CODEX FULL COURSE: From Zero to Deployed App (2026)v

Key Takeaways

This is the ultimate Codex guide for beginners. My name is David Andre and I've spent well over a thousand hours coding with AI. I also used Codex to build Vectal, my AI startup, which got acquired last year for seven figures. And I'm one of the first people on YouTube to begin teaching Codex. I've helped hundreds of thousands of people learn how to use Codex. So, if they can do it, so can you. Now, to start using Codex, you don't need any coding experience. I'm going to walk you through everything step by step. And by the end of this course, you'll be able to build anything. You will have the skills to turn any idea into a real software application that's fully deployed on the internet in less than an hour. Codex will also make you a lot more productive. And not just in coding, in most areas of life and business. And I know that because I'm speaking from experience. So here is what we're going to cover. First, the fundamentals. installing Codex, setting up your environment, my personal codec setup, and understanding how it works within IDE. Then we'll build a real app together from scratch and deploy it live so it's available on the internet to your friends, family or potential customers. And we will do all of that just by speaking in plain English to Codex. After that, we'll go into the advanced stuff. Sub agents, skills, automations, git workshries, MCP servers, cloud agents, and more. Yeah, this really is the ultimate Codex guide. So, let's get into it. Now, there are actually four different ways to use Codex. But the first one I'm going to show you, which is the one that absolutely anybody can use, no matter how bad your computer is, is the Codex CLI. This is the official documentation by OpenAI. I'm going to link it below the video. And luckily, there's a simple oneline install command, right? So if we copy this and open your terminal, which you can do by typing terminal on spotlight search on Mac OS or on Windows, press Windows plus R and type in cmd. So open the terminal, copy this command and paste this in. Boom. This will install Codex CLI on your machine. And yes, it really is that easy. After that, what we need to do is simply type in the word CEX to launch the Codex CLI. So first, it shows you that there's update. So obviously we want to use the latest version. So we're going to update. And usually these updates are very fast. There we go. Those 2.4 seconds. So I'm going to type clear again and type Codex again to launch the Codex CLI. Now it will want you to login and we have three different options. If you have CHBD subscription, that is the best, right? It works on Go as well. That's like $8 a month, but ideally you would have plus or pro. But again, even the $8 a month subscription works. If you don't have a chair GB subscription, I highly recommend you getting one. It's one of the most efficient ways to spend money in AI. In fact, I think this $20 subscription is probably the single most valuable $20 subscription in all of AI. Also, you can see that there's a new plan, Pro. Before, the Pro plan costed $200, but now they have two different versions, $100 and $200 a month, just like Cloud Code. You know, OpenAI got inspired by Enthropic. So, if you want a bit more usage than the plus plan, this is the best time in history to begin using Codex because the pro plan now starts at $100 a month. That said, if you don't have any of the plans, you have two more options. Signing up with device code or API key. Now, honestly, device code only works if you have another device that's signed in. So, really the only two relevant options are signing in with your CI GBD subscription or the API key. Now, since most people already have CHIGD subscription and if you're watching this channel, hopefully you have some form of Chad GBD subscription, I'm going to show you this. So, hit enter and it's going to redirect you to a website where you need to authenticate with the same account you use as in Chad GBD. Now, if you have multiple subscriptions, for example, I have a CH GBD teams account where I have, you know, my team, my people, my employees, and then I have a personal one which I have the $200 a month pro plan. If that's you, then just select the plan which has higher usage. So for me, obviously the pro plan is the highest plan. So I'm going to select that and click continue. Boom. There it is. Sign into COX. We can close this page. So if we open the terminal again, we should be logged in. Sign in with your CHP account. Okay. So we can hit enter and we should be able to access the CEX agent. There we go. And let me try sending a prompt to see if it works. Hey, there it is. Who are you? So we are chatting with Codex through the Codex CLI. This is the command line interface and this is like the OpenAI's competitor to CL code. But Codex is without a doubt more powerful at building complex applications, solving deep errors and running for longer without making mistakes than cloud code. So, if you aren't using Codex in 2026, you really are missing out because Cloud Code, while it's good, it's not as powerful as Codex at many of the use cases. And by the way, if you're having any troubles with the Codex install if the oneliner doesn't work or, you know, is is throwing some errors, then you might need to check if you have Node.js installed on your system. So, again, I'm going to link this below the video, but this is super simple. Here, you just select the latest version. You you