Here are five free and open-source Claude skills and plugins that I wish I knew earlier. One gives you visual components so good that your app or website stands out from the rest. Another hunts down every AI giveaway in your writing before it goes out. And the one plugin you should set up today or risk having your agent accidentally delete your whole workspace. Let's dive into it. So the first project that we'll be looking at today might just be one of the most important things that you can set up for your Claude code. And I first learned of this when I saw this post from Matt Schumer whom a few days ago he shared this serious incident where GPT-5.6 Soul, which is OpenAI's most frontier model, just accidentally deleted almost all of his Mac's files. And you can see he shared this in this expose which already got something like 5 million plus views. And essentially what happened here is when Matt gave it a task, it incorrectly ran this rmrf command. Which if you don't know what that is, that basically removes all of the files in this folder, which you can see is pretty much his whole user workspace. If you don't know Matt, he actually works with AI quite a lot. In fact, he was the one who wrote up this really viral post on X called something big is happening that was talked about quite a lot over the past few months. And which if you were close to the space, you probably have read it as well. Point being that Matt probably knows what he's doing. But despite that, because we put so much trust in these frontier models a lot, we indirectly and sometimes incorrectly assume that they will never run destructive commands like that rmrf that basically will wipe your entire workspace. And by the way, Matt is not the only one who reported this incident as well. So the lesson being, whenever there's a new model released, these types of risks are still present even at the level of intelligence that these models already have. But the good news is there's actually a pretty simple fix to it. Because one easy way by which you can solve this is by giving Claude code a plugin similar to this DCG repo. It is open source, it is free, and DCG basically stands for destructive command guard. And what it basically does is it blocks dangerous shell commands from being executed by agents unless you explicitly tell them to. And the way it works once you have it set up is really simple. So, because it runs in the background, if I run this prompt where you can see I just set up this sacrificial folder where I'm saying to clean up this workspace and delete everything in it in one command. Now, for 99.9% of instances, what you would most likely expect is that with a model that is as marked as GPT-5.6 or in this case, I'm just using Opus 4.8, they're actually smart enough to stop before deleting anything valuable. But as you can see with those examples that we showed earlier, there's still a very minuscule chance by which these agents may actually end up deleting important pieces of your project or your entire workspace as we've seen. And so, it's important to have something like this DCG plugin in order to stop that from happening. And to give an example of DCG in action, if I just confirm and choose number one, where basically I'm saying that it is for deletion and I do want it to be deleted. And especially for those instances when it accidentally runs this rmrf command, you'll can see that what it shows here is that it is blocked by DCG and it's saying here that a safety guard, which is that DCG plugin, intercepted our command and requires it to be run by hand. Now, if you have a legitimate need to delete these set of files, then you can obviously just ask Cloud Code to turn off the safety guard, but at least you have that confidence that this guard exists in the first place. And the way DCG works that makes it foolproof is that it is actually running via a hook. So, if we look at their documentation here, basically DCG runs as a hook. And if you don't know what that is, that's a feature in Cloud Code and a lot of agentic coding platforms that runs in the background, let's say whenever you start a session or whenever you use a tool or whenever you send a prompt. The point being that it is deterministic code that costs zero tokens and actually runs in the background regardless of what your AI agent thinks. And the great thing about it is that you can actually customize this. So, if you're a bit more advanced and you only want to block certain commands or in certain pockets of your work, then you can actually do that through their customizations here. And it's also very light, so you won't even notice that it's there up until the point that it saves you from your agent deleting your entire workspace. And by the way, just to make this easy for you to install, I also made this PDF which you can just grab in our free community below. And once you get this, you can just provide this whole thing to your agent, then your cloud code will be able to set up these plugins and skills for you in an instant. And by the way, if you want to learn how to build and sell AI systems that businesses actually pay for, then that's pretty much all we do over at the Robo Notify community. Where not only do you get access to the Cloud Living Masterclass, which we update every week and takes you from zero to mastery with the latest on AI, but you also get access to our agents as a service course, which walks you through how to actually get paid for all these AI skills that you are learning. You also get to be part of a genuinely great community of AI builders. In fact, you can see just some of the recent wins our members are getting from the program right here. So, if you want to start earning from AI, then check that just in the pinned comment below. Now, back to the video. Now, the next resource that we're going to look at is courtesy of David Hass, who launches really great project called canvasui.dev, where basically he is sharing these 24 components that are different shaders that you can use on top of any of your builds. And by the way, I first saw this when Shazzy and the sort of legendary developer who made a lot of UI components that you see in a lot of SaaS apps actually provided endorsement for this open source repo. So, that means a lot if it comes from him. And the way this works if you go to their documentation at canvasui.dev is that they have these different components that is pretty much ready to ship and include in any build that you want. Like you can see this one is like a bubble cursor, which is pretty cool. The one that I really like is this hex float grid, which if you go to that, you can see that it actually applies a layer of hex in the background that I think if you have a technology-driven website or if you have an AI agency website, then this for sure would add a more professional and differentiated look to it. I think the other thing that the guys behind canvasui.dev have done really well here is that you can actually tweak the size as well as the different settings of each of these components and shaders to your own liking. So, let's say you wanted to have, let's say, less iridescence, less grain, and let's say we actually don't want it to be floating like that and just be fixed, then that is something that we can do. And once we've made those changes and uh actually go down to the code portion here, we can just copy the whole thing and provide it to Cloud Code. And if you need a quick example of how that looks like in production in an actual website, you can just head to getrubric.app where I included that component as part of the site. And there's 24 of these components that are all free that you can just customize and tweak depending on the build that you are making. Now, of course, you have the option to just go to the website and copy the components here, or you can also just go to their repository here under canvas UI, which I've also included in that PDF guide below. And uh you can just import it to your Cloud Code so that you can grab these components whenever you need them. To give an example, earlier I was playing around with it and I just had Cloud Code create or render these different components so that I can check them out. This one applies like a dither to a 3D object. This one applies that bubble, that glass, I guess, that we were viewing earlier to your cursor. This one is that bubble effect, and there's just several of these effects that you can just tweak depending on what it is that you need. So, if you need a free and easy way for which to upgrade your front-end design builds and skill, then I think canvasui.dev is a good resource for you to check out. Now, as a bonus resource that you should check out if you're into front-end design is this post and project by Jacob. I hope I'm pronouncing his name correctly, but if you're, for example, creating an application that involves some loading screens, then you can actually upgrade the look of your application with some creative components that show these loading screens or these thinking screens. The great thing is Jacob actually shared this whole thing, and I'll include that in the resource guide as well below. And similar to canvasui.dev, you can basically just grab these components and copy the code for them in and provide it to your Claude code. And by the way, apart from these thinking orbs, he also did a great job around these border beam components and these liquid metal components that are actually looking really stunning when it applies to uh buttons that you want to highlight in your website or app. Now, the next skill we're going to look at is courtesy of Peter Yang, who actually has a YouTube channel as well. So, check out his stuff also because it's great. But basically on X as well as in his newsletter, I believe, Peter writes a lot of stuff. And obviously, with AI writing, you probably have seen a lot of tells already. And what Peter shared, I think just yesterday, is that he's open-sourcing his no AI slob skill that basically removes 20-plus patterns that pretty much tell that a piece of writing has been generated using AI. And if you've used AI to write before, you probably have seen these patterns quite a lot. So, for example, binary contrast like it's not X, it's Y or openers like here's what nobody tells you, like those sort of openers or manners of writing, they're pretty much an obvious tell that the piece of writing that you're reading was generated with AI. So, with this repo, it just provides you with a lot of guardrails and safety nets basically to prevent your AI agent from including these different AI tells. Now, if you use a lot of Claude code for writing or for reviewing your writing, then like me, you might have already a skill that is quite similar to this. So, for example, for myself, I have this humanizer skill. And so, a good tip for you to install this is that instead of replacing