There's other subjective checks that we have to let it do things where it like opens up the app, and it scrolls through, and it reads through, and it screenshots, and it verifies, but a lot of those checks, like I said are more subjective. And so that's where iterating on these skills takes a lot of time because you're constantly running the skill, analyzing the output, giving feedback, running the skill again, analyzing the output, giving feedback, and you're basically just like training the skill where every single time you use it, it gets better. But the thing about verification is there's always a way. You might be thinking to yourself, "Uh, well, I don't really know how I would have agents verify that process." Agents can do anything. They can use your computer, they can use your browser, they can look at things, they can listen to things, they can do anything. So just think about it like this, if you assigned that task to a human, what would you do to basically give it the stamp of approval, and then just explain that to the agent, and that's your verification. That doesn't mean it's going to always come out 100%, but it's going to be more like 95% rather than giving you something that started off at like 75 or 80. Okay, and moving on to number five, we have Walk It Down. And when I say walk it down, I mean walking it down the model list. So, whether you're using Claude Code or Codex, we all know that different models have different strengths and we all know that models that have the strongest strengths are the most expensive. So, for example, Astra is more expensive than Sol, but more capable than Sol. Sol is more expensive than Terra, but more capable than Terra. So, the idea is you build a skill and you're getting this good output and maybe you were testing that with Astra. Okay, well, let's run that same skill on Sol. Are we getting the same result? If we're getting the same result and it's cheaper, okay, let's test it on Terra now. Are we getting the same result? Is it cheaper? Okay, let's test it on Luna. There's no reason to be running a skill with Astra if you could run it with Luna and get the exact same output. But it is important to test it because not every skill fits that. I've had some skills where I run them on Haiku or Luna because they're just super simple. For example, when I give my agent a YouTube video and I say, "Hey, I need a YouTube video description and timestamps," that I can do with Luna and doing it with Astra is overkill. But for this x-article thing, I typically like to use at least Sol or Astra because it just has a better understanding of screenshotting things and putting it in the right spot and cropping it down and even blurring things out. It's so much better with the browser use with Sol and Astra. And then, once you've landed on a model that you like, you can take that one step further and you can walk it down the effort level. So, start it off on maybe High, and if it's not good enough, then move it up a little bit. And if it is good, then move it down to Medium. And just keep finding, basically, the simplest model or the lightest and cheapest model that still executes at the level of quality that you're looking for. And then moving on here to number six, we have the Bike Method, which is I've kind of already alluded to it multiple times throughout this video so far. But really, it's just the idea to me that your skill is never done. Every single time you run the skill, you're going to improve it. You're going to say, "Hey, here's what I really liked, here's what I didn't like." Update the skill so that next time it's better. Almost every single time I run a skill, I give it feedback and I tell it to update. And if you think about this, like you're teaching a kid to ride a bike, that's what a skill really should feel like. You start off very cautious, you're watching everything, you're guiding them, you—you've got your hand on this the the steering wheel, steering wheel, I meant like handlebars, and you're—you're right there. And then what happens is you give feedback, you say like, "Okay, that was good, but you know, you were leaning a bit too far to the left, make sure your weight's more in the center." And then you keep going. Eventually, you take off the training wheels because the skill is getting better and you're getting more trust in the skill, and then you give more feedback again, and then eventually you take off the elbow pads, right? And then you keep giving feedback. But that doesn't mean that you just take off the kid's helmet and let them bike on the highway. You're not going to do something reckless like that. You're also not just going to shoot them off down the road and then go inside and take a nap. You're still going to be in in some way, watching or having certain guard rails in place in order to make you feel comfortable and in order for you to have more trust in the skill itself. And I truly believe in the idea that there's no such thing as a finished product unless your skill is super, super deterministic and it runs perfect every time and you don't have to worry about it. But when the skills are more on the judgement side of the spectrum, you know, as we talked about up here, the more your skills live over here, the less I believe that you can actually have a finished skill, because like I said, every time things might happen a little bit differently, and every time your process might change or new models might drop, and you're just constantly going to be giving feedback every time you use it. Okay, so let's hop into Codex here and we can take a look at how this actually works. Now, when you are inside of your project, and if you're inside of your AI OS, which is usually what I'm building in, the Codex skills will live in a folder called the .agents, and then inside of the .agents, there's a skill—or sorry, not a skill—there's another folder called .skills. And if you're doing this in Cloud Code, the skills live in a .cloud and then in a folder called skills. Now, these can be transferable. I basically just tell Codex or Cloud, "Hey, see all the skills in there and bring them over here, too," so it just I have duplicates of them. But that's where they actually live inside of your project, unless you have them at a global level, which just means they're living locally across all of your Codex projects, rather than just inside of your AI OS. So that's where they live, and you can always just say, "Hey, you know, Codex, I have this skill, I don't remember where it is, could you show me the file path?" And it will find it, and then you can move it if you want. But anyways, that's where they live. And you can see I actually just ran this x-article skill on a video, so it says your x-article draft is saved, I'll click into it, and I actually did obviously already publish this one, but I didn't even have to change anything. I basically just clicked in here, I read it all, and I published it. And you can see that as it's going through, it already formatted everything, and it used the video of course. It took screenshots of things and it like put them in the right spot to line up with what was actually being said in the article, and it even added this little, um, I don't know what you want to call it like a spotlight. It added that. I didn't actually do that, but it will basically crop things out and it will add things like that into my actual article because I've told it in the skill to do things like that. And you can see that for this specific video, this ran on GPT-6 Astra. So what I'm going to do real quick is I'm going to walk this down. I'm going to give this same prompt to Sol and Terra and tell it to make the x-article and it's going to use the skill and we'll see how those outputs compare to Astra's output. Because this is a pretty complex skill, it has to transcribe the video, it has to open up the article, it has to make the thumbnail, it has to format everything, and it has to screenshot everything and put it in and then actually like drag it around. So there's a lot of browser use, and there is a ton of judgement inside of this, like I said this is not a deterministic skill. So you can see right now I have Sol, Terra, and Luna all running. It's been about 10 minutes, so I'll just check with you guys when this is done. But I just wanted to show you how all of them because of the skill, because of the verification are viewing images. You can see how many times Luna has viewed these images here, because it's taking screenshots and it's trying to understand, okay, which of these should be worked into the article and where and why. Same thing down here with Sol, obviously this one is a super, super visual-heavy, sort of skill. But because I've worked that in there, all of them are viewing images no matter what. And so it'll be really interesting to see how they decide to place them and, you know, what images they actually choose to put in there, because X limits you to a certain amount of images per article. So, I'm going to let these run, we'll compare the results, and we'll kind of talk about how walking it down has changed my perspective on this specific skill. All right, so you can see now that we have Luna and Terra have finished up. Sol is still working, so we'll start reviewing these two, and then we'll hop over to Sol. Hopefully it'll be done by the time we're looking through these. So, Luna took 28 minutes and 33 seconds. Let's open up the draft and see what we are working with. So we have, um, the thumbnail in there. We have "I designed a one-person million-dollar business with Claude." Um, the title is the same as like the opening line, which I don't love. There's also this weird spacing was right there. I don't know if it tried to like center it, but I do notice that. Anyways, we have sort of like the TL;DR, we have these different headers here. I'm yet to see an image. Here comes the first screenshot. So this is talking about, um, the three different types of ideas to compare. And in the screenshot we do see all three, so that matches up pretty well. Um, we see Agent Report Card as well as down here. So, really what I'm checking for right now is that it all sounds like me. So this is kind of trained on like my LinkedIn style guide and my YouTube style guide and like the way that I speak, but I'm also making sure that the images that are put into the article, the screenshots, are cropped. There's other subjective checks that we have to let it do things where it like opens up the app, and it scrolls through, and it reads through, and it screenshots, and it verifies, but a lot of those checks, like I said are more subjective. And so that's where iterating on these skills takes a lot of time because you're constantly running the skill, analyzing the output, giving feedback, running the skill again, analyzing the output, giving feedback, and you're basically just like training the skill where every single time you use it, it gets better. But the thing about verification is there's always a way. You might be thinking to yourself, "Uh, well, I don't really know how I would have agents verify that process." Agents can do anything. They can use your computer, they can use your browser, they can look at things, they can listen to things, they can do anything. So just think about it like this, if you assigned that task to a human, what would you do to basically give it the stamp of approval, and then just explain that to the agent, and that's your verification. That doesn't mean it's going to always come out 100%, but it's going to be more like 95% rather than giving you something that started off at like 75 or 80. Okay, and moving on to number five, we have Walk It Down. And when I say walk it down, I mean walking it down the model list. So, whether you're using Claude Code or Codex, we all know that different models have different strengths and we all know that models that have the strongest strengths are the most expensive. So, for example, Astra is more expensive than Sol, but more capable than Sol. Sol is more expensive than Terra, but more capable than Terra. So, the idea is you build a skill and you're getting this good output and maybe you were testing that with Astra. Okay, well, let's run that same skill on Sol. Are we getting the same result? If we're getting the same result and it's cheaper, okay, let's test it on Terra now. Are we getting the same result? Is it cheaper? Okay, let's test it on Luna. There's no reason to be running a skill with Astra if you could run it with Luna and get the exact same output. But it is important to test it because not every skill fits that. I've had some skills where I run them on Haiku or Luna because they're just super simple. For example, when I give my agent a YouTube video and I say, "Hey, I need a YouTube video description and timestamps," that I can do with Luna and doing it with Astra is overkill. But for this x-article thing, I typically like to use at least Sol or Astra because it just has a better understanding of screenshotting things and putting it in the right spot and cropping it down and even blurring things out. It's so much better with the browser use with Sol and Astra. And then, once you've landed on a model that you like, you can take that one step further and you can walk it down the effort level. So, start it off on maybe High, and if it's not good enough, then move it up a little bit. And if it is good, then move it down to Medium. And just keep finding, basically, the simplest model or the lightest and cheapest model that still executes at the level of quality that you're looking for. And then moving on here to number six, we have the Bike Method, which is I've kind of already alluded to it multiple times throughout this video so far. But really, it's just the idea to me that your skill is never done. Every single time you run the skill, you're going to improve it. You're going to say, "Hey, here's what I really liked, here's what I didn't like." Update