don't have to change that. here. Select your operating system and then run these terminal commands one by one in your terminal. Now to check if you have Node.js, just open the terminal and type in Node-V. And there we go. This is how you know whether NodeJS is installed on your system. If it's not showing any version, then you need to install it. That said, let me show you what I've built for my team and my business with the help of Codex. So, here is one app. I call it YouTube alpha that we use for titles, thumbnails and analyzing outline videos. As you can see, this is very advanced way to create uh thumbnails. You upload an image and you can do prompts and based on Nanobana 2 and Nanobana Pro, it generates variations and you can get to a thumbnail like that. Something that inside of Google Studio, it would take you a lot of time because it constantly fails. It can only generate one variation at a time. And we have these different canvases that you can share with other people on the team that allow you to simply create variations of uh different thumbnails. And this is super efficient and we use it for every single video. And uh it was built largely with Codex. Here's another application that we use for tracking some of our funnels and seeing the metrics. It's integrated with type form and Calendarly. It was built mostly with Codex like probably 90% Codex, 10% cloth code. And this is allowing us to see where the biggest drop off is in our business and what we need to fix. Also known as datadriven decisions. So stuff like this is absolutely invaluable to me as the CEO. And these are the types of applications that we're able to build in one two three hours and get it fully deployed on the internet with the help of Codex. So if you stick through this entire video, you will have the skills to build app like this for your own life and business. And the obvious advantage of this is that number one, you don't have to pay for other people's software that might be mediocre and not specialized to your use case. And number two, you can turn any idea you have into a real application in a matter of couple hours. This is the power of Codex. Okay, so let's go back to the terminal. I'm going to type in Codex again to launch it. And the very first thing you need to do is select the AI model. Now, you can see that I already have GBD 5.4 for selected. But if you don't, just make sure to type in /model. Boom. And make sure to select the latest one. Okay? Do not use GPD 5.2. Do not use 5.3 Spark or 5.3 Codex. And definitely do not use a mini model. Like, god forbid, please do not use mini models. Use the best model, okay? Which is the latest one, which is GPT 5.4 as of the time of this recording. If you watch this maybe in a month, in in two months, maybe it's GPD 5.5. Whatever is the latest model by OpenAI, you'll see it right here in the description. They tell you what is the latest model. So use the best one. Hit enter. And then you need to select the reasoning effort. How much time does it spend on thinking, right? So there's four options. Low, medium, high, and extra high. Medium is default. And I would say that's actually a good default because for most things, you don't need it to run for many minutes, right? Low I would suggest you never use. This is simply uh too little reasoning and it will underperform. So I would never use low. Medium is good default especially if you don't want to burn your limits too fast. High is also a great starting point. Most of the time I do have it set to high actually. And the beauty of high is that if it needs to, it can run for two, three, four minutes if it if the task requires it. But if it's a simple change, it'll just do it in five seconds. Now then we have extra high. This I would only use for the most complex errors, right? When you when you're struggling with a deep bug that you haven't been able to fix, use extra high. Or if you're doing a massive refactor, like a very risky refactor of your entire back end, use extra high. But for most things, I think high is the sweet spot and that's what I'm going to select. Now, the way we did it right now, Codex is running on the root of my machine. So, if I tell it where on our MacBook are you running, it's going to run the pwd command and it's going to check where it's running. You can see it's user/david. So, this is the root level of my MacBook, which is not recommended unless you want to do changes like cleaning up your downloads folder or deleting some old applications. Yeah, for these it would be good running it on root level. But if you want to build something like an application, you know, do some AI coding, it's much better to launch it in a specific folder. So what I'm going to do is I'm going to press Ctrl C a few times and this will kill the process. Right? So again Ctrl C even on Mac OS not command C and this will kill COX process and I'm just going to type clear and then we need to go into a specific folder which means changing the directory cd. This is a very simple terminal command that all of you should learn. Even if you're not a developer there's no disadvantage to knowing this command, right? So CD and you can go for example to documents and if you press tab it's going to prefill the rest of the folder name. Right now if you want to go back up you can do CD dot dot and it goes one level up. So to go to a specific folder just type in CD and the folders name. But if you want to go back up type in CD dot. Now you might be thinking okay David but what if I