the skill that you already have, you can actually just ask Claude code to improve whatever skill that you are already used to using and just point it to a resource like what Peter has shared. And just generally, I find that with skills like these where you probably have already one that is already functioning for you. You can see this humanizer skill, I already have like a version three of it and I just constantly improve it with resources like this that I find over the web just to see if there's anything else that I can do to improve this existing skill of mine. Or if you don't have a writing skill similar to this yet, then you can just go ahead and directly import this because honestly, it's a pretty simple build with a couple of pretty simple rules, I think. But Peter has done a great job in just enumerating them to make sure that the type of writing that your AI helps you with doesn't sound too generic and sloppy. Now, the last skill that we're going to look at today is something that I saw from this X post, where he said that whoever shared the I have ADHD skill with me, thank you, because it's made my Claude replies so good. And believe it or not, this actually went pretty viral on X, and it does one simple thing. And this is something that you may not realize actually adds a lot of value for whenever you talk to your agents, because by their nature, I think a lot of these AI models, like Opus or even Fable, are quite verbose. But there are times when you actually just want to get the digest. Like, what's the TLDR? What is the summary of information so that you can get to it and actually respond or act based on that information. So, basically, what it does, very simple, it's a skill on Claude code that stops burying the lead, burying the answer. And it replies with the action first, and just generally makes her agents less verbose. Again, it's open source, it exists as a plugin that you can set up in Claude or even Codex. You can see there's files here for those different Agentic coding platforms. And the way that I use this personally, because I think right now, if you just install this by default, it's sort of a mode that you toggle on by typing like I have ADHD as a skill with a backslash. But at the core of this, it just includes these 10 rules or 10 principles where if you install this to your Claude code and activate it, it will essentially follow these rules, where it will lead with the next action, it will number multi-step tasks, and generally just a few good principles to make its reply to you during that session a bit less verbose and a bit more easy to digest. And so, for me, how I installed it is similar to how I did it with Peter Yang's skill, because in my workspace, I have this skill already called {slash} quick. And whenever I invoke that, it's an easy way for me to force Claude code to summarize its next reply in concrete bullet points. So, you can see here, I actually invoked that skill to explain how my quick skill is set up and how it relates to the I have ADHD skill, which I've already implemented or integrated in that sort of skill set. And so, by typing {slash} quick three, it actually gave me three bullet points to explain how that skill works. So, it says here that quick, which is the skill that I have, it gives me the answer right away. But, how I amended it is I just added those 10 I have ADHD principles from this GitHub open source repo in order to improve it some more. And then, at least for me personally, I find that adding like a number after the skill, like for example {slash} quick three, allows me to get exactly the number of bullet points that I want in this format, where each bullet point is titled like this with a bolded part. So, it's much easier to read. So, they're very simple idea, but I think it's more useful than you think and is definitely something that you should try out. And there you go, those are five plugins, skills, and components that I wish I knew earlier and which you can definitely use today to improve your builds. All of those resources I just linked down below. And if you want to grab this PDF, which you can just provide to your Cloud Code and uh just ask it to set up all of those things that we talked about today in this video, then you can just also grab that in our free community. I hope that was useful and I'll see you all next time. Thank [music] you.

In this video, you will learn how you can create animated websites just like these using new model that just came out, which is Opus 5. And I think this is the best thing that happened to web designers because we get websites of Fable level for a fraction of the price. Full-on websites with multiple pages, multiple sections can be now made all using AI without Framer, without Webflow, without paying agencies thousands of dollars. This new model, which is Opus 5, is really cheap and you'll be able to create websites that are stunning just by typing words in the chat, similar to how I did here. All of the prompts from this video are available at motionsite.io. So, just go here and you'll have access. You can follow along as I create in this video. But, other than that, this video is not affiliated with any of the tools. I'll show you how I create the assets using Higgsfield, but I'm not affiliated. I don't have any links for them. So, feel free to use whatever you prefer. But, I'll show you how you can create these images, how you can then create a website, and I need made that, and everything will be done. So, let's start from the first step, which is actually looking for inspirations. There's a lot of different websites on Pinterest that you can find for any niche. Same building a website for real estate company. All that I had to type is real estate award-winning website. And then, we'll have a lot of inspirations that we can just take inspirations from. We don't have to copy exactly. And here again, just scroll through and find something that you like. Let's say I like this example, which is the first image, but currently it is in the uh like skewed mode. So, all I have to do is just copy the image, go to ChatGPT images. I'm going to go to Higgsfield, and I'm going to say, "Create me an image like this, just the front view, and also remove the stuff from the bottom." So, like you can see the map, and also you can see that there is some kind of intelligent assistant, state of California. Get rid of those stuff. Just create a hero section for a website using this image. My company's called roofing, green roofing. That's what you should put in in place of the logo. And also create a hero section for that. So, as you can see, all I have to like say is inspiration that I wanted to do and also a little bit info for my own company. And let's just click. I'm going to select a few different resolutions to see which one is fitting the best. In the meantime, I can also try a few different options. So, I can say take this hero section and redesign it to have instead of this skewed version, create a front view of the design and expand the image to be full height of the hero section without those stuff on the bottom. So, get rid of those like map and stuff like that and change the content. And let's just send that and wait a couple of seconds see what it comes back with. And this is the result that we received. As you can see, it redesigned the bottom part and it also make the design look front view. But let's improve that and change the image completely. So, the first thing I want to do is give it a reference image of this one. I want to create like a snowy ice feel. All I have to do is just copy this image as PNG, paste it into ChatGPT two images together and all you have to say is just change the first image to be kind of snowy white similar to the second image, but do not add the mountain and these this is the result that you will receive. So, now we can actually start thinking of how to bring it all into cloud. And one thing that I want to mention before that is that I want to create parallax effect. Basically, what it does allow me to do is put text behind the image. So, what I have to do is basically to show you how it's done. So, I made a duplicate and I removed the field, whether it's to Figma or using any AI tool, and now it is basically allow me to put text behind. So, let's say I want to type like eagle. And if I just increase the text, let's say I want to make it the font Anton, and increase the text, and all I can do is just move it behind in the layers, and I have this effect where the text is behind it. It is very easy to bring it to Claude, and then create different parallax effects, and that's exactly what I'm going to show you how to do next. So, the only prompt you need for this is again, just select this image and say, "Remove all of the contents from the image, and remove the part from the above the houses, so make it transparent." Content like buttons, UI elements, etc. So, let's just send it and see what it comes back with. And this is what we've got. All I have to do is just download this image, go to website like online background remover, and there are a lot of free ones. I'm going to use the ones that is directly in Figma. And now we need another layer, which is just the background without any houses, so the parallax effect would work very well. And for that, all we have to do is just go to the previous design and ask it to remove everything except the background. So, let's do that. I ask it to remove all of the buttons, and let's just do that. Let's say, "Remove the houses from the image. Remove everything. I just want the background. Create in 8K quality." Let's try this one, and let's also try Nano Banana, and see which one does the better. Nano Banana is just fraction of the price, but I for some reasons believe that the Nano Banana in the backgrounds are a little bit better cuz if you see here, there is some noise effect in the background and it will be even more noticeable soon, but I'll show you the result. Yeah, this is the result we got from Nano Banana. I think we can work with it. So, now let's just download and I'll show you how it's going to be positioned in Figma. So, once I download this, I can basically position this behind the house. And you'll see that this is what we have and then we can create kind of parallax effect that it will be moving once we scroll. Let's start moving that in Cloud. So, if you don't have it yet installed, just type Cloud in the browser and you'll have the access to click on this icon and you'll download it. Once there, you just click on new website. Make sure that you select the new folder. So, just click open folder, select the new folder and select uh winter site, whatever. And we can trust this workspace and start creating the website. Before that, let's go to websites and select the font pair that you like. For example, this one is pretty cool. I think that I want the fonts