the skill so that next time it's better. Almost every single time I run a skill, I give it feedback and I tell it to update. And if you think about this, like you're teaching a kid to ride a bike, that's what a skill really should feel like. You start off very cautious, you're watching everything, you're guiding them, you—you've got your hand on this the the steering wheel, steering wheel, I meant like handlebars, and you're—you're right there. And then what happens is you give feedback, you say like, "Okay, that was good, but you know, you were leaning a bit too far to the left, make sure your weight's more in the center." And then you keep going. Eventually, you take off the training wheels because the skill is getting better and you're getting more trust in the skill, and then you give more feedback again, and then eventually you take off the elbow pads, right? And then you keep giving feedback. But that doesn't mean that you just take off the kid's helmet and let them bike on the highway. You're not going to do something reckless like that. You're also not just going to shoot them off down the road and then go inside and take a nap. You're still going to be in in some way, watching or having certain guard rails in place in order to make you feel comfortable and in order for you to have more trust in the skill itself. And I truly believe in the idea that there's no such thing as a finished product unless your skill is super, super deterministic and it runs perfect every time and you don't have to worry about it. But when the skills are more on the judgement side of the spectrum, you know, as we talked about up here, the more your skills live over here, the less I believe that you can actually have a finished skill, because like I said, every time things might happen a little bit differently, and every time your process might change or new models might drop, and you're just constantly going to be giving feedback every time you use it. Okay, so let's hop into Codex here and we can take a look at how this actually works. Now, when you are inside of your project, and if you're inside of your AI OS, which is usually what I'm building in, the Codex skills will live in a folder called the .agents, and then inside of the .agents, there's a skill—or sorry, not a skill—there's another folder called .skills. And if you're doing this in Cloud Code, the skills live in a .cloud and then in a folder called skills. Now, these can be transferable. I basically just tell Codex or Cloud, "Hey, see all the skills in there and bring them over here, too," so it just I have duplicates of them. But that's where they actually live inside of your project, unless you have them at a global level, which just means they're living locally across all of your Codex projects, rather than just inside of your AI OS. So that's where they live, and you can always just say, "Hey, you know, Codex, I have this skill, I don't remember where it is, could you show me the file path?" And it will find it, and then you can move it if you want. But anyways, that's where they live. And you can see I actually just ran this x-article skill on a video, so it says your x-article draft is saved, I'll click into it, and I actually did obviously already publish this one, but I didn't even have to change anything. I basically just clicked in here, I read it all, and I published it. And you can see that as it's going through, it already formatted everything, and it used the video of course. It took screenshots of things and it like put them in the right spot to line up with what was actually being said in the article, and it even added this little, um, I don't know what you want to call it like a spotlight. It added that. I didn't actually do that, but it will basically crop things out and it will add things like that into my actual article because I've told it in the skill to do things like that. And you can see that for this specific video, this ran on GPT-6 Astra. So what I'm going to do real quick is I'm going to walk this down. I'm going to give this same prompt to Sol and Terra and tell it to make the x-article and it's going to use the skill and we'll see how those outputs compare to Astra's output. Because this is a pretty complex skill, it has to transcribe the video, it has to open up the article, it has to make the thumbnail, it has to format everything, and it has to screenshot everything and put it in and then actually like drag it around. So there's a lot of browser use, and there is a ton of judgement inside of this, like I said this is not a deterministic skill. So you can see right now I have Sol, Terra, and Luna all running. It's been about 10 minutes, so I'll just check with you guys when this is done. But I just wanted to show you how all of them because of the skill, because of the verification are viewing images. You can see how many times Luna has viewed these images here, because it's taking screenshots and it's trying to understand, okay, which of these should be worked into the article and where and why. Same thing down here with Sol, obviously this one is a super, super visual-heavy, sort of skill. But because I've worked that in there, all of them are viewing images no matter what. And so it'll be really interesting to see how they decide to place them and, you know, what images they actually choose to put in there, because X limits you to a certain amount of images per article. So, I'm going to let these run, we'll compare the results, and we'll kind of talk about how walking it down has changed my perspective on this specific skill. All right, so you can see now that we have Luna and Terra have finished up. Sol is still working, so we'll start reviewing these two, and then we'll hop over to Sol. Hopefully it'll be done by the time we're looking through these. So, Luna took 28 minutes and 33 seconds. Let's open up the draft and see what we are working with. So we have, um, the thumbnail in there. We have "I designed a one-person million-dollar business with Claude." Um, the title is the same as like the opening line, which I don't love. There's also this weird spacing was right there. I don't know if it tried to like center it, but I do notice that. Anyways, we have sort of like the TL;DR, we have these different headers here. I'm yet to see an image. Here comes the first screenshot. So this is talking about, um, the three different types of ideas to compare. And in the screenshot we do see all three, so that matches up pretty well. Um, we see Agent Report Card as well as down here. So, really what I'm checking for right now is that it all sounds like me. So this is kind of trained on like my LinkedIn style guide and my YouTube style guide and like the way that I speak, but I'm also making sure that the images that are put into the article, the screenshots, are cropped. There's other subjective checks that we have to let it do things where it like opens up the app, and it scrolls through, and it reads through, and it screenshots, and it verifies, but a lot of those checks, like I said are more subjective. And so that's where iterating on these skills takes a lot of time because you're constantly running the skill, analyzing the output, giving feedback, running the skill again, analyzing the output, giving feedback, and you're basically just like training the skill where every single time you use it, it gets better. But the thing about verification is there's always a way. You might be thinking to yourself, "Uh, well, I don't really know how I would have agents verify that process." Agents can do anything. They can use your computer, they can use your browser, they can look at things, they can listen to things, they can do anything. So just think about it like this, if you assigned that task to a human, what would you do to basically give it the stamp of approval, and then just explain that to the agent, and that's your verification. That doesn't mean it's going to always come out 100%, but it's going to be more like 95% rather than giving you something that started off at like 75 or 80. Okay, and moving on to number five, we have Walk It Down. And when I say walk it down, I mean walking it down the model list. So, whether you're using Claude Code or Codex, we all know that different models have different strengths and we all know that models that have the strongest strengths are the most expensive. So, for example, Astra is more expensive than Sol, but more capable than Sol. Sol is more expensive than Terra, but more capable than Terra. So, the idea is you build a skill and you're getting this good output and maybe you were testing that with Astra. Okay, well, let's run that same skill on Sol. Are we getting the same result? If we're getting the same result and it's cheaper, okay, let's test it on Terra now. Are we getting the same result? Is it cheaper? Okay, let's test it on Luna. There's no reason to be running a skill with Astra if you could run it with Luna and get the exact same output. But it is important to test it because not every skill fits that. I've had some skills where I run them on Haiku or Luna because they're just super simple. For example, when I give my agent a YouTube video and I say, "Hey, I need a YouTube video description and timestamps," that I can do with Luna and doing it with Astra is overkill. But for this x-article thing, I typically like to use at least Sol or Astra because it just has a better understanding of screenshotting things and putting it in the right spot and cropping it down and even blurring things out. It's so much better with the browser use with Sol and Astra. And then, once you've landed on a model that you like, you can take that one step further and you can walk it down the effort level. So, start it off on maybe High, and if it's not good enough, then move it up a little bit. And if it is good, then move it down to Medium. And just keep finding, basically, the simplest model or the lightest and cheapest model that still executes at the level of quality that you're looking for. And then moving on here to number six, we have the Bike Method, which is I've kind of already alluded to it multiple times throughout this video so far. But really, it's just the idea to me that your skill is never done. Every single time you run the skill, you're going to improve it. You're going to say, "Hey, here's what I really liked, here's what I didn't like." Update the skill so that next time it's better. Almost every single time I run a skill, I give it feedback and I tell it to update. And if you think about this, like you're teaching a kid to ride a bike, that's what a skill really should feel like. You start off very cautious, you're watching everything, you're guiding them, you—you've got your hand on this the the steering wheel, steering wheel, I meant like handlebars, and you're—you're right there. And then what happens is you give feedback, you say like, "Okay, that was good, but you know, you were leaning a bit too far to the left, make sure your weight's more in the center." And then you keep going. Eventually, you take off the training wheels because the skill is getting better and you're getting more trust in the skill, and then you give more feedback again, and then eventually you take off the elbow pads, right? And then you keep giving feedback. But that doesn't mean that you just take off the kid's helmet and let them bike on the highway. You're not going to do something reckless like that. You're also not just going to shoot them off down the road and then go inside and take a nap. You're still going to be in in some way, watching or having certain guard rails in place in order to make you feel comfortable and in order for you to have more trust in the skill itself. And I truly believe in the idea that there's no such thing as a finished product unless your skill is super, super deterministic and it runs perfect every time and you don't have to worry about it. But when the skills are more on the judgement side of the spectrum, you know, as we talked about up here, the more your skills live over here, the less I believe that you can actually have a finished skill, because like I said, every time things might happen a little bit differently, and every time your process might change or new models might drop, and you're just constantly going to be giving feedback every time you use it. Okay, so let's hop into Codex here and we can take a look at how this actually works. Now, when you are inside of your project, and if you're inside of your AI OS, which is usually what I'm building in, the Codex skills will live in a folder called the .agents, and then inside of the .agents, there's a skill—or sorry, not a skill—there's another folder called .skills. And if you're doing this in Cloud Code, the skills live in a .cloud and then in a folder called skills. Now, these can be transferable. I basically just tell Codex or Cloud, "Hey, see all the skills in there and bring them over here, too," so it just I have duplicates of them. But that's where they actually live inside of your project, unless you have them at a global level, which just means they're living locally across all of your Codex projects, rather than just inside of your AI OS. So that's where they live, and you can always just say, "Hey, you know, Codex, I have this skill, I don't remember where it is, could you show me the file path?" And it will find it, and then you can move it if you want. But anyways, that's where they live. And you can see I actually just ran this x-article skill on a video, so it says your x-article draft is saved, I'll click into it, and I actually did obviously already publish this one, but I didn't even have to change anything. I basically just clicked in here, I read it all, and I published it. And you can see that as it's going through, it already formatted everything, and it used the video of course. It took screenshots of things and it like put them in the right spot to line up with what