don't know all the folders right? We're on the root level. I just open the terminal. I don't know where this is located in my MacBook. First if you want to know where this is located. Type in pwd. This is what Codex did to check where we are. Let's do clear. But if you want to see what is in this directory, like what are the different files here and folders, you can type in ls and this will list out everything in this directory. So you can see that we have desktop documents downloads. So for a project like this, you probably want to go into documents, right? So you do ls to see everything on this level, all of the folders that are available. And you do cd and you start typing the ones you want to go to. Boom. And now I'm in documents. Let me type in clear again. LS to list out everything in documents, right? There's a lot, but I know where I want to go to. I'm going to go to CD David. Boom. David. That's my personal folder for my YouTube channel. And then I'm going to do ls again. Okay. I have a bunch of stuff. So I'm going to go cd into video files. Okay, that's clear. And then I'm going to CD. And then I'm going to ls again. And I'm going to CD into the Codex beginner course, which is the video you're watching right now. Now that we're in this specific folder, we can finally launch Codex with restrictions of it being in this folder. I'm going to do that right now. Codex. Now, it will ask you, do you trust the contents of this directory? Right. So, make sure you're on your computer and you trust it. I'm going to hit yes because it's my computer and I trust it. And this is something that you might think, okay, do I need to be using the terminal? You know, is this too advanced? I literally showed you two commands, ls and cd. ls to list out the files and folders. CD stands for change directory. Go into a specific folder. That's all you need to know because for each project you build, you should be in a separate folder. Don't be one of these people that like saves everything on the desktop. That's completely amateur and you should avoid that at all cost. Just have a specific folder for each coding project you're going to build and you're good to go. Now, the very first thing you want Codex to do, no matter what you're building, is to create agents.mmd file. So, I'm going to say create an agents.mmd file in this project and keep it empty for now. Just put a header. This is the ultimate system prompt that all of the agents follow. In fact, this is not just for Codex. This is a global convention used by over 60,000 different open source projects. Basically, it's like a readme file for agents. Now, if you use cloud code, you will know the cloud MD file and that's basically the only project that doesn't follow agents.mmd. all the other agents whether that is AM VS Code, Gemini CLI, Cursor, Warp, Kilo Code, Ader, Factory, Codeex, Rue Code, Windsurf, all the other ones use agents.mmd. And here's what that can look like. It's basically a single markdown file that is a system prompt, right, that the agents should follow. Now, you might be saying, "Okay, David, but how do I know where to start? Like, I don't have agents.md file." Don't worry, I created a free one for all of you. This is a preset that I use that you can just build up on top of, right? So, I'm going to link this as a free GitHub gist below the video and you can just copy it, you know, click, you can click raw, get the raw file and just copy control A. Boom. Then you can go into your codeex in the terminal and say like now put in the following text. Going to wrap it with XML tags like this. This is a pro tip for context engineering. any large paste you're going to do, just wrap it with XML text like this. Boom. Text text and it allows the agent to better see where's the start and where's the end. So, I just literally took the preset from GitHub gist. And as you can see, there needs to be a bit updated like the project name, target user, your skill level, and stuff like that so that it's relevant to your project. But this is already a very decent agentmd file that's going to make the agent behave better, respond in a more friendly way, and follow best practices. So again, this is going to be linked below the video. It's completely free. Take it and use it. And by the way, if you want to access more presets and resources like this that I use myself for coding and that I use to build vectal and sell it for $1.8 million, make sure to join the new society inside of the classroom. We just released a brand new course, a 3-week outline that's going to take you from a complete beginner to master AI coding. That includes cloth code, codex, cursor, deployments to versel, superbase, and anything else you need to truly be able to build any type of software with AI. In these three weeks, you will go from a complete beginner to someone who has the skills to build anything with AI. In fact, let me show you just how the outline looks like. Each week is split into these granular modules between one, three, sometimes four minutes. And these modules are as step by step as it goes, showing you every single step of the setup, all the mistakes to avoid, giving you all the presets, so you can just copy paste them and start using them yourself. And this is designed for people who've never built anything with any coding background, no experience with AI coding agents, nothing. You will learn absolutely all the things you need to know to not only be able to build any type of software you want, but also deploy it on the