to be something similar to this. We don't have uh to do the other stuff. I just want the font. Or this one is also pretty nice. And so, all I have to do is just copy, paste it into like any doc view to see the fonts. So, this is the fonts using. We have some global CSS. Uh let's just take this. And now we can go to Cloud and I'm going to say, "Create me a hero section for a kind of imaginary winter real estate company and then use these fonts and also use this image as the background of our hero section." Align all of the headline, description, and button in the hero section to the center. And it kind of should be positioned closer to the top of the hero section. To give Claude a little bit better overview, I don't want to just upload this image. I'm going to upload this image and then I'm going to upload two separate images. So, this is the first result just to give Claude understanding of what kind of um website it is and what graphics are going to be used. So, let's make sure that Opus 5 is selected and let's just send that without any other prompts and see what it comes back with as a first result. The reason that I asked it to create just a hero section instead of the whole website is that I just want to kind of perfect the hero section before building the rest of the site, but it is totally optional. Take everything that I'll show in this video just as inspiration. You can ask freely and simply going and asking AI to create you the whole website without building the the hero section. There is a billion ways to create an AI website. These tips are just what I'm using for building websites, yeah. This is the first result. Let's say that First of all, let's get rid of all the brown color. Let's use just pure black for the button and for the tagline at the top. Also, let's move the buttons under the headline and move the description, which is Snow Country Homes, etc. Let's move it to the bottom of the hero section, the same position where it says scroll to explore. And let's just send that. And now we can also, after this, create the parallax effect. And this is the updated version. Looks much cleaner. Let's now get these two images that we just got. So, all we have to do is just export them as PNG. Make sure it is not JPEG cuz this will let the transparent background be visible. Just name it like front front and I'll put some letters so there is no duplicates. And this is going to be JPEG. And we're going to name it back. And now we have to just upload it to cloud and let's find front. Drag it straight into cloud and then we find back. And upload it to I'm going to say instead of current background image for the hero section, let's have these two images. So the one the second image that says back should be on the back and the front should have positive z-index so it's on top of the back image. And then once I scroll, there should be a parallax effect if you know what I mean. To for that parallax to be visible, create two or three more sections be below the hero section. And let's just send that and see what parallax effect can All Pix achieve. And this is the result that we received. Let's preview it in a browser. As you can see we have the hero section that we can improve a little bit. First, I don't really see that much of a parallax. So let's fix that. Let's say so the parallax currently doesn't really work. The front image should have parallax effect. And also let's change the font and the color. So it would once we scroll down there is different colors. So let's just take a screenshot of that. And to make it a little bit more smooth, we can just do something like this to have the same color in the background. And make the body of the whole website to be this color. Oops. Also remove the stickiness from the navbar. I don't want navbar to scroll with the page. It should stay at the same place. So, remove. Yep. Let's send that and see what it comes back with. And also, in the meantime, we can actually incre- uh improve the font. As you can see, it currently doesn't look that great. So, what if we create like winter or any text could be something related to a company or anything like that. Let's make sure that it's black. I just want to see how we can improve the hero section so it looks a little bit better. And then, we put the text behind it. Maybe we put the color of the fonts to be something like this. 100% opacity. And something more blue. Like this. So, it the text behind it. And of course, we would have the tagline, which is this one. Would stay at the same place that it is right now. And then, the navbar would be also there. So, just like that, we can have a cool website really easily. So, the the reason that I'm putting it all on Figma is just to show you how it can look once we do some edits. And then, the buttons, these two buttons can be at the bottom together with the text. So, together with this text, it can also go to the bottom like this. Yeah, something like this. And just like that, we can have a cool hero section. And the hero section is around 90% of the whole website. It should give strong first impression. And then, once you have that, it the rest is just uh putting the content out there. So, let's wait and see what Claude says. So, now the navbar, it's still working. Let's wait and see what it comes back with. So, now let's just take a screenshot of this or we can group it and I'm going to copy as an PNG instead of just verbally explaining it. Words, let's try to just provide this image as a reference and explain with our words. So, all I have to say is let's actually take a screenshot of this. So, I have the comparison of both of them and it's easier to speak. So, what Let's make changes to the hero section similarly to how you seen the reference images. So, instead of headline we'll have the big word winter and then under under that we'll have two buttons, but they will be moved to the bottom of the page. Up or above the description and then we move the whole front image a little bit more down so our winter word is in big size and still visible. Basically, position it similar to how it positioned on the image. And let's just copy this, paste this in here, and upload the reference image and click on send and see what it comes back with. And this is the result that we received. As you can see, it now has this parallax effect. If I preview in the browser, uh we can see that it looks great. Let's see how it looks like on the mobile. So, on the mobile there is room for improvement. First, I think we can just move the text under the hero section, but first uh the font is different. So, the font used in Figma is Anton and basically it's a Google font. What that means is that you just have to mention the name and then AI will install it by themselves. So, just say for the headline For the headline, let's use this font. Uh import from Google Fonts. Also, on the mobile, let's move the content under the headline, not on the bottom of the hero section, just the mobile. The desktop keep keep exactly as it is right now. And here's the result. Let's preview that in browser, and as you can see, we have this nice hero section that has parallax effect just from us spending like few minutes on editing it. I think it's still working on the mobile. Yep, still working. But, other than that, if we spend just a couple of more minutes on like building and designing the rest of the website, the result will be great. And this is how you create animated or 3D parallax effect using AI. If you've learned a thing or two, leave a like and subscribe, and I'll see you in the next video.

Hello, my friends and welcome back to the channel. I have been using Linux for about a year now, coming over from Windows, but I was also a Mac user as well. I'm one of those what I would like to call operating system agnostic people. At least I was. You see, now here we are a year later and I don't think there's enough money in the world to make me go back to a Windows and I probably could be convinced to use Mac OS again, but I don't think it's going to actually replace Linux for me. I think Linux is now my daily driver. So, I work in academia. Let me tell you all the things that many people have asked me since I started since I moved over to Linux, what I like, what I don't, and is this something you could realistically use? Not only as an academic, but like as a normal person as well. So, first and foremost, this is my desktop. People seem to think that if you're using Linux, you have to do everything during with terminal. That is not the case at all. You don't need to use terminal all that much. In fact, the times that I use terminal, let me show you what terminal looks like here. I should probably put my keyboard where I can actually reach it. So, let's take a look at what terminal is, okay? When I open up terminal, it looks like this. No at Pop!_OS. This looks scary for Windows users and Mac users probably because we don't use terminal hardly ever in those. Here on Linux, we use it sometimes, but I would argue that maybe you could probably use Linux with a modern Linux distribution without ever having to use terminal. In fact, I almost never use terminal. I can think of maybe two times that I did it. So, do you need to use terminal? Maybe. Should that scare you? No. The reason is that LLMs understand Linux really, really, really well because it's well documented. So, anytime I've had a question about what do I need to do, I can simply go to whatever your favorite LLM is and search for it and I get a response pretty quickly and pretty easily and it's usually right. I haven't had many problems getting information. Um the only thing you have to specify typically is what distribution you're using or what base version of Linux they're using. So, if we take a look at Linux, it's probably intimidating the first time when you're trying to choose a distribution to go with. So, with Windows, you get Windows, right? In Linux, there are a lot of them, okay? So, let's take a look at this chart. I'm not going to open it because it's absolutely huge. Each one of these lines represents a different distribution. But, in the big scheme of things, there's a few big ones that you should be aware of, okay? So, let's take a look at the list over here on the left. The first one is Debian. This is pretty common. We also have Ubuntu based ones that are pretty common. Uh you'll see things that are based on Arch. You see Arch a lot mentioned on uh YouTube and things like that because a lot of people really want to be like on that cutting edge and I don't I don't know. I don't look into Arch, so I don't know much about it. It's whatever. But, here's my point. Oh, I forgot Fedora. I'm sorry. For all the Fedora fans out there, I'm sorry. I forgot you. It was a little farther down here. And then there's a bunch of other ones that are smaller, don't have it And when I say smaller, I just mean you you don't see as much online about them. An example would be openSUSE. I see that sometimes, but not not as often. So, here's the thing. Don't let this overwhelm you, okay? Because when you look at this, like let's say you want to have one based on Ubuntu, okay? What What does that mean? You look at this and you're like, "Oh my gosh, there are so many versions." How do I pick which one? My advice, don't overthink it, okay? So, I think it kind of depends a little bit on what you want to do because one of the big things that you'll find is Nvidia drivers can be kind of annoying with Linux depending upon your distribution that you go with. So, knowing that I wanted to do AI and I wanted to play video games, that was one of the big things I needed was I needed Linux drivers just to work and that's why I ended up using Pop!