was actually being said in the article, and it even added this little, um, I don't know what you want to call it like a spotlight. It added that. I didn't actually do that, but it will basically crop things out and it will add things like that into my actual article because I've told it in the skill to do things like that. And you can see that for this specific video, this ran on GPT-6 Astra. So what I'm going to do real quick is I'm going to walk this down. I'm going to give this same prompt to Sol and Terra and tell it to make the x-article and it's going to use the skill and we'll see how those outputs compare to Astra's output. Because this is a pretty complex skill, it has to transcribe the video, it has to open up the article, it has to make the thumbnail, it has to format everything, and it has to screenshot everything and put it in and then actually like drag it around. So there's a lot of browser use, and there is a ton of judgement inside of this, like I said this is not a deterministic skill. So you can see right now I have Sol, Terra, and Luna all running. It's been about 10 minutes, so I'll just check with you guys when this is done. But I just wanted to show you how all of them because of the skill, because of the verification are viewing images. You can see how many times Luna has viewed these images here, because it's taking screenshots and it's trying to understand, okay, which of these should be worked into the article and where and why. Same thing down here with Sol, obviously this one is a super, super visual-heavy, sort of skill. But because I've worked that in there, all of them are viewing images no matter what. And so it'll be really interesting to see how they decide to place them and, you know, what images they actually choose to put in there, because X limits you to a certain amount of images per article. So, I'm going to let these run, we'll compare the results, and we'll kind of talk about how walking it down has changed my perspective on this specific skill. All right, so you can see now that we have Luna and Terra have finished up. Sol is still working, so we'll start reviewing these two, and then we'll hop over to Sol. Hopefully it'll be done by the time we're looking through these. So, Luna took 28 minutes and 33 seconds. Let's open up the draft and see what we are working with. So we have, um, the thumbnail in there. We have "I designed a one-person million-dollar business with Claude." Um, the title is the same as like the opening line, which I don't love. There's also this weird spacing was right there. I don't know if it tried to like center it, but I do notice that. Anyways, we have sort of like the TL;DR, we have these different headers here. I'm yet to see an image. Here comes the first screenshot. So this is talking about, um, the three different types of ideas to compare. And in the screenshot we do see all three, so that matches up pretty well. Um, we see Agent Report Card as well as down here. So, really what I'm checking for right now is that it all sounds like me. So this is kind of trained on like my LinkedIn style guide and my YouTube style guide and like the way that I speak, but I'm also making sure that the images that are put into the article, the screenshots, are cropped. There's other subjective checks that we have to let it do things where it like opens up the app, and it scrolls through, and it reads through, and it screenshots, and it verifies, but a lot of those checks, like I said are more subjective. And so that's where iterating on these skills takes a lot of time because you're constantly running the skill, analyzing the output, giving feedback, running the skill again, analyzing the output, giving feedback, and you're basically just like training the skill where every single time you use it, it gets better. But the thing about verification is there's always a way. You might be thinking to yourself, "Uh, well, I don't really know how I would have agents verify that process." Agents can do anything. They can use your computer, they can use your browser, they can look at things, they can listen to things, they can do anything. So just think about it like this, if you assigned that task to a human, what would you do to basically give it the stamp of approval, and then just explain that to the agent, and that's your verification. That doesn't mean it's going to always come out 100%, but it's going to be more like 95% rather than giving you something that started off at like 75 or 80. Okay, and moving on to number five, we have Walk It Down. And when I say walk it down, I mean walking it down the model list. So, whether you're using Claude Code or Codex, we all know that different models have different strengths and we all know that models that have the strongest strengths are the most expensive. So, for example, Astra is more expensive than Sol, but more capable than Sol. Sol is more expensive than Terra, but more capable than Terra. So, the idea is you build a skill and you're getting this good output and maybe you were testing that with Astra. Okay, well, let's run that same skill on Sol. Are we getting the same result? If we're getting the same result and it's cheaper, okay, let's test it on Terra now. Are we getting the same result? Is it cheaper? Okay, let's test it on Luna. There's no reason to be running a skill with Astra if you could run it with Luna and get the exact same output. But it is important to test it because not every skill fits that. I've had some skills where I run them on Haiku or Luna because they're just super simple. For example, when I give my agent a YouTube video and I say, "Hey, I need a YouTube video description and timestamps," that I can do with Luna and doing it with Astra is overkill. But for this x-article thing, I typically like to use at least Sol or Astra because it just has a better understanding of screenshotting things and putting it in the right spot and cropping it down and even blurring things out. It's so much better with the browser use with Sol and Astra. And then, once you've landed on a model that you like, you can take that one step further and you can walk it down the effort level. So, start it off on maybe High, and if it's not good enough, then move it up a little bit. And if it is good, then move it down to Medium. And just keep finding, basically, the simplest model or the lightest and cheapest model that still executes at the level of quality that you're looking for. And then moving on here to number six, we have the Bike Method, which is I've kind of already alluded to it multiple times throughout this video so far. But really, it's just the idea to me that your skill is never done. Every single time you run the skill, you're going to improve it. You're going to say, "Hey, here's what I really liked, here's what I didn't like." Update the skill so that next time it's better. Almost every single time I run a skill, I give it feedback and I tell it to update. And if you think about this, like you're teaching a kid to ride a bike, that's what a skill really should feel like. You start off very cautious, you're watching everything, you're guiding them, you—you've got your hand on this the the steering wheel, steering wheel, I meant like handlebars, and you're—you're right there. And then what happens is you give feedback, you say like, "Okay, that was good, but you know, you were leaning a bit too far to the left, make sure your weight's more in the center." And then you keep going. Eventually, you take off the training wheels because the skill is getting better and you're getting more trust in the skill, and then you give more feedback again, and then eventually you take off the elbow pads, right? And then you keep giving feedback. But that doesn't mean that you just take off the kid's helmet and let them bike on the highway. You're not going to do something reckless like that. You're also not just going to shoot them off down the road and then go inside and take a nap. You're still going to be in in some way, watching or having certain guard rails in place in order to make you feel comfortable and in order for you to have more trust in the skill itself. And I truly believe in the idea that there's no such thing as a finished product unless your skill is super, super deterministic and it runs perfect every time and you don't have to worry about it. But when the skills are more on the judgement side of the spectrum, you know, as we talked about up here, the more your skills live over here, the less I believe that you can actually have a finished skill, because like I said, every time things might happen a little bit differently, and every time your process might change or new models might drop, and you're just constantly going to be giving feedback every time you use it. Okay, so let's hop into Codex here and we can take a look at how this actually works. Now, when you are inside of your project, and if you're inside of your AI OS, which is usually what I'm building in, the Codex skills will live in a folder called the .agents, and then inside of the .agents, there's a skill—or sorry, not a skill—there's another folder called .skills. And if you're doing this in Cloud Code, the skills live in a .cloud and then in a folder called skills. Now, these can be transferable. I basically just tell Codex or Cloud, "Hey, see all the skills in there and bring them over here, too," so it just I have duplicates of them. But that's where they actually live inside of your project, unless you have them at a global level, which just means they're living locally across all of your Codex projects, rather than just inside of your AI OS. So that's where they live, and you can always just say, "Hey, you know, Codex, I have this skill, I don't remember where it is, could you show me the file path?" And it will find it, and then you can move it if you want. But anyways, that's where they live. And you can see I actually just ran this x-article skill on a video, so it says your x-article draft is saved, I'll click into it, and I actually did obviously already publish this one, but I didn't even have to change anything. I basically just clicked in here, I read it all, and I published it. And you can see that as it's going through, it already formatted everything, and it used the video of course. It took screenshots of things and it like put them in the right spot to line up with what was actually being said in the article, and it even added this little, um, I don't know what you want to call it like a spotlight. It added that. I didn't actually do that, but it will basically crop things out and it will add things like that into my actual article because I've told it in the skill to do things like that. And you can see that for this specific video, this ran on GPT-6 Astra. So what I'm going to do real quick is I'm going to walk this down. I'm going to give this same prompt to Sol and Terra and tell it to make the x-article and it's going to use the skill and we'll see how those outputs compare to Astra's output. Because this is a pretty complex skill, it has to transcribe the video, it has to open up the article, it has to make the thumbnail, it has to format everything, and it has to screenshot everything and put it in and then actually like drag it around. So there's a lot of browser use, and there is a ton of judgement inside of this, like I said this is not a deterministic skill. So you can see right now I have Sol, Terra, and Luna all running. It's been about 10 minutes, so I'll just check with you guys when this is done. But I just wanted to show you how all of them because of the skill, because of the verification are viewing images. You can see how many times Luna has viewed these images here, because it's taking screenshots and it's trying to understand, okay, which of these should be worked into the article and where and why. Same thing down here with Sol, obviously this one is a super, super visual-heavy, sort of skill. But because I've worked that in there, all of them are viewing images no matter what. And so it'll be really interesting to see how they decide to place them and, you know, what images they actually choose to put in there, because X limits you to a certain amount of images per article. So, I'm going to let these run, we'll compare the results, and we'll kind of talk about how walking it down has changed my perspective on this specific skill. All right, so you can see now that we have Luna and Terra have finished up. Sol is still working, so we'll start reviewing these two, and then we'll hop over to Sol. Hopefully it'll be done by the time we're looking through these. So, Luna took 28 minutes and 33 seconds. Let's open up the draft and see what we are working with. So we have, um, the thumbnail in there. We have "I designed a one-person million-dollar business with Claude." Um, the title is the same as like the opening line, which I don't love. There's also this weird spacing was right there. I don't know if it tried to like center it, but I do notice that. Anyways, we have sort of like the TL;DR, we have these different headers here. I'm yet to see an image. Here comes the first screenshot. So this is talking about, um, the three different types of ideas to compare. And in the screenshot we do see all three, so that matches up pretty well. Um, we see Agent Report Card as well as down here. So, really what I'm checking for right now is that it all sounds like me. So this is kind of trained on like my LinkedIn style guide and my YouTube style guide and like the way that I speak, but I'm also making sure that the images that are put into the article, the screenshots, are cropped. There's other subjective checks that we have to let it do things where it like opens up the app, and it scrolls through, and it reads through, and it screenshots, and it verifies, but a lot of those checks, like I said are more subjective. And so that's where iterating on these skills takes a lot of time because