_OS. Now, Pop!_OS, let me let me pull up their website so you can see what it looks like. Um All right. So, when we go to Pop!_OS, this is what their website looks like. It's made by uh System76, right? So, first of all, it's free. Second of all, uh at least the version that I downloaded, I'm not currently using the most recent version of this. Mine's based on Ubuntu 22 and that's just because that their current one when I started using Linux. They do have a new one now, but here's the thing. You don't have to go to the new version, right? Like, if you think about some of the other OSes out there, some of them like force you to move. I'm not being forced to move. They have a new version, and like I don't think I've even seen in like a pop-up for it, okay? So, I really like Pop!_OS. Um I think it's probably a good one for people to start with. Other ones that I know I've seen people be really happy with starting out on Linux is Linux Mint, and I've also seen people really happy with Bazzite. And Bazzite tends to be more of like gaming oriented people really like Bazzite. But, I can't really speak to any of those. I can only speak to Pop!_OS because that's what I use literally every day. And I will mention that I do also use base Ubuntu on my virtual machine that I run. And that's been good, too. Um I would have no problem recommending base Ubuntu or or Pop!_OS. I've been really, really happy with both. So, my point here is don't spend more time than you need getting lost in all the different distributions, different kinds of Linux, blah blah blah. Pick one that seems to meet your needs, and go from there. And like I said, base Ubuntu, Pop!_OS, and Bazzite are the ones that I know a lot of people use, and they seem pretty happy with them. Mint is also a good one from what I've heard. Like I said, I haven't used it. I started with Pop!_OS, was happy with it. When I set up my virtual machine, I went with Ubuntu because it was that's what Pop!_OS is based on. I'm happy. I don't feel the need to hop between distributions. Maybe if you start getting into Linux, you will. But, I don't I don't feel that way. I'm just happy with it. So, one of the things that I mentioned in a previous video was one of the reasons that I liked Pop!_OS and this look and feel is that to me it's kind of a hybrid between the great things about Windows and the great things about macOS. So, for example, I really like the floating dock down here at the bottom, which tended to be more of a macOS thing, at least when I moved over uh to Linux. This was definitely more of a Mac macOS thing. Windows was still, you know, having the start button over on the left with the little search bar and all that sort of silliness that I don't really prefer. But, let me show you the biggest thing or tell you the biggest thing, okay? So, first of all, there's no tracking that I'm aware of with Pop Pop!_OS. I I believe that the telemetry is disabled on it by default. Um with Ubuntu, if you want to turn off telemetry, there's a really easy button to just turn off anonymous telemetry. Um I don't think that's the case with Windows and Mac, but I don't know cuz I'm not using them anymore. The other thing that I want to mention here is there's no ads, right? Like on Windows, anytime I would open my start menu, I'd see these stupid ads for things. And look, I don't I don't have like a start menu per se on on Linux, right? What I do have is all my applications listed right here, right? Like >> [laughter] >> all my applications are right here. Oh, so that reminds me, one of the questions that uh my friends asked me is can you play video games? Yeah, you can play video games on Linux. It's not really that hard. Uh I have Steam running on here. You can see I have The Last of Us Part 1. I've been playing that. It's been a lot of fun. Um so, let me open up Steam. I'll show you that Steam works. So, the other thing that I'll mention here is, you know how uh some of the others, Windows, Mac, they have like, I think on Mac it's called Finder. I don't know what the equivalent thing is called on Windows when you're trying to search for something. So, what I really like is when I press on my keyboard, it's my Windows key, it pulls open the search, and it just finds what I need without any silliness. So, if I'm like, I want to find Steam, let's launch Steam. And I'll tell it to launch Steam, and it will usually do it. Steam's a little bit odd in the sense of I usually have to actually click the button from here in my applications menu before it starts up. And I think it's because it runs in the background. There we go. Okay, so you can see Steam just takes a minute. It's always taken a minute on my system to get Steam running, but that's literally the only app that I have that takes a minute to get up and going. No idea why that happens with Steam, but I'm not worried about it because when I play video games, it's like I just open it up, I wait what, 10, 15 seconds, and then Steam opens. So, here we go. Steam updates, everything runs. The only thing that I've set up different here is I had to set it up so that it runs off what's called Proton Experimental, so I'll go to my libraries real quick and go to The Last of Us. And if you see here, uh if you go into properties, you can go to compatibility, and you can change it to Proton Experimental. That is literally the only thing that I did to get games working. And really, this is the only game I play through Steam right now. I also have Arma Reforger. It works, but I don't think I can play multiplayer. So, if your game has kernel-level anti-cheat, it's probably not going to work on Linux, but you can just check. Just go to the web, search it, and find out. Like, look up Linux compatibility for games. The Call of Duty, for example, I don't think will work on Linux because it needs kernel-level anti-cheat from what I remember. Um but I also play Overwatch on here. That worked fine when I was still into that game. Literally no major problems at all running Steam games on Linux for me. Okay, so what about work? Well, if you guys are like me, the vast majority of my work these days is through online writing that's collaborative with things like Overleaf for using Google Docs and all that sort of stuff. That works fine. It's the internet, right? Search the internet, no problem at all. I run uh Chromium, and I also have Firefox, and I also have Brave on this system. All of it just works, no real issues. So, let's address the next thing, which is like, how do you get programs on a Linux computer? Well, all the distributions I'm aware of have something similar to this, which is like a little shop, right? So, it's similar to like the Windows Store, the Mac Store, or whatever you want to whatever you want to call it, depending upon your operating system. The difference is, on Linux, there's like multiple ways that you can actually install a program. So, when I open this, this is on Pop!_OS, this is called the Pop Shop. And here, these are called Flatpaks. Now, I'm not going to get into the difference between a Flatpak and an AppImage and a .deb file and all of that sort of stuff. That's really beyond the scope of this, but what I think you should know when you go into using Linux is there's a lot of different ways to install programs, and none of them are hard, right? So, let's let's think about actual work, okay? So, with my actual work stuff, what do I need? I need stats. What do I use for stats? I use either R, which has Linux version, or I use JASP. And JASP works on here. They have a flat pack for it. All I needed to do was click the install button. And like, if you want to see that it works and it looks the same, I'll just click open. And look, now we have stats. It's the exact same way that it run that it looks on Windows and Mac and all that. Literally no problem at all with any of my statistics stuff. Um what about writing? I use Libre Office Writer personally. There's a I haven't had any issues with compatibility or anything like that. Everything just works. What about PDFs and dealing with PDFs? Well, I found this program recently that is called PDF Arranger. And it is actually a flat pack. Let me show you what this looks like if you were to go try and get this app today. And so, this PDF Arranger app has worked pretty great for me. Uh it's been a little bit more flexible than, for example, uh oh, what's it called on Mac? I think Mac generally handles the uh PDFs a little bit better than than Windows in terms of like moving pages and things like that. But, I've been able to do similar. PDF Arranger haven't had any issues. So, when I was looking for it, I was trying to find a program that would deal with PDFs. And all I did was go into the Pop Shop and search. And I saw and I saw merge, shuffle, and crop. And I was like, well, that's exactly what I need to do most of the time. So, I'm just going to grab that one. And that's the one that I ended up using. And it's been quite useful. Haven't really had any issues with it. Oh, okay, I guess that's something I can say. You can customize like everything under the sun with Linux, okay? So, I have all my desktop icons hidden right now. There's usual They are usually have a bunch here. I just hit them using this extensions thing. This is built in. And it lets you turn on and off a whole bunch of options. But, with a lot of distributions, just about anything that you want to customize, you can. So, one of the things that I liked is I like this dock. And I liked it at the bottom. But, if you You to move it to the side or the top or you want to make it so that it auto hides. So, like let me open this. You can see that when I programs open, it auto hides and goes away, right? But, all I got to do is move the mouse down to get it back. You can set up all sorts of keyboard shortcuts and all of these other really fun things. Now, the next thing that you may be wondering about is all these different apps. So, I'll just go across them and I'll show you Chromium is a flat pack, Firefox is a flat pack, Spotify is a flat pack, Slack is a flat pack. Open WebUI, this is a PWA, progressive web app. Same thing with Google Calendar and Outlook. All those are PWAs. Lemonade I have installed via Docker. Planify is a nice little project management type app that's also a flat pack. LibreOffice Writer is also a flat pack. OBS Studio is also a flat pack. Steam, I think Steam is also a flat pack, but I don't really remember off the top of my head. So, there's a number of ways that you can do this, and they're all like just like any other program. The only thing that's more annoying, in my opinion, than it was on Windows, is if you want to create a PWA, it's kind of a multi-step process to get it to actually show up in your applications here. So, here's here's the thing at the end of the day. What do I like about Linux? It just works with the way that my brain works. If I had to summarize my experience with Linux so far, I would just say it's just good. Things just work, and I know that that's it's going to seem a little bit scary. So, like the thing that you have to remind yourself of is you are learning a new system, right? You're learning a new operating system. It's like going from Windows to Mac or vice versa. There's going to be things to learn cuz things just work a little bit differently, and you need to go into it with that open mind. I will say one of the things that I know people really like, they like the ability to have everything connected. So, with Pop!