you're constantly running the skill, analyzing the output, giving feedback, running the skill again, analyzing the output, giving feedback, and you're basically just like training the skill where every single time you use it, it gets better. But the thing about verification is there's always a way. You might be thinking to yourself, "Uh, well, I don't really know how I would have agents verify that process." Agents can do anything. They can use your computer, they can use your browser, they can look at things, they can listen to things, they can do anything. So just think about it like this, if you assigned that task to a human, what would you do to basically give it the stamp of approval, and then just explain that to the agent, and that's your verification. That doesn't mean it's going to always come out 100%, but it's going to be more like 95% rather than giving you something that started off at like 75 or 80. Okay, and moving on to number five, we have Walk It Down. And when I say walk it down, I mean walking it down the model list. So, whether you're using Claude Code or Codex, we all know that different models have different strengths and we all know that models that have the strongest strengths are the most expensive. So, for example, Astra is more expensive than Sol, but more capable than Sol. Sol is more expensive than Terra, but more capable than Terra. So, the idea is you build a skill and you're getting this good output and maybe you were testing that with Astra. Okay, well, let's run that same skill on Sol. Are we getting the same result? If we're getting the same result and it's cheaper, okay, let's test it on Terra now. Are we getting the same result? Is it cheaper? Okay, let's test it on Luna. There's no reason to be running a skill with Astra if you could run it with Luna and get the exact same output. But it is important to test it because not every skill fits that. I've had some skills where I run them on Haiku or Luna because they're just super simple. For example, when I give my agent a YouTube video and I say, "Hey, I need a YouTube video description and timestamps," that I can do with Luna and doing it with Astra is overkill. But for this x-article thing, I typically like to use at least Sol or Astra because it just has a better understanding of screenshotting things and putting it in the right spot and cropping it down and even blurring things out. It's so much better with the browser use with Sol and Astra. And then, once you've landed on a model that you like, you can take that one step further and you can walk it down the effort level. So, start it off on maybe High, and if it's not good enough, then move it up a little bit. And if it is good, then move it down to Medium. And just keep finding, basically, the simplest model or the lightest and cheapest model that still executes at the level of quality that you're looking for. And then moving on here to number six, we have the Bike Method, which is I've kind of already alluded to it multiple times throughout this video so far. But really, it's just the idea to me that your skill is never done. Every single time you run the skill, you're going to improve it. You're going to say, "Hey, here's what I really liked, here's what I didn't like." Update the skill so that next time it's better. Almost every single time I run a skill, I give it feedback and I tell it to update. And if you think about this, like you're teaching a kid to ride a bike, that's what a skill really should feel like. You start off very cautious, you're watching everything, you're guiding them, you—you've got your hand on this the the steering wheel, steering wheel, I meant like handlebars, and you're—you're right there. And then what happens is you give feedback, you say like, "Okay, that was good, but you know, you were leaning a bit too far to the left, make sure your weight's more in the center." And then you keep going. Eventually, you take off the training wheels because the skill is getting better and you're getting more trust in the skill, and then you give more feedback again, and then eventually you take off the elbow pads, right? And then you keep giving feedback. But that doesn't mean that you just take off the kid's helmet and let them bike on the highway. You're not going to do something reckless like that. You're also not just going to shoot them off down the road and then go inside and take a nap. You're still going to be in in some way, watching or having certain guard rails in place in order to make you feel comfortable and in order for you to have more trust in the skill itself. And I truly believe in the idea that there's no such thing as a finished product unless your skill is super, super deterministic and it runs perfect every time and you don't have to worry about it. But when the skills are more on the judgement side of the spectrum, you know, as we talked about up here, the more your skills live over here, the less I believe that you can actually have a finished skill, because like I said, every time things might happen a little bit differently, and every time your process might change or new models might drop, and you're just constantly going to be giving feedback every time you use it. Okay, so let's hop into Codex here and we can take a look at how this actually works. Now, when you are inside of your project, and if you're inside of your AI OS, which is usually what I'm building in, the Codex skills will live in a folder called the .agents, and then inside of the .agents, there's a skill—or sorry, not a skill—there's another folder called .skills. And if you're doing this in Cloud Code, the skills live in a .cloud and then in a folder called skills. Now, these can be transferable. I basically just tell Codex or Cloud, "Hey, see all the skills in there and bring them over here, too," so it just I have duplicates of them. But that's where they actually live inside of your project, unless you have them at a global level, which just means they're living locally across all of your Codex projects, rather than just inside of your AI OS. So that's where they live, and you can always just say, "Hey, you know, Codex, I have this skill, I don't remember where it is, could you show me the file path?" And it will find it, and then you can move it if you want. But anyways, that's where they live. And you can see I actually just ran this x-article skill on a video, so it says your x-article draft is saved, I'll click into it, and I actually did obviously already publish this one, but I didn't even have to change anything. I basically just clicked in here, I read it all, and I published it. And you can see that as it's going through, it already formatted everything, and it used the video of course. It took screenshots of things and it like put them in the right spot to line up with what was actually being said in the article, and it even added this little, um, I don't know what you want to call it like a spotlight. It added that. I didn't actually do that, but it will basically crop things out and it will add things like that into my actual article because I've told it in the skill to do things like that. And you can see that for this specific video, this ran on GPT-6 Astra. So what I'm going to do real quick is I'm going to walk this down. I'm going to give this same prompt to Sol and Terra and tell it to make the x-article and it's going to use the skill and we'll see how those outputs compare to Astra's output. Because this is a pretty complex skill, it has to transcribe the video, it has to open up the article, it has to make the thumbnail, it has to format everything, and it has to screenshot everything and put it in and then actually like drag it around. So there's a lot of browser use, and there is a ton of judgement inside of this, like I said this is not a deterministic skill. So you can see right now I have Sol, Terra, and Luna all running. It's been about 10 minutes, so I'll just check with you guys when this is done. But I just wanted to show you how all of them because of the skill, because of the verification are viewing images. You can see how many times Luna has viewed these images here, because it's taking screenshots and it's trying to understand, okay, which of these should be worked into the article and where and why. Same thing down here with Sol, obviously this one is a super, super visual-heavy, sort of skill. But because I've worked that in there, all of them are viewing images no matter what. And so it'll be really interesting to see how they decide to place them and, you know, what images they actually choose to put in there, because X limits you to a certain amount of images per article. So, I'm going to let these run, we'll compare the results, and we'll kind of talk about how walking it down has changed my perspective on this specific skill. All right, so you can see now that we have Luna and Terra have finished up. Sol is still working, so we'll start reviewing these two, and then we'll hop over to Sol. Hopefully it'll be done by the time we're looking through these. So, Luna took 28 minutes and 33 seconds. Let's open up the draft and see what we are working with. So we have, um, the thumbnail in there. We have "I designed a one-person million-dollar business with Claude." Um, the title is the same as like the opening line, which I don't love. There's also this weird spacing was right there. I don't know if it tried to like center it, but I do notice that. Anyways, we have sort of like the TL;DR, we have these different headers here. I'm yet to see an image. Here comes the first screenshot. So this is talking about, um, the three different types of ideas to compare. And in the screenshot we do see all three, so that matches up pretty well. Um, we see Agent Report Card as well as down here. So, really what I'm checking for right now is that it all sounds like me. So this is kind of trained on like my LinkedIn style guide and my YouTube style guide and like the way that I speak, but I'm also making sure that the images that are put into the article, the screenshots, are cropped. There's other subjective checks that we have to let it do things where it like opens up the app, and it scrolls through, and it reads through, and it screenshots, and it verifies, but a lot of those checks, like I said are more subjective. And so that's where iterating on these skills takes a lot of time because you're constantly running the skill, analyzing the output, giving feedback, running the skill again, analyzing the output, giving feedback, and you're basically just like training the skill where every single time you use it, it gets better. But the thing about verification is there's always a way. You might be thinking to yourself, "Uh, well, I don't really know how I would have agents verify that process." Agents can do anything. They can use your computer, they can use your browser, they can look at things, they can listen to things, they can do anything. So just think about it like this, if you assigned that task to a human, what would you do to basically give it the stamp of approval, and then just explain that to the agent, and that's your verification. That doesn't mean it's going to always come out 100%, but it's going to be more like 95% rather than giving you something that started off at like 75 or 80. Okay, and moving on to number five, we have Walk It Down. And when I say walk it down, I mean walking it down the model list. So, whether you're using Claude Code or Codex, we all know that different models have different strengths and we all know that models that have the strongest strengths are the most expensive. So, for example, Astra is more expensive than Sol, but more capable than Sol. Sol is more expensive than Terra, but more capable than Terra. So, the idea is you build a skill and you're getting this good output and maybe you were testing that with Astra. Okay, well, let's run that same skill on Sol. Are we getting the same result? If we're getting the same result and it's cheaper, okay, let's test it on Terra now. Are we getting the same result? Is it cheaper? Okay, let's test it on Luna. There's no reason to be running a skill with Astra if you could run it with Luna and get the exact same output. But it is important to test it because not every skill fits that. I've had some skills where I run them on Haiku or Luna because they're just super simple. For example, when I give my agent a YouTube video and I say, "Hey, I need a YouTube video description and timestamps," that I can do with Luna and doing it with Astra is overkill. But for this x-article thing, I typically like to use at least Sol or Astra because it just has a better understanding of screenshotting things and putting it in the right spot and cropping it down and even blurring things out. It's so much better with the browser use with Sol and Astra. And then, once you've landed on a model that you like, you can take that one step further and you can walk it down the effort level. So, start it off on maybe High, and if it's not good enough, then move it up a little bit. And if it is good, then move it down to Medium. And just keep finding, basically, the simplest model or the lightest and cheapest model that still executes at the level of quality that you're looking for. And then moving on here to number six, we have the Bike Method, which is I've kind of already alluded to it multiple times throughout this video so far. But really, it's just the idea to me that your skill is never done. Every single time you run the skill, you're going to improve it. You're going to say, "Hey, here's what I really liked, here's what I didn't like." Update the skill so that next time it's better. Almost every single time I run a skill, I give it feedback and I tell it to update. And if you think about this, like you're teaching a kid to ride a bike, that's what a skill really should feel like. You start off very cautious, you're watching everything, you're guiding them, you—you've got your hand on this the the steering wheel, steering wheel, I meant like handlebars, and you're—you're right there. And then what happens is you give feedback, you say like, "Okay, that was good, but you know, you were leaning a bit too far to the left, make sure your weight's more in the center." And then you keep going. Eventually, you take off the training wheels because the skill is getting better and you're getting more trust in the skill, and then you give more feedback again, and then eventually you take off the elbow pads, right? And then you keep giving feedback. But that doesn't mean that you just take off the kid's helmet and let them bike on the highway. You're not going to do something reckless like that. You're also not just going to shoot them off down the road and then go inside and take a nap. You're still going to be in in some way, watching or having certain guard rails in place in order to make you feel comfortable and in order for you to have more trust in the skill itself. And I truly believe in the idea that there's no such thing as a finished product unless your skill is super, super deterministic and it runs perfect every time and you don't have to worry about it. But when the skills are more on the judgement side of the spectrum, you know, as we talked about up here, the more your skills live over here, the less I believe that you can actually have a finished skill, because like I said, every time things might happen a little bit differently, and every time your process might change or new models might drop, and you're just constantly going to be giving feedback every time you use it. Okay, so let's hop into Codex here and we can take a look at how this actually works. Now, when you are inside of your project, and if you're inside of your AI OS, which is usually what I'm building in, the Codex skills will live in a folder called the .agents, and then inside of the .agents, there's a skill—or sorry, not a skill—there's another folder called .skills. And if you're doing this in Cloud Code, the skills live in a .cloud and then in a folder called skills. Now, these can be transferable. I basically just tell Codex or Cloud, "Hey, see all the skills in there and bring them over here, too," so it just I have duplicates of them. But that's where they actually live inside of your project, unless you have them at a global level, which just means they're living locally across all of your Codex projects, rather than just inside of your AI OS. So that's where they live, and you can always just say, "Hey, you know, Codex, I have this skill, I don't remember where it is, could you show me the file path?" And it will find it, and then you can move it if you want. But anyways, that's where they live. And you can see I actually just ran this x-article skill on a video, so it says your x-article draft is saved, I'll click into it, and I actually did obviously already publish this one, but I didn't even have to change anything. I basically just clicked in here, I read it all, and I published it. And you can see that as it's going through, it already formatted everything, and it used the video of course. It took screenshots of things and it like put them in the right spot to line up with what was actually being said in the article, and it even added this little, um, I don't know what you want to call it like a spotlight. It added that. I didn't actually do that, but it will basically crop things out and it will add things like that into my actual article because I've told it in the skill to do things like that. And you can see that for this specific video, this ran on GPT-6 Astra. So what I'm going to do real quick is I'm going to walk this down. I'm going to give this same prompt to Sol and Terra and tell it to make the x-article and it's going to use the skill and we'll see how those outputs compare to Astra's output. Because this is a pretty complex skill, it has to transcribe the video, it has to open up the article, it has to make the thumbnail, it has to format everything, and it has to screenshot everything and put it in and then actually like drag it around. So there's a lot of browser use, and there is a ton of judgement inside of this, like I said this is not a deterministic skill. So you can see right now I have Sol, Terra, and Luna all running. It's been about 10 minutes, so I'll just check with you guys when this is done. But I just wanted to show you how all of them because of the skill, because of the verification are viewing images. You can see how many times Luna has viewed these images here, because it's taking screenshots and it's trying to understand, okay, which of these should be worked into the article and where and why. Same thing down here with Sol, obviously this one is a super, super visual-heavy, sort of skill. But because I've worked that in there, all of them are viewing images no matter what. And so it'll be really interesting to see how they decide to place them and, you know, what images they actually choose to put in there, because X limits you to a certain amount of images per article. So, I'm going to let these run, we'll compare the results, and we'll kind of talk about how walking it down has changed my perspective on this specific skill. All right, so you can see now that we have Luna and Terra have finished up. Sol is still working, so we'll start reviewing these two, and then we'll hop over to Sol. Hopefully it'll be done by the time we're looking through these. So, Luna took 28 minutes and 33 seconds. Let's open up the draft and see what we are working with. So we have, um, the thumbnail in there. We have "I designed a one-person million-dollar business with Claude." Um, the title is the same as like the opening line, which I don't love. There's also this weird spacing was right there. I don't know if it tried to like center it, but I do notice that. Anyways, we have sort of like the TL;DR, we have these different headers here. I'm yet to see an image. Here comes the first screenshot. So this is talking about, um, the three different types of ideas to compare. And in the screenshot we do see all three, so that matches up pretty well. Um, we see Agent Report Card as well as down here. So, really what I'm checking for right now is that it all sounds like me. So this is kind of trained on like my LinkedIn style guide and my YouTube style guide and like the way that I speak, but I'm also making sure that the images that are put into the article, the screenshots, are cropped. There's other subjective checks that we have to let it do things where it like opens up the app, and it scrolls through, and it reads through, and it screenshots, and it verifies, but a lot of those checks, like I said are more subjective. And so that's where iterating on these skills takes a lot of time because you're constantly running the skill, analyzing the output, giving feedback, running the skill again, analyzing the output, giving feedback, and you're basically just like training the skill where every single time you use it, it gets better. But the thing about verification is there's always a way. You might be thinking to yourself, "Uh, well, I don't really know how I would have agents verify that process." Agents can do anything. They can use your computer, they can use your browser, they can look at things, they can listen to things, they can do anything. So just think about it like this, if you assigned that task to a human, what would you do to basically give it the stamp of approval, and then just explain that to the agent, and that's your verification. That doesn't mean it's going to always come out 100%, but it's going to be more like 95% rather than giving you something that started off at like 75 or 80. Okay, and moving on to number five, we have Walk It Down. And when I say walk it down, I mean walking it down the model list. So, whether you're using Claude Code or Codex, we all know that different models have different strengths and we all know that models that have the strongest strengths are the most expensive. So, for example, Astra is more expensive than Sol, but more capable than Sol. Sol is more expensive than Terra, but more capable than Terra. So, the idea is you build a skill and you're getting this good output and maybe you were testing that with Astra. Okay, well, let's run that same skill on Sol. Are we getting the same result? If we're getting the same result and it's cheaper, okay, let's test it on Terra now. Are we getting the same result? Is it cheaper? Okay, let's test it on Luna. There's no reason to be running a skill with Astra if you could run it with Luna and get the exact same output. But it is important to test it because not every skill fits that. I've had some skills where I run them on Haiku or Luna because they're just super simple. For example, when I give my agent a YouTube video and I say, "Hey, I need a YouTube video description and timestamps," that I can do with Luna and doing it with Astra is overkill. But for this x-article thing, I typically like to use at least Sol or Astra because it just has a better understanding of screenshotting things and putting it in the right spot and cropping it down and even blurring things out. It's so much better with the browser use with Sol and Astra. And then, once you've landed on a model that you like, you can take that one step further and you can walk it down the effort level. So, start it off on maybe High, and if it's not good enough, then move it up a little bit. And if it is good, then move it down to Medium. And just keep finding, basically, the simplest model or the lightest and cheapest model that still executes at the level of quality that you're looking for. And then moving on here to number six, we have the Bike Method, which is I've kind of already alluded to it multiple times throughout this video so far. But really, it's just the idea to me that your skill is never done. Every single time you run the skill, you're going to improve it. You're going to say, "Hey, here's what I really liked, here's what I didn't like." Update the skill so that next time it's better. Almost every single time I run a skill, I give it feedback and I tell it to update. And if you think about this, like you're teaching a kid to ride a bike, that's what a skill really should feel like. You start off very cautious, you're watching everything, you're guiding them, you—you've got your hand on this the the steering wheel, steering wheel, I meant like handlebars, and you're—you're right there. And then what happens is you give feedback, you say like, "Okay, that was good, but you know, you were leaning a bit too far to the left, make sure your weight's more in the center." And then you keep going. Eventually, you take off the training wheels because the skill is getting better and you're getting more trust in the skill, and then you give more feedback again, and then eventually you take off the elbow pads, right? And then you keep giving feedback. But that doesn't mean that you just take off the kid's helmet and let them bike on the highway. You're not going to do something reckless like that. You're also not just going to shoot them off down the road and then go inside and take a nap. You're still going to be in in some way, watching or having certain guard rails in place in order to make you feel comfortable and in order for you to have more trust in the skill itself. And I truly believe in the idea that there's no such thing as a finished product unless your skill is super, super deterministic and it runs perfect every time and you don't have to worry about it. But when the skills are more on the judgement side of the spectrum, you know, as we talked about up here, the more your skills live over here, the less I believe that you can actually have a finished skill, because like I said, every time things might happen a little bit differently, and every time your process might change or new models might drop, and you're just constantly going to be giving feedback every time you use it. Okay, so let's hop into Codex here and we can take a look at how this actually works. Now, when you are inside of your project, and if you're inside of your AI OS, which is usually what I'm building in, the Codex skills will live in a folder called the .agents, and then inside of the .agents, there's a skill—or sorry, not a skill—there's another folder called .skills. And if you're doing this in Cloud Code, the skills live in a .cloud and then in a folder called skills. Now, these can be transferable. I basically just tell Codex or Cloud, "Hey, see all the skills in there and bring them over here, too," so it just I have duplicates of them. But that's where they actually live inside of your project, unless you have them at a global level, which just means they're living locally across all of your Codex projects, rather than just inside of your AI OS. So that's where they live, and you can always just say, "Hey, you know, Codex, I have this skill, I don't remember where it is, could you show me the file path?" And it will find it, and then you can move it if you want. But anyways, that's where they live. And you can see I actually just ran this x-article skill on a video, so it says your x-article draft is saved, I'll click into it, and I actually did obviously already publish this one, but I didn't even have to change anything. I basically just clicked in here, I read it all, and I published it. And you can see that as it's going through, it already formatted everything, and it used the video of course. It took screenshots of things and it like put them in the right spot to line up with what was actually being said in the article, and it even added this little, um, I don't know what you want to call it like a spotlight. It added that. I didn't actually do that, but it will basically crop things out and it will add things like that into my actual article because I've told it in the skill to do things