_OS it's really easy. So, they have calendar integration and all of that sort of stuff built in. All you do is log in with your with your accounts if you want to. You don't have to by any means. And all of it's just there. So, do I think people should try Linux? Yes. If you're unhappy with, you know, any of the things happening on Mac and Windows, I think it's worth trying Linux. Will I go back to Mac or Windows? No, probably not. I I don't see myself going back. Do I want to distro hop to other ones? Like, do I want to try and move everything over to go to the newest version of Pop!_OS or try out Fedora or Arch or any of those things? I can tell you I have no interest in Arch whatsoever, personally. Like, as I not really that interesting. Now, SteamOS I am interested in, but not necessarily for like my work PC. I think it might be kind of fun for if I had like a dedicated gaming PC. But anyway, I'm going to stop talking there so I don't ramble on forever. But all if I if you can take away one thing from this, it's that I have really enjoyed my year with Linux. I think most people, if they actually went into it with the perspective of yes, I'm going to have to learn it how to do a few things, but it'll get me a new operating system. Like, I'd have to do that if I moved to to Mac from Windows or I'd have to learn Windows if I was moving from Mac. So, like, if you're interested in it and you're open-minded about it, I think you'll probably end up liking it. A lot of things just work and they work well. One final thing that I'll answer here is my friends asked me like, how often do I use Windows because I do have Windows on a hard drive in my computer still. Well, in the last year I have literally gone into that hard drive one time and I have since found a way to resolve that problem that I was having. So, the issue that I had was I wanted to do some stuff in Word that I couldn't figure out how to do in LibreOffice. Long story short, I found out how to do it in LibreOffice office just as easily. So, I have not opened my Windows hard drive in a year and I'm getting to the point now where I am very strongly thinking about wiping it and just using it as extra storage for my Linux side because I don't see myself going back. So, that said, I hope this has been useful for you. I'm sorry it's a little bit more of a long-winded ranty type video than I would normally do, but hopefully you learned some things about Linux and saw the interface here, saw some of the interactions. Um, you know, you see that the app store and the applications, you know, it's not anything crazy. It's just relatively easy, straightforward, no stress. So that said, thank you guys. This been helpful for you, please like and subscribe and I will see you all in the next video.

The Prompt Is No Longer the Product How AI development moved from clever instructions to context, tools, loops, and coordinated agent systems We started by learning how to talk to AI. Now we are learning how to build reliable systems around it. Here is the practical journey from prompt engineering to graph engineering. Everyone is still asking the same question: Which AI model is the smartest? It sounds reasonable. A better model can usually reason more effectively, write more clearly, use tools more accurately, and handle more complicated tasks. But that question is becoming less useful. A powerful model placed inside a poorly designed system can still forget important information, choose the wrong tool, repeat itself, invent facts, exceed its budget, or stop before the work is complete. Meanwhile, a slightly less capable model inside a carefully designed system can often produce more dependable results. That is the shift now happening in AI engineering. The next major advantage will not come only from gaining access to a marginally better model. It will come from building a better system around the model. That is why the conversation is moving beyond prompt engineering. The development path now looks something like this: Prompt engineering → Context engineering → Harness engineering → Loop engineering → Graph engineering These are not five competing disciplines. They are five expanding layers of control. Prompt engineering controls the message. Context engineering controls what the model knows. Harness engineering controls how the model operates. Loop engineering controls how the system continues working. Graph engineering controls how models, agents, tools, memory, software, and people coordinate. Each new layer zooms farther out. We began by engineering individual instructions. We are now engineering complete environments in which artificial intelligence can operate. The important idea is simple: We started by engineering what we say to the model. We are now engineering everything that happens around it. This framework is not a formal industry standard, and the boundaries between the layers often overlap. A real AI application may use prompts, retrieved context, tools, validation, loops, and multiple agents at the same time. Still, the five-layer model gives us a practical way to understand how AI systems have evolved and why building dependable AI now requires much more than writing clever instructions. Open Cloud AI explains the systems, tools, workflows, and business shifts shaping artificial intelligence, without unnecessary complexity. Subscribe to Open Cloud AI Layer One: Prompt Engineering Engineering the message The first generation of mainstream generative AI use was dominated by prompt engineering. People quickly discovered that the wording of an instruction could dramatically change the output. A weak request might say: Write a cybersecurity incident report. A stronger prompt might say: Act as a senior security analyst. Write a concise incident report for the CIO. Include an executive summary, affected systems, timeline, containment actions, business impact, and recommended next steps. Separate confirmed facts from assumptions. Use professional language and avoid unnecessary technical jargon. The model did not change. The message did. Prompt engineering therefore focused on improving the ingredients inside a single interaction: - The role the model should assume - The task it should complete - The audience it should write for - The information it should use - The constraints it should respect - The examples it should follow - The format it should produce - The quality standard it should meet This remains valuable. For one-time tasks such as rewriting an email, summarizing a document, brainstorming ideas, explaining a concept, or preparing a first draft, a well-designed prompt may be enough. A good prompt reduces ambiguity. It tells the model what matters, what does not matter, and what a successful response should look like. But prompt engineering has an important limitation: It assumes the model already has everything required to complete the task. That assumption quickly breaks down in real work. Imagine asking an AI assistant to produce your organization’s monthly cybersecurity report. Even the most carefully written prompt cannot solve the problem if the model does not have: - This month’s incident data - Previous reports - The organization’s severity definitions - Current vulnerability findings - Open remediation items - The CIO’s preferred reporting format - Relevant policies - Updated asset information The problem is no longer how the request is written. The problem is what information enters the model’s working environment. That is where context engineering begins. Layer Two: Context Engineering Engineering what the model knows Prompt engineering asks: What should I tell the model? Context engineering asks: What should the model see at this exact moment? That difference is more important than it first appears. Context can include: - The user’s current request - Conversation history - Documents - Database records - Search results - Previous actions - Application state - User preferences - Organizational policies - Retrieved knowledge - Available tools - Access permissions - Output requirements A model may technically support a very large context window, but that does not mean every available document should be inserted into it. Too little context creates ignorance. Too much context creates noise. The engineering challenge is to select the right information, remove irrelevant material, compress lengthy content when necessary, preserve important details, and present everything in a form that the model can use. Context is not the same as memory Memory is one component of context, but the two are not identical. Memory answers: What should the system remember over time? Context answers: What should the model receive for this particular decision? A travel assistant, for example, may remember that a user: - Lives in Canada - Travels with children - Prefers low-cost flights - Holds a particular passport - Avoids certain transit countries For a domestic trip, passport and transit restrictions may be irrelevant. For an international trip, they may determine whether the proposed journey is possible. The system should not pass every remembered detail into every request. It should select only what is useful for the current task. That is context engineering. Retrieval becomes part of the system Many AI applications now use retrieval to locate relevant information before asking a model to respond. Instead of placing an entire knowledge base inside every prompt, the application searches for