like that. And you can see that for this specific video, this ran on GPT-6 Astra. So what I'm going to do real quick is I'm going to walk this down. I'm going to give this same prompt to Sol and Terra and tell it to make the x-article and it's going to use the skill and we'll see how those outputs compare to Astra's output. Because this is a pretty complex skill, it has to transcribe the video, it has to open up the article, it has to make the thumbnail, it has to format everything, and it has to screenshot everything and put it in and then actually like drag it around. So there's a lot of browser use, and there is a ton of judgement inside of this, like I said this is not a deterministic skill. So you can see right now I have Sol, Terra, and Luna all running. It's been about 10 minutes, so I'll just check with you guys when this is done. But I just wanted to show you how all of them because of the skill, because of the verification are viewing images. You can see how many times Luna has viewed these images here, because it's taking screenshots and it's trying to understand, okay, which of these should be worked into the article and where and why. Same thing down here with Sol, obviously this one is a super, super visual-heavy, sort of skill. But because I've worked that in there, all of them are viewing images no matter what. And so it'll be really interesting to see how they decide to place them and, you know, what images they actually choose to put in there, because X limits you to a certain amount of images per article. So, I'm going to let these run, we'll compare the results, and we'll kind of talk about how walking it down has changed my perspective on this specific skill. All right, so you can see now that we have Luna and Terra have finished up. Sol is still working, so we'll start reviewing these two, and then we'll hop over to Sol. Hopefully it'll be done by the time we're looking through these. So, Luna took 28 minutes and 33 seconds. Let's open up the draft and see what we are working with. So we have, um, the thumbnail in there. We have "I designed a one-person million-dollar business with Claude." Um, the title is the same as like the opening line, which I don't love. There's also this weird spacing was right there. I don't know if it tried to like center it, but I do notice that. Anyways, we have sort of like the TL;DR, we have these different headers here. I'm yet to see an image. Here comes the first screenshot. So this is talking about, um, the three different types of ideas to compare. And in the screenshot we do see all three, so that matches up pretty well. Um, we see Agent Report Card as well as down here. So, really what I'm checking for right now is that it all sounds like me. So this is kind of trained on like my LinkedIn style guide and my YouTube style guide and like the way that I speak, but I'm also making sure that the images that are put into the article, the screenshots, are cropped. There's other subjective checks that we have to let it do things where it like opens up the app, and it scrolls through, and it reads through, and it screenshots, and it verifies, but a lot of those checks, like I said are more subjective. And so that's where iterating on these skills takes a lot of time because you're constantly running the skill, analyzing the output, giving feedback, running the skill again, analyzing the output, giving feedback, and you're basically just like training the skill where every single time you use it, it gets better. But the thing about verification is there's always a way. You might be thinking to yourself, "Uh, well, I don't really know how I would have agents verify that process." Agents can do anything. They can use your computer, they can use your browser, they can look at things, they can listen to things, they can do anything. So just think about it like this, if you assigned that task to a human, what would you do to basically give it the stamp of approval, and then just explain that to the agent, and that's your verification. That doesn't mean it's going to always come out 100%, but it's going to be more like 95% rather than giving you something that started off at like 75 or 80. Okay, and moving on to number five, we have Walk It Down. And when I say walk it down, I mean walking it down the model list. So, whether you're using Claude Code or Codex, we all know that different models have different strengths and we all know that models that have the strongest strengths are the most expensive. So, for example, Astra is more expensive than Sol, but more capable than Sol. Sol is more expensive than Terra, but more capable than Terra. So, the idea is you build a skill and you're getting this good output and maybe you were testing that with Astra. Okay, well, let's run that same skill on Sol. Are we getting the same result? If we're getting the same result and it's cheaper, okay, let's test it on Terra now. Are we getting the same result? Is it cheaper? Okay, let's test it on Luna. There's no reason to be running a skill with Astra if you could run it with Luna and get the exact same output. But it is important to test it because not every skill fits that. I've had some skills where I run them on Haiku or Luna because they're just super simple. For example, when I give my agent a YouTube video and I say, "Hey, I need a YouTube video description and timestamps," that I can do with Luna and doing it with Astra is overkill. But for this x-article thing, I typically like to use at least Sol or Astra because it just has a better understanding of screenshotting things and putting it in the right spot and cropping it down and even blurring things out. It's so much better with the browser use with Sol and Astra. And then, once you've landed on a model that you like, you can take that one step further and you can walk it down the effort level. So, start it off on maybe High, and if it's not good enough, then move it up a little bit. And if it is good, then move it down to Medium. And just keep finding, basically, the simplest model or the lightest and cheapest model that still executes at the level of quality that you're looking for. And then moving on here to number six, we have the Bike Method, which is I've kind of already alluded to it multiple times throughout this video so far. But really, it's just the idea to me that your skill is never done. Every single time you run the skill, you're going to improve it. You're going to say, "Hey, here's what I really liked, here's what I didn't like." Update the skill so that next time it's better. Almost every single time I run a skill, I give it feedback and I tell it to update. And if you think about this, like you're teaching a kid to ride a bike, that's what a skill really should feel like. You start off very cautious, you're watching everything, you're guiding them, you—you've got your hand on this the the steering wheel, steering wheel, I meant like handlebars, and you're—you're right there. And then what happens is you give feedback, you say like, "Okay, that was good, but you know, you were leaning a bit too far to the left, make sure your weight's more in the center." And then you keep going. Eventually, you take off the training wheels because the skill is getting better and you're getting more trust in the skill, and then you give more feedback again, and then eventually you take off the elbow pads, right? And then you keep giving feedback. But that doesn't mean that you just take off the kid's helmet and let them bike on the highway. You're not going to do something reckless like that. You're also not just going to shoot them off down the road and then go inside and take a nap. You're still going to be in in some way, watching or having certain guard rails in place in order to make you feel comfortable and in order for you to have more trust in the skill itself. And I truly believe in the idea that there's no such thing as a finished product unless your skill is super, super deterministic and it runs perfect every time and you don't have to worry about it. But when the skills are more on the judgement side of the spectrum, you know, as we talked about up here, the more your skills live over here, the less I believe that you can actually have a finished skill, because like I said, every time things might happen a little bit differently, and every time your process might change or new models might drop, and you're just constantly going to be giving feedback every time you use it. Okay, so let's hop into Codex here and we can take a look at how this actually works. Now, when you are inside of your project, and if you're inside of your AI OS, which is usually what I'm building in, the Codex skills will live in a folder called the .agents, and then inside of the .agents, there's a skill—or sorry, not a skill—there's another folder called .skills. And if you're doing this in Cloud Code, the skills live in a .cloud and then in a folder called skills. Now, these can be transferable. I basically just tell Codex or Cloud, "Hey, see all the skills in there and bring them over here, too," so it just I have duplicates of them. But that's where they actually live inside of your project, unless you have them at a global level, which just means they're living locally across all of your Codex projects, rather than just inside of your AI OS. So that's where they live, and you can always just say, "Hey, you know, Codex, I have this skill, I don't remember where it is, could you show me the file path?" And it will find it, and then you can move it if you want. But anyways, that's where they live. And you can see I actually just ran this x-article skill on a video, so it says your x-article draft is saved, I'll click into it, and I actually did obviously already publish this one, but I didn't even have to change anything. I basically just clicked in here, I read it all, and I published it. And you can see that as it's going through, it already formatted everything, and it used the video of course. It took screenshots of things and it like put them in the right spot to line up with what was actually being said in the article, and it even added this little, um, I don't know what you want to call it like a spotlight. It added that. I didn't actually do that, but it will basically crop things out and it will add things like that into my actual article because I've told it in the skill to do things like that. And you can see that for this specific video, this ran on GPT-6 Astra. So what I'm going to do real quick is I'm going to walk this down. I'm going to give this same prompt to Sol and Terra and tell it to make the x-article and it's going to use the skill and we'll see how those outputs compare to Astra's output. Because this is a pretty complex skill, it has to transcribe the video, it has to open up the article, it has to make the thumbnail, it has to format everything, and it has to screenshot everything and put it in and then actually like drag it around. So there's a lot of browser use, and there is a ton of judgement inside of this, like I said this is not a deterministic skill. So you can see right now I have Sol, Terra, and Luna all running. It's been about 10 minutes, so I'll just check with you guys when this is done. But I just wanted to show you how all of them because of the skill, because of the verification are viewing images. You can see how many times Luna has viewed these images here, because it's taking screenshots and it's trying to understand, okay, which of these should be worked into the article and where and why. Same thing down here with Sol, obviously this one is a super, super visual-heavy, sort of skill. But because I've worked that in there, all of them are viewing images no matter what. And so it'll be really interesting to see how they decide to place them and, you know, what images they actually choose to put in there, because X limits you to a certain amount of images per article. So, I'm going to let these run, we'll compare the results, and we'll kind of talk about how walking it down has changed my perspective on this specific skill. All right, so you can see now that we have Luna and Terra have finished up. Sol is still working, so we'll start reviewing these two, and then we'll hop over to Sol. Hopefully it'll be done by the time we're looking through these. So, Luna took 28 minutes and 33 seconds. Let's open up the draft and see what we are working with. So we have, um, the thumbnail in there. We have "I designed a one-person million-dollar business with Claude." Um, the title is the same as like the opening line, which I don't love. There's also this weird spacing was right there. I don't know if it tried to like center it, but I do notice that. Anyways, we have sort of like the TL;DR, we have these different headers here. I'm yet to see an image. Here comes the first screenshot. So this is talking about, um, the three different types of ideas to compare. And in the screenshot we do see all three, so that matches up pretty well. Um, we see Agent Report Card as well as down here. So, really what I'm checking for right now is that it all sounds like me. So this is kind of trained on like my LinkedIn style guide and my YouTube style guide and like the way that I speak, but I'm also making sure that the images that are put into the article, the screenshots, are cropped. There's other subjective checks that we have to let it do things where it like opens up the app, and it scrolls through, and it reads through, and it screenshots, and it verifies, but a lot of those checks, like I said are more subjective. And so that's where iterating on these skills takes a lot of time because you're constantly running the skill, analyzing the output, giving feedback, running the skill again, analyzing the output, giving feedback, and you're basically just like training the skill where every single time you use it, it gets better. But the thing about verification is there's always a way. You might be thinking to yourself, "Uh, well, I don't really know how I would have agents verify that process." Agents can do anything. They can use your computer, they can use your browser, they can look at things, they can listen to things, they can do anything. So just think about it like this, if you assigned that task to a human, what would you do to basically give it the stamp of approval, and then just explain that to the agent, and that's your verification. That doesn't mean it's going to always come out 100%, but it's going to be more like 95% rather than giving you something that started off at like 75 or 80. Okay, and moving on to number five, we have Walk It Down. And when I say walk it down, I mean walking it down the model list. So, whether you're using Claude Code or Codex, we all know that different models have different strengths and we all know that models that have the strongest strengths are the most expensive. So, for example, Astra is more expensive than Sol, but more capable than Sol. Sol is more expensive than Terra, but more capable than Terra. So, the idea is you build a skill and you're getting this good output and maybe you were testing that with Astra. Okay, well, let's run that same skill on Sol. Are we getting the same result? If we're getting the same result and it's cheaper, okay, let's test it on Terra now. Are we getting the same result? Is it cheaper? Okay, let's test it on Luna. There's no reason to be running a skill with Astra if you could run it with Luna and get the exact same output. But it is important to test it because not every skill fits that. I've had some skills where I run them on Haiku or Luna because they're just super simple. For example, when I give my agent a YouTube video and I say, "Hey, I need a YouTube video description and timestamps," that I can do with Luna and doing it with Astra is overkill. But for this x-article thing, I typically like to use at least Sol or Astra because it just has a better understanding of screenshotting things and putting it in the right spot and cropping it down and even blurring things out. It's so much better with the browser use with Sol and Astra. And then, once you've landed on a model that you like, you can take that one step further and you can walk it down the effort level. So, start it off on maybe High, and if it's not good enough, then move it up a little bit. And if it is good, then move it down to Medium. And just keep finding, basically, the simplest model or the lightest and cheapest model that still executes at the level of quality that you're looking for. And then moving on here to number six, we have the Bike Method, which is I've kind of already alluded to it multiple times throughout this video so far. But really, it's just the idea to me that your skill is never done. Every single time you run the skill, you're going to improve it. You're going to say, "Hey, here's what I really liked, here's what I didn't like." Update the skill so that next time it's better. Almost every single time I run a skill, I give it feedback and I tell it to update. And if you think about this, like you're teaching a kid to ride a bike, that's what a skill really should feel like. You start off very cautious, you're watching everything, you're guiding them, you—you've got your hand on this the the steering wheel, steering wheel, I meant like handlebars, and you're—you're right there. And then what happens is you give feedback, you say like, "Okay, that was good, but you know, you were leaning a bit too far to the left, make sure your weight's more in the center." And then you keep going. Eventually, you take off the training wheels because the skill is getting better and you're getting more trust in the skill, and then you give more feedback again, and then eventually you take off the elbow pads, right? And then you keep giving feedback. But that doesn't mean that you just take off the kid's helmet and let them bike on the highway. You're not going to do something reckless like that. You're also not just going to shoot them off down the road and then go inside and take a nap. You're still going to be in in some way, watching or having certain guard rails in place in order to make you feel comfortable and in order for you to have more trust in the skill itself. And I truly believe in the idea that there's no such thing as a finished product unless your skill is super, super deterministic and it runs perfect every time and you don't have to worry about it. But when the skills are more on the judgement side of the spectrum, you know, as we talked about up here, the more your skills live over here, the less I believe that you can actually have a finished skill, because like I said, every time things might happen a little bit differently, and every time your process might change or new models might drop, and you're just constantly going to be giving feedback every time you use it. Okay, so let's hop into Codex here and we can take a look at how this actually works. Now, when you are inside of your project, and if you're inside of your AI OS, which is usually what I'm building in, the Codex skills will live in a folder called the .agents, and then inside of the .agents, there's a skill—or sorry, not a skill—there's another folder called .skills. And if you're doing this in Cloud Code, the skills live in a .cloud and then in a folder called skills. Now, these can be transferable. I basically just tell Codex or Cloud, "Hey, see all the skills in there and bring them over here, too," so it just I have duplicates of them. But that's where they actually live inside of your project, unless you have them at a global level, which just means they're living locally across all of your Codex projects, rather than just inside of your AI OS. So that's where they live, and you can always just say, "Hey, you know, Codex, I have this skill, I don't remember where it is, could you show me the file path?" And it will find it, and then you can move it if you want. But anyways, that's where they live. And you can see I actually just ran this x-article skill on a video, so it says your x-article draft is saved, I'll click into it, and I actually did obviously already publish this one, but I didn't even have to change anything. I basically just clicked in here, I read it all, and I published it. And you can see that as it's going through, it already formatted everything, and it used the video of course. It took screenshots of things and it like put them in the right spot to line up with what was actually being said in the article, and it even added this little, um, I don't know what you want to call it like a spotlight. It added that. I didn't actually do that, but it will basically crop things out and it will add things like that into my actual article because I've told it in the skill to do things like that. And you can see that for this specific video, this ran on GPT-6 Astra. So what I'm going to do real quick is I'm going to walk this down. I'm going to give this same prompt to Sol and Terra and tell it to make the x-article and it's going to use the skill and we'll see how those outputs compare to Astra's output. Because this is a pretty complex skill, it has to transcribe the video, it has to open up the article, it has to make the thumbnail, it has to format everything, and it has to screenshot everything and put it in and then actually like drag it around. So there's a lot of browser use, and there is a ton of judgement inside of this, like I said this is not a deterministic skill. So you can see right now I have Sol, Terra, and Luna all running. It's been about 10 minutes, so I'll just check with you guys when this is done. But I just wanted to show you how all of them because of the skill, because of the verification are viewing images. You can see how many times Luna has viewed these images here, because it's taking screenshots and it's trying to understand, okay, which of these should be worked into the article and where and why. Same thing down here with Sol, obviously this one is a super, super visual-heavy, sort of skill. But because I've worked that in there, all of them are viewing images no matter what. And so it'll be really interesting to see how they decide to place them and, you know, what images they actually choose to put in there, because X limits you to a certain amount of images per article. So, I'm going to let these run, we'll compare the results, and we'll kind of talk about how walking it down has changed my perspective on this specific skill. All right, so you can see now that we have Luna and Terra have finished up. Sol is still working, so we'll start reviewing these two, and then we'll hop over to Sol. Hopefully it'll be done by the time we're looking through these. So, Luna took 28 minutes and 33 seconds. Let's open up the draft and see what we are working with. So we have, um, the thumbnail in there. We have "I designed a one-person million-dollar business with Claude." Um, the title is the same as like the opening line, which I don't love. There's also this weird spacing was right there. I don't know if it tried to like center it, but I do notice that. Anyways, we have sort of like the TL;DR, we have these different headers here. I'm yet to see an image. Here comes the first screenshot. So this is talking about, um, the three different types of ideas to compare. And in the screenshot we do see all three, so that matches up pretty well. Um, we see Agent Report Card as well as down here. So, really what I'm checking for right now is that it all sounds like me. So this is kind of trained on like my LinkedIn style guide and my YouTube style guide and like the way that I speak, but I'm also making sure that the images that are put into the article, the screenshots, are cropped. There's other subjective checks that we have to let it do things where it like opens up the app, and it scrolls through, and it reads through, and it screenshots, and it verifies, but a lot of those checks, like I said are more subjective. And so that's where iterating on these skills takes a lot of time because you're constantly running the skill, analyzing the output, giving feedback, running the skill again, analyzing the output, giving feedback, and you're basically just like training the skill where every single time you use it, it gets better. But the thing about verification is there's always a way. You might be thinking to yourself, "Uh, well, I don't really know how I would have agents verify that process." Agents can do anything. They can use your computer, they can use your browser, they can look at things, they can listen to things, they can do anything. So just think about it like this, if you assigned that task to a human, what would you do to basically give it the stamp of approval, and then just explain that to the agent, and that's your verification. That doesn't mean it's going to always come out 100%, but it's going to be more like 95% rather than giving you something that started off at like 75 or 80. Okay, and moving on to number five, we have Walk It Down. And when I say walk it down, I mean walking it down the model list. So, whether you're using Claude Code or Codex, we all know that different models have different strengths and we all know that models that have the strongest strengths are the most expensive. So, for example, Astra is more expensive than Sol, but more capable than Sol. Sol is more expensive than Terra, but more capable than Terra. So, the idea is you build a skill and you're getting this good output and maybe you were testing that with Astra. Okay, well, let's run that same skill on Sol. Are we getting the same result? If we're getting the same result and it's cheaper, okay, let's test it on Terra now. Are we getting the same result? Is it cheaper? Okay, let's test it on Luna. There's no reason to be running a skill with Astra if you could run it with Luna and get the exact same output. But it is important to test it because not every skill fits that. I've had some skills where I run them on Haiku or Luna because they're just super simple. For example, when I give my agent a YouTube video and I say, "Hey, I need a YouTube video description and timestamps," that I can do with Luna and doing it with Astra is overkill. But for this x-article thing, I typically like to use at least Sol or Astra because it just has a better understanding of screenshotting things and putting it in the right spot and cropping it down and even blurring things out. It's so much better with the browser use with Sol and Astra. And then, once you've landed on a model that you like, you can take that one step further and you can walk it down the effort level. So, start it off on maybe High, and if it's not good enough, then move it up a little bit. And if it is good, then move it down to Medium. And just keep finding, basically, the simplest model or the lightest and cheapest model that still executes at the level of quality that you're looking for. And then moving on here to number six, we have the Bike Method, which is I've kind of already alluded to it multiple times throughout this video so far. But really, it's just the idea to me that your skill is never done. Every single time you run the skill, you're going to improve it. You're going to say, "Hey, here's what I really liked, here's what I didn't like." Update the skill so that next time it's better. Almost every single time I run a skill, I give it feedback and I tell it to update. And if you think about this, like you're teaching a kid to ride a bike, that's what a skill really should feel like. You start off very cautious, you're watching everything, you're guiding them, you—you've got your hand on this the the steering wheel, steering wheel, I meant like handlebars, and you're—you're right there. And then what happens is you give feedback, you say like, "Okay, that was good, but you know, you were leaning a bit too far to the left, make sure your weight's more in