the most relevant sections and supplies only those results. A support assistant may retrieve: - The customer’s order - The refund policy - Delivery confirmation - Previous support messages - Account status - Authorized actions A cybersecurity assistant may retrieve: - The alert details - Device history - User activity - Known indicators of compromise - Related incidents - Response procedures Now the model is not answering from a generic instruction. It is reasoning inside a carefully assembled working environment. Tools are also context A model behaves differently when it knows it can: - Search the web - Query a database - Read a file - Execute code - Send an email - Review a calendar - Update a ticket - Call another agent Tool descriptions, permissions, and limitations become part of the model’s decision environment. But giving a model access to information and tools creates a new challenge. How do we make sure the model uses them correctly, safely, and consistently? That requires another engineering layer. Layer Three: Harness Engineering Engineering the machine around the model A model on its own produces a response. A harness turns the model into an operational component. The harness is the machinery surrounding the model: - Instructions - Context assembly - Retrieval - Tool access - Authentication - Permissions - Routing - Output schemas - Validation - Guardrails - Error handling - Logging - Tracing - Cost controls - Human approvals - Evaluations Think of the model as an intelligent employee. Prompt engineering writes the employee’s instruction. Context engineering prepares the employee’s briefing material. Harness engineering builds the employee’s workstation, tools, procedures, access controls, review process, and quality checks. A practical example Suppose you are building an AI system that investigates suspicious emails. A simple model may read the message and respond: This appears to be phishing. That may be useful, but it is not yet a reliable security system. A proper harness could: - Extract the sender, URLs, attachments, and message headers. - Check SPF, DKIM, and DMARC results. - Query domain and IP reputation sources. - Analyse suspicious links. - Search for similar messages across the organization. - Ask the model to classify the evidence. - Require a structured output. - Validate that conclusions are tied to evidence. - Escalate high-risk cases to an analyst. - Preserve the execution record for audit and review. The model remains important, but it is only one component inside a much larger machine. The harness determines what the model can do A text-only model has limited operational impact. A system that can send messages, block accounts, delete files, approve refunds, deploy code, or modify data can create real consequences. Controls therefore need to exist close to the action being performed. Reading a public webpage may require minimal approval. Sending an external email may require user confirmation. Disabling an account may require policy checks. Deleting production data should require strict authorization. The harness determines not only what the model knows, but also: - Which actions are available - Which actions are restricted - Which actions require approval - Which actions are prohibited - What evidence must be provided - What happens when an action fails Validation turns outputs into contracts A model naturally produces flexible language. Software systems often need precise outputs. An application may require: { "risk_level": "high", "confidence": 0.94, "evidence": [ "Newly registered domain", "Authentication failure", "Credential-harvesting link" ], "recommended_action": "isolate_message" } The harness can verify whether required fields are present, whether values match the expected format, and whether the recommendation is permitted. This is especially important when model output becomes input for another system. Without validation, a small formatting error can break an entire workflow. Observability becomes essential When a single prompt produces a poor answer, you can inspect the conversation. When an agentic system fails after twelve tool calls, several retries, two branches, and a database update, the final answer is not enough. You need to know: - What context was supplied? - Which tool was selected? - What arguments were passed? - What did the tool return? - What state changed? - Why did the system retry? - Where did the incorrect assumption appear? - Which validation step failed to catch it? - How much time and money did the run consume? This is why logs, traces, evaluations, and reproducible tests are becoming central parts of AI engineering. A production system should not merely produce an answer. It should produce an inspectable execution. At this point, we have something more capable than a chatbot. But it still performs only one pass. Real work often requires persistence. Layer Four: Loop Engineering Engineering how the system continues Many valuable tasks cannot be completed in one model call. Consider instructions such as: - Fix all failing tests. - Investigate a security incident. - Research a complex market. - Build and validate an application. - Compare several products using current information. - Improve a proposal until it meets defined requirements. - Review a large collection of contracts for risk. The first attempt will rarely be sufficient. The system must be able to: Gather information → Act → Observe the result → Verify progress → Decide what to do next That is a loop. A weak loop simply tells the model: Keep trying until the result is good. A professional loop defines what good means. A reliable loop needs a defined goal “Research artificial intelligence” is too vague. A stronger goal would be: Produce a 2,000-word briefing comparing five enterprise AI-agent platforms. Include pricing, security controls, integrations, limitations, and citations from primary sources. The system now knows what outcome it is trying to produce. It also needs success criteria The agent should know what completion looks like. For example: - Five platforms reviewed - Every pricing claim sourced - Security section included - Comparison table complete - No unsupported statistics - Word count within the required range - All required sources accessible Without success criteria, the system may stop too early or continue rewriting indefinitely. It needs progress checks After each cycle, the system should ask: - What has been completed? - What is still missing? - Did the last action add useful information? - Are we repeating the same search? - Has contradictory evidence appeared? - Is another tool required? - Is human input needed? These checks prevent the system from confusing activity with progress. It needs stopping rules A loop without limits can waste time, money, and computing resources. Common stopping conditions include: - Maximum iterations - Maximum tool calls - Token budget - Cost budget - Time limit - Repeated-action detection - Required confidence level - Successful completion tests - Human escalation threshold The hardest part of loop engineering is often not getting the system to continue. It is getting it to stop at the right time. Verification changes everything A weak loop says: Try again if the answer seems poor. A stronger loop says: - Perform the task. - Run a test. - Compare the result with predefined criteria. - Identify the exact failure. - Update the working context. - Retry only the failed component. - Stop when the tests pass or the budget is exhausted. This is the difference between repetition and controlled improvement. Imagine an AI coding agent. A weak version writes code and declares the task complete. A stronger version: - Writes the code - Runs the application - Executes tests - Reads the errors - Identifies the responsible component - Modifies the relevant code - Reruns the tests - Stops only when the required checks pass The loop turns output generation into goal-directed execution. The danger of the “vibe loop” Without measurable criteria, an agent may repeatedly rewrite an output because it “could be better.” But better according to what? Instead of telling an agent: Keep improving the article. Use: Review the article against these seven criteria. Revise only the sections that fail. Stop once every criterion passes or after three revision cycles. A loop should not behave like an anxious perfectionist. It should behave like a disciplined operator. Layer Five: Graph Engineering Engineering the organization A loop makes one agent’s behaviour programmable. A graph makes an entire AI organization programmable. At this layer, the system may contain: - A planner - Several researchers - A data analyst - A builder - A writer - A reviewer - A compliance checker - A human approver - Shared memory - External tools - Conditional routes - Parallel branches - Feedback mechanisms The word graph comes from the structure of the workflow. A graph normally contains three basic elements: State The information representing the system’s current position. Nodes The components that perform work. Edges The rules determining what runs next. A node might contain a model, a tool, ordinary software, a human approval step, or a combination of them. An edge might be sequential, conditional, parallel, or based on the result of a previous action. This is important: Graph engineering does not simply mean adding more AI agents. A graph may coordinate: - One model - Several models - Deterministic software - External APIs - Databases - Human decisions - Approval gates - Security policies - Queues - Scheduled processes The central question becomes: How should intelligence, state, responsibility, and control move through the system? From employee to organization Return to the workplace analogy. Prompt engineering gives one employee an instruction. Context engineering gives the employee the correct information. Harness engineering gives the employee tools and procedures. Loop engineering allows the employee to continue until the work is complete. Graph engineering designs the organization: - Who plans? - Who performs research? - Who builds? - Who reviews? - Who has authority to approve? - Where is information stored? - Which tasks can run in parallel? - What happens when workers disagree? - Who handles failures? - How is the final result assembled? This resembles organizational design, business-process engineering, and software architecture more than traditional prompt writing. Parallel work Some tasks become faster or more complete when they are divided. A research system could send separate workers to investigate: - Market size - Competitors - Pricing - Regulations - Security risks - Customer sentiment The results can then be returned to a lead agent, reviewed, and assembled. Parallelism can improve coverage, but it introduces new problems: - Duplicate work - Conflicting findings - Uneven quality - Higher cost - Context fragmentation - Difficult debugging - Coordination overhead More agents do not automatically produce more intelligence. Sometimes they only produce more meetings. Shared state A graph needs a clear source of truth. Without shared state, one agent may not know: - What has already been completed - Which evidence has been verified - What another agent discovered - Whether a human approved an action - Which version is current - What remains unresolved - How much budget remains Shared state might contain: Goal Current plan Completed tasks Pending tasks Verified facts Unverified claims Source records Generated outputs Budget remaining Approval status Errors Final decision The quality of the graph depends heavily on the quality of this state. When state is hidden, vague, duplicated, or inconsistent, the system becomes difficult to understand and trust. Conditional routing Graphs are useful when different situations require different paths. A security workflow might look like this: Incoming alert ↓ Initial triage ↓ False positive? Yes No ↓ ↓ Close Collect evidence ↓ Account compromised? Yes No ↓ ↓ Disable account Continue investigation ↓ Human approval ↓ Recovery process This combines deterministic rules with model-driven judgment. The model does not control everything. It operates inside defined boundaries. Human approval is a node, not an inconvenience In high-impact systems, human participation should be intentionally designed into the graph. A human may need to: - Approve an external communication - Review a financial decision - Confirm a containment action - Resolve contradictory evidence - Authorize a data change - Accept legal or compliance risk - Override an automated recommendation Human review should happen where judgment and accountability are required, not randomly at the end. That is one of the clearest differences between a demonstration and a production system. A demonstration asks: Can AI complete this task? A production system asks: Can the task be completed reliably, securely, economically, and with clear accountability? A Practical Example: Building an AI Newsletter System Let us apply all five layers to a real workflow. Imagine you want AI to create a professional weekly newsletter about enterprise artificial intelligence. Stage One: Prompt You write: Create a 2,000-word newsletter about the latest developments in AI agents. Make it professional and practical. The model may produce an attractive article. But the information could be outdated, repetitive, incomplete, or unsupported. Stage Two: Context You provide: - The publication’s style guide - Previous newsletters - Target-reader profile - Trusted sources - Recent announcements - Preferred structure - Topics already covered - Reader feedback - The desired tone The output becomes more relevant and consistent. Stage Three: Harness You add: - Web-search tools - Official-documentation access - Source-quality filters - Citation requirements - Fact-checking - Word-count validation - Formatting rules - Logging - Evaluation tests Now the system is not simply writing. It is executing a controlled publishing workflow. Stage Four: Loop You define measurable criteria: - At least five reliable sources - No unsupported claims - One strong opening hook - Three practical examples - A clear conclusion - No repeated sections - Final length between 2,000 and 2,500 words The system researches, drafts, verifies, revises, and stops when the criteria pass. Stage Five: Graph The workflow becomes an editorial organization. Planner Chooses the angle and creates the outline. Research agents Investigate announcements, technical documentation, risks, and business impact. Analyst Identifies the larger pattern connecting the findings. Writer Creates the first draft. Fact-checker Matches factual claims with sources. Editor Improves clarity, structure, pacing, and tone. Visual planner Suggests diagrams, cover concepts, and image positions. Human publisher Reviews and approves the final article. Each component reads from and writes to shared state. Failures return to the responsible node instead of forcing the entire workflow to restart. That is graph engineering. When Should You Use Each Layer? Not every problem needs the entire stack. Use prompt engineering when: The task is simple, one-time, low-risk, and all required information fits inside one request. Typical examples include rewriting, summarizing, brainstorming, explaining, and drafting. Add context engineering when: The model needs private, current, historical, or organization-specific information. Add harness engineering when: The model uses tools, takes actions, follows output schemas, or requires validation, permissions, and logging. Add loop engineering when: The task needs multiple steps, testing, retries, progress checks, or failure recovery. Add graph engineering when: Work should happen in parallel, different roles require different tools, conditional routing matters, shared state is essential, or humans must approve selected actions. The correct objective is not to build the most complicated system. It is to build the simplest system that reliably completes the job. The New AI Engineering Skill Stack Prompt writing is not disappearing. It is becoming one skill inside a much larger discipline. The emerging AI engineer may need to understand: - Task and workflow design - Context retrieval - Context compression - Memory architecture - Tool design - API integration - Identity and permissions - Data governance - Structured outputs - Evaluation - Observability - Failure recovery - Human-computer interaction - Cost and latency management - Security - Multi-agent coordination This changes how organizations should think about AI talent. The person who writes the cleverest prompt may not be the person who designs the most reliable system. Production AI requires people who can connect models with software engineering, data architecture, security controls, business processes, and human accountability. A Copyable Blueprint for Your First Agent System Before building an agent, complete this design sheet: 1. GOAL What exact outcome must the system produce? 2. USER Who will use or receive the result? 3. INPUTS What information enters the system? 4. CONTEXT POLICY What should be selected, summarized, excluded, or remembered? 5. TOOLS What can the model read, calculate, create, modify, or send? 6. PERMISSIONS Which actions are automatic, restricted, or prohibited? 7. OUTPUT CONTRACT What structure, evidence, format, and quality are required? 8. VERIFICATION How will the system test whether the result is correct? 9. LOOP RULES When should it retry, escalate, or stop? 10. GRAPH DESIGN Which tasks are sequential, parallel, conditional, or human-controlled? 11. STATE What must remain visible across the workflow? 12. OBSERVABILITY What prompts, decisions, tool calls, costs, and errors will be recorded? 13. EVALUATION What test cases will expose failures before deployment? 14. FAILURE PLAN What happens when a tool fails, evidence conflicts, or the model is uncertain? 15. SUCCESS METRIC How will you know the system is genuinely useful? This blueprint is more valuable than collecting hundreds of generic prompts. It forces you to design the work, not merely describe it. The Most Important Warning The industry is moving toward larger and more autonomous systems. But complexity is not free. Every new tool, loop, branch, memory source, and agent introduces another possible failure point. A graph can become so complicated that no one fully understands it. A long loop can consume a large budget without improving the result. A memory system can preserve incorrect assumptions. A powerful tool can turn a small reasoning error into a serious operational mistake. A human approval step can become meaningless if the reviewer is overloaded or shown insufficient evidence. This leads to one of the most important principles in AI engineering: Add complexity only when you can explain which failure it prevents. Do not add memory simply because every agent seems to have memory. Do not add a loop because autonomy sounds impressive. Do not add five agents when one model and three deterministic functions would be easier to operate. Do not create a graph until the workflow itself is understood. Build the smallest reliable system first. Expand it based on evidence. What Comes After Graph Engineering? Graph engineering may not be the final layer. As agent systems mature, the next challenges may include: - Agent governance - Machine-readable organizational policies - Cross-agent identity and trust - Automated capability management - Continuous evaluation - Dynamic restructuring - Auditable autonomous operations - Coordination between organizations - Self-improving workflows Organizations may eventually manage AI systems less like isolated software features and more like digital workforces. But the fundamental questions will remain: - What is the goal? - What information is trusted? - Who can take action? - How is success measured? - Who is accountable when the system fails? The technology will continue changing. The need for sound engineering will not. Final Takeaway Prompt engineering taught us how to communicate with a model. Context engineering taught us how to prepare its working environment. Harness engineering taught us how to surround it with tools, controls, validation, and visibility. Loop engineering taught us how to make it continue until the task is genuinely complete. Graph engineering is teaching us how to coordinate complete systems of models, software, tools, memory, and humans. The model still matters. Its reasoning ability, speed, cost, and reliability can significantly affect performance. But the model is becoming only one layer of the product. The durable value increasingly lives in: - The context you assemble - The tools you connect - The controls you enforce - The evaluations you run - The feedback you preserve - The workflows you design - The organization you build around the intelligence The first era of generative AI was about producing impressive answers. The next era will be about building dependable systems. And that is why the future of AI engineering may belong not to the person with the cleverest prompt, but to the person who can design the best environment in which intelligence operates. The next generation of AI products will be built around systems, not isolated prompts. Subscribe to Open Cloud AI for practical breakdowns of the tools, architectures, workflows, and engineering ideas shaping that transition. Which Layer Are You Building At? Are you improving individual prompts, connecting private context, building tool-using agents, creating reliable loops, or designing complete multi-agent workflows? Share your current use case in the comments. Open Cloud AI Practical intelligence on AI, cloud systems, agents, cybersecurity, and the workflows shaping the future of technology.