the center." And then you keep going. Eventually, you take off the training wheels because the skill is getting better and you're getting more trust in the skill, and then you give more feedback again, and then eventually you take off the elbow pads, right? And then you keep giving feedback. But that doesn't mean that you just take off the kid's helmet and let them bike on the highway. You're not going to do something reckless like that. You're also not just going to shoot them off down the road and then go inside and take a nap. You're still going to be in in some way, watching or having certain guard rails in place in order to make you feel comfortable and in order for you to have more trust in the skill itself. And I truly believe in the idea that there's no such thing as a finished product unless your skill is super, super deterministic and it runs perfect every time and you don't have to worry about it. But when the skills are more on the judgement side of the spectrum, you know, as we talked about up here, the more your skills live over here, the less I believe that you can actually have a finished skill, because like I said, every time things might happen a little bit differently, and every time your process might change or new models might drop, and you're just constantly going to be giving feedback every time you use it. Okay, so let's hop into Codex here and we can take a look at how this actually works. Now, when you are inside of your project, and if you're inside of your AI OS, which is usually what I'm building in, the Codex skills will live in a folder called the .agents, and then inside of the .agents, there's a skill—or sorry, not a skill—there's another folder called .skills. And if you're doing this in Cloud Code, the skills live in a .cloud and then in a folder called skills. Now, these can be transferable. I basically just tell Codex or Cloud, "Hey, see all the skills in there and bring them over here, too," so it just I have duplicates of them. But that's where they actually live inside of your project, unless you have them at a global level, which just means they're living locally across all of your Codex projects, rather than just inside of your AI OS. So that's where they live, and you can always just say, "Hey, you know, Codex, I have this skill, I don't remember where it is, could you show me the file path?" And it will find it, and then you can move it if you want. But anyways, that's where they live. And you can see I actually just ran this x-article skill on a video, so it says your x-article draft is saved, I'll click into it, and I actually did obviously already publish this one, but I didn't even have to change anything. I basically just clicked in here, I read it all, and I published it. And you can see that as it's going through, it already formatted everything, and it used the video of course. It took screenshots of things and it like put them in the right spot to line up with what was actually being said in the article, and it even added this little, um, I don't know what you want to call it like a spotlight. It added that. I didn't actually do that, but it will basically crop things out and it will add things like that into my actual article because I've told it in the skill to do things like that. And you can see that for this specific video, this ran on GPT-6 Astra. So what I'm going to do real quick is I'm going to walk this down. I'm going to give this same prompt to Sol and Terra and tell it to make the x-article and it's going to use the skill and we'll see how those outputs compare to Astra's output. Because this is a pretty complex skill, it has to transcribe the video, it has to open up the article, it has to make the thumbnail, it has to format everything, and it has to screenshot everything and put it in and then actually like drag it around. So there's a lot of browser use, and there is a ton of judgement inside of this, like I said this is not a deterministic skill. So you can see right now I have Sol, Terra, and Luna all running. It's been about 10 minutes, so I'll just check with you guys when this is done. But I just wanted to show you how all of them because of the skill, because of the verification are viewing images. You can see how many times Luna has viewed these images here, because it's taking screenshots and it's trying to understand, okay, which of these should be worked into the article and where and why. Same thing down here with Sol, obviously this one is a super, super visual-heavy, sort of skill. But because I've worked that in there, all of them are viewing images no matter what. And so it'll be really interesting to see how they decide to place them and, you know, what images they actually choose to put in there, because X limits you to a certain amount of images per article. So, I'm going to let these run, we'll compare the results, and we'll kind of talk about how walking it down has changed my perspective on this specific skill. All right, so you can see now that we have Luna and Terra have finished up. Sol is still working, so we'll start reviewing these two, and then we'll hop over to Sol. Hopefully it'll be done by the time we're looking through these. So, Luna took 28 minutes and 33 seconds. Let's open up the draft and see what we are working with. So we have, um, the thumbnail in there. We have "I designed a one-person million-dollar business with Claude." Um, the title is the same as like the opening line, which I don't love. There's also this weird spacing was right there. I don't know if it tried to like center it, but I do notice that. Anyways, we have sort of like the TL;DR, we have these different headers here. I'm yet to see an image. Here comes the first screenshot. So this is talking about, um, the three different types of ideas to compare. And in the screenshot we do see all three, so that matches up pretty well. Um, we see Agent Report Card as well as down here. So, really what I'm checking for right now is that it all sounds like me. So this is kind of trained on like my LinkedIn style guide and my YouTube style guide and like the way that I speak, but I'm also making sure that the images that are put into the article, the screenshots, are cropped. There's other subjective checks that we have to let it do things where it like opens up the app, and it scrolls through, and it reads through, and it screenshots, and it verifies, but a lot of those checks, like I said are more subjective. And so that's where iterating on these skills takes a lot of time because you're constantly running the skill, analyzing the output, giving feedback, running the skill again, analyzing the output, giving feedback, and you're basically just like training the skill where every single time you use it, it gets better. But the thing about verification is there's always a way. You might be thinking to yourself, "Uh, well, I don't really know how I would have agents verify that process." Agents can do anything. They can use your computer, they can use your browser, they can look at things, they can listen to things, they can do anything. So just think about it like this, if you assigned that task to a human, what would you do to basically give it the stamp of approval, and then just explain that to the agent, and that's your verification. That doesn't mean it's going to always come out 100%, but it's going to be more like 95% rather than giving you something that started off at like 75 or 80. Okay, and moving on to number five, we have Walk It Down. And when I say walk it down, I mean walking it down the model list. So, whether you're using Claude Code or Codex, we all know that different models have different strengths and we all know that models that have the strongest strengths are the most expensive. So, for example, Astra is more expensive than Sol, but more capable than Sol. Sol is more expensive than Terra, but more capable than Terra. So, the idea is you build a skill and you're getting this good output and maybe you were testing that with Astra. Okay, well, let's run that same skill on Sol. Are we getting the same result? If we're getting the same result and it's cheaper, okay, let's test it on Terra now. Are we getting the same result? Is it cheaper? Okay, let's test it on Luna. There's no reason to be running a skill with Astra if you could run it with Luna and get the exact same output. But it is important to test it because not every skill fits that. I've had some skills where I run them on Haiku or Luna because they're just super simple. For example, when I give my agent a YouTube video and I say, "Hey, I need a YouTube video description and timestamps," that I can do with Luna and doing it with Astra is overkill. But for this x-article thing, I typically like to use at least Sol or Astra because it just has a better understanding of screenshotting things and putting it in the right spot and cropping it down and even blurring things out. It's so much better with the browser use with Sol and Astra. And then, once you've landed on a model that you like, you can take that one step further and you can walk it down the effort level. So, start it off on maybe High, and if it's not good enough, then move it up a little bit. And if it is good, then move it down to Medium. And just keep finding, basically, the simplest model or the lightest and cheapest model that still executes at the level of quality that you're looking for. And then moving on here to number six, we have the Bike Method, which is I've kind of already alluded to it multiple times throughout this video so far. But really, it's just the idea to me that your skill is never done. Every single time you run the skill, you're going to improve it. You're going to say, "Hey, here's what I really liked, here's what I didn't like." Update the skill so that next time it's better. Almost every single time I run a skill, I give it feedback and I tell it to update. And if you think about this, like you're teaching a kid to ride a bike, that's what a skill really should feel like. You start off very cautious, you're watching everything, you're guiding them, you—you've got your hand on this the the steering wheel, steering wheel, I meant like handlebars, and you're—you're right there. And then what happens is you give feedback, you say like, "Okay, that was good, but you know, you were leaning a bit too far to the left, make sure your weight's more in the center." And then you keep going. Eventually, you take off the training wheels because the skill is getting better and you're getting more trust in the skill, and then you give more feedback again, and then eventually you take off the elbow pads, right? And then you keep giving feedback. But that doesn't mean that you just take off the kid's helmet and let them bike on the highway. You're not going to do something reckless like that. You're also not just going to shoot them off down the road and then go inside and take a nap. You're still going to be in in some way, watching or having certain guard rails in place in order to make you feel comfortable and in order for you to have more trust in the skill itself. And I truly believe in the idea that there's no such thing as a finished product unless your skill is super, super deterministic and it runs perfect every time and you don't have to worry about it. But when the skills are more on the judgement side of the spectrum, you know, as we talked about up here, the more your skills live over here, the less I believe that you can actually have a finished skill, because like I said, every time things might happen a little bit differently, and every time your process might change or new models might drop, and you're just constantly going to be giving feedback every time you use it. Okay, so let's hop into Codex here and we can take a look at how this actually works. Now, when you are inside of your project, and if you're inside of your AI OS, which is usually what I'm building in, the Codex skills will live in a folder called the .agents, and then inside of the .agents, there's a skill—or sorry, not a skill—there's another folder called .skills. And if you're doing this in Cloud Code, the skills live in a .cloud and then in a folder called skills. Now, these can be transferable. I basically just tell Codex or Cloud, "Hey, see all the skills in there and bring them over here, too," so it just I have duplicates of them. But that's where they actually live inside of your project, unless you have them at a global level, which just means they're living locally across all of your Codex projects, rather than just inside of your AI OS. So that's where they live, and you can always just say, "Hey, you know, Codex, I have this skill, I don't remember where it is, could you show me the file path?" And it will find it, and then you can move it if you want. But anyways, that's where they live. And you can see I actually just ran this x-article skill on a video, so it says your x-article draft is saved, I'll click into it, and I actually did obviously already publish this one, but I didn't even have to change anything. I basically just clicked in here, I read it all, and I published it. And you can see that as it's going through, it already formatted everything, and it used the video of course. It took screenshots of things and it like put them in the right spot to line up with what was actually being said in the article, and it even added this little, um, I don't know what you want to call it like a spotlight. It added that. I didn't actually do that, but it will basically crop things out and it will add things like that into my actual article because I've told it in the skill to do things like that. And you can see that for this specific video, this ran on GPT-6 Astra. So what I'm going to do real quick is I'm going to walk this down. I'm going to give this same prompt to Sol and Terra and tell it to make the x-article and it's going to use the skill and we'll see how those outputs compare to