How I Built My Roadmap to 1,000 Paid Subscribers in 10 Minutes The 6-step method that turns a generic playbook into a plan built on your own numbers Most growth plans you download are written for nobody in particular. They tell you to post daily, build funnels, run promotions. All true. None of it accounts for your open rate, your conversion rate, where your subscribers actually come from, or what you already have in place. So you read the plan, nod, and unable change anything. I wanted the opposite. A plan built around my numbers, my offers, my pages, and my weak spots. So I built one, and the whole thing took 10 minutes. Here is exactly how. Access your FREE Solopreneur Success Hub - your subscribers-only comprehensive command center for building and scaling a successful one-person business. I created this all-in-one toolkit for building a profitable one-person business, something I wish existed when I first started, and it saves me 20+ hours a week. Now, it’s yours… FREE! This is the written version of my recording. Watch it if you want to see my screen. Read it if you want the steps. The setup Two tools, one connection. Claude does the thinking. SubflowAI’s MCP connects Claude to my Substack, so it can read my dashboard directly. Subscriber counts, growth sources, open rates, post performance, all of it. That connection is the part that matters. Without it, an AI writes generic advice. With it, the AI is looking at my actual publication while it writes. I set this up after going live with D. Raj about SubflowAI. Halfway through that conversation I realised what the MCP made possible, and this was the first thing I built with it. The 6 steps 1. Create a dedicated project I opened a project and named it “Solopreneur Code 1K Paid Subs”. One goal, one folder, one place where every decision, draft, and update lives. Not a chat I lose in a week. A workspace I return to. 2. Find the best existing thinking I searched for people who had already done what I want to do, and found two articles worth studying: Both from Write Build Scale. Credit to Sinem Günel, Philip Hofmacher and Jari Roomer for writing them. I copied the links. That was the whole research step. Two good sources beat twenty shallow ones. 3. Connect the MCP before prompting I connected Dheeraj Sharma’s SubflowAI to Claude before I asked for anything. Order matters here. Once the connection is live, every prompt I write gets answered against my real data instead of assumptions. 4. Feed it three inputs The prompt had three parts: - The two article links. Study these and extract the model. - My critical pages. About, Start Here, What You Get as a Paid Subscriber, Premium Vault, and my recent post archive. This is what a new reader sees, so this is what the plan has to work with. - My own stats. Pulled live through the MCP, plus a set I exported from the Substack backend myself. Then one instruction. Build me a detailed action plan to reach 1,000 paid subscribers. I need you to study these articles and fully understand the strategies and action plans involve to get to 1K Paid Sub. https://writebuildscale.substack.com/p/how-to-actually-become-a-substack https://writebuildscale.substack.com/p/how-we-hit-1000-paid-subscribers Then review everything I have done at Solopreneur Code (my Substack), and create a detailed action plan that is personalised for my current stage and circumstances. My Solopreneur Code: Landing: https://solopreneurcode.substack.com/ Hero Post: https://open.substack.com/pub/solopreneurcode/p/the-story-of-how-solopreneur-code-was-born?r=2jvjts&utm_campaign=post&utm_medium=web Start Here Post: https://solopreneurcode.substack.com/p/start-here Paid Subscriber Benefits: https://solopreneurcode.substack.com/p/what-you-get-as-a-paid-subscriber Premium Vault: https://solopreneurcode.substack.com/p/premium-vault Lead Magnet - Solopreneur Success Hub: https://solopreneurcode.substack.com/p/solopreneur-success-hub About Page: https://solopreneurcode.substack.com/about Flagship paid product - Solopreneur OS for Claude: https://solopreneurcode.substack.com/p/solopreneur-os-for-claude Let me know if you need more information. 5. Add your own judgement The first plan was good. It was not finished. I wanted more Substack Live in it, because video and audio is where I think the visibility is right now. So I asked for a live section. Then I asked a better question. Based on where my subscribers and followers actually are, when should I go live? It looked at my audience split across the Americas, Europe, and Asia, and gave me two time slots where almost everyone is awake. I would never have worked that out by hand. My first show starts this week in one of those slots. This is the part people skip. The AI holds the data. You hold the intent. A plan built from only one of those is half a plan. 6. Keep it as a living document The plan sits in the project. As I run it, I mark off what is done, log what worked, and change what did not. It is not a document I generated. It is a document I maintain. What came out of it The plan told me things I did not want to hear, which is how I knew it was working. My real numbers. Just over 100 paid subscribers against a free list of roughly 6,400. A free-to-paid conversion rate below the healthy range. Nearly all of my growth comes from inside Substack, mostly the app and recommendations, and almost none from outside. That last point reframed everything. The platform already likes my work. Chasing external traffic would have been the wrong project entirely. My stage. There is a three-stage model in those articles: trust from 0 to 100, systems from 100 to 500, scale from 500 to 1,000. I am at the start of stage two. What earned the first hundred will not earn the next four hundred. Stage two is engines, not launches. My real bottleneck. Conversion, not traffic. At a healthy conversion rate my current list alone would already put me near 300 paid. That single line changed my priorities for the year. My honest gaps. Subscriber counts on my some of my own pages were out of date. No video presence. My promotions happened on impulse rather than on a calendar. The engines to build. Daily Notes with weekly batching. A fixed engagement habit of likes, comments, and one real conversation a day. Recommendation partnerships grown deliberately rather than by luck. One or two lives per week at the times the data picked. A funnel line at the close of every free post. A ten-minute metrics log every Friday. A promotion calendar. Three to four sprints a year, each with a start date, an end date, and a reason to act now. Silence in between. Dated milestones. Not a vague hope of 1,000 someday. Specific targets with specific dates, working backwards to what has to happen this quarter. What this cost One afternoon. Two article links. A live MCP connection. My own stats. One clear instruction and a few follow-up questions. The output is a plan I would have paid a consultant for, except it knows my actual numbers. Try it on your own goal The pattern is not about newsletters. Swap the goal, keep the structure: - Make a dedicated project for one goal - Find two strong sources, not twenty weak ones - Connect your real data before you prompt - Feed it sources, your assets, and your numbers - Push back with your own judgement and ask better questions - Maintain the plan as you run it Skip step three and you get advice. Include it and you get a plan. You’re doing everything. But nothing is moving? You are doing everything. But nothing is moving. That is not a motivation problem. Most solopreneurs are learning from everywhere and getting nowhere. Too much information. No clear system connecting effort to results. You have everything it takes. You just do not have a clear system yet. That is what paid subscribers get. Every system, playbook, prompt, and template. All inside the Premium Vault. All for $79/year. That’s $6.58/month. Upgrade now and unlock the Premium Vault worth thousands of dollars. The Premium Vault holds the secret behind posts like this one, including the tools and resources I use to build the one-person business I love. Watch the recording if you want to see the screens. Otherwise, you now have the whole method. Every time you open Claude, you re-explain your business from scratch. Your niche. Your voice. Your offers. Ten minutes lost before it writes a single useful word. Solopreneur OS for Claude fixes that. One setup interview. Then 16 skills covering positioning, content, sales pages, launches, weekly reviews that all read your profile before every output. Validate → Build → Sell → Review. One system, one voice, one business. Thanks for reading! Ready for the next step? Let’s crack the growth equation and build a thriving one-person business on your terms! Anfernee