internet and start getting users. And in these three weeks, you will have the skill set to do that. So again, if you're serious about AI and if you want to master AI coding, make sure to join the new society. It's going to be linked below the video. Okay. So the next thing I'm going to show you is image input. You can actually attach images, screenshots directly into the Codex CLI. Now we should start building something right because right now so far we created agents.mmd but we haven't really chosen a project what to build now for apps the woman targeting apps are insane like it sounds bad but women do spend a lot more money and you can build the dumbest [ __ ] imaginable and it will work. So this girl got to 300k MR with a manifestation app. I don't even know what that means, but I know what 300K MR means. And what I'm trying to say is that this is the biggest opportunity to build apps is stuff like that's not really obvious to us because each instinct, you know, most of my audience like 98% are men. The instinct is to build something for yourself. you know, you you want to build a tool that you would use and you go into these highly competitive industries, but there's so much potential building these mobile apps or or it doesn't have to be a mobile app, could be a web app as well, but targeting women and you don't even need to add that much AI. I guarantee you this app is like the simplest thing imaginable architecturally, but it's just branded in a way that works and it clicks and 300K MR something that will be absolutely life-changing to all of us, right? So we're going to take this approach and actually I have a really good idea which is proven by being in the field let's just put it that way of showing seeing how we would look like. So AI images are much newer, especially since the rise of Nanobana Pro and Nanobana 2, than AI text, right? Most people understand AI text. So what we're going to do is we're going to use the untapped potential of AI images to build an app. And this is something that literally every single woman or girl is wondering how she would look like if she had a bit smaller nose or if she had a lip fill or stuff like that, right? And whether this type of stuff should be promoted or not, that's an entirely different conversation. But I think it's a good idea from the technical standpoint because it's going to allow us to use AI images and it's a pretty interesting project to build. Let's put it that way. So I will say I want to build an app where you can upload an image of yourself and see how you would look like with different facial changes. Smaller nose, lip filler, foxy eye lift, faceelift, Botox, different skin tone, and at least 10 plus other things that women often do or would like to see how they would look on them before doing anything. The target audience is women 18 to there's no limit really but let's say 18 to 55 years old. Update agents.md accordingly and also create readme.md file. This is something you should have in the root of every project is what's called a readme file. This is what displays on GitHub right away. And later in the video I'll also show you how to actually get this on GitHub because that's essential. There we go. Codex has updated it. And now I'm going to do the visual style. So what's another thing that women like is Pinterest. So I'm going to just say that you know whatever this I'm going to copy this. I'm going to shoot a visual input. You can do Ctrl +V. Not command V. Control V. Command V doesn't work. I don't know why. So just do Ctrl +V to attach this image. And I'm going to say this is what the layout of the app should look like after the user uploads her face. It should generate all kinds of different predefined changes to her face. Again, at least 15 to 20 different images. Study this layout and save it into both MD files. Okay. So, this is how image input looks like. You can literally take a screenshot of something maybe of a website you like some style you want to copy even an error or the layout you don't want to you don't like the layout of your current site take an image and attach it with your prompt. This is something that so many people are underutilizing. The ability to attach an image input is very very OP. And by the way if you don't know how to take a full screenshot here is how on Mac OS command shift and free takes a screenshot of the whole screen. So you can just copy and paste it into an AI chatbot or into Codex itself. On Windows it's print screen key. You should have a dedicated key for that or win plus print screen. And on Linux it's print screen as well. So this is how you take screenshots. And when I code with AI, I attach screenshots all the time. Like every other prompt includes at least one attached image. It really is underrated thing that beginners are not doing. Now another thing I'm going to show you is web search. This is one of the built-in tools into Codex, but again, very underutilized. So, I'm going to say use the web search tool to find the 20 plus most common facial cosmetic surgeries women do and save them into a single bullet list into readme.md file. Boom. So, now Codex will browse the web, right? It has a lot of knowledge built in because the model has a lot of it baked into the training data. But for stuff like this or maybe you want to check the latest documentation, maybe you are using like a obscure framework and you want to make sure that you are using the correct docs then tell it to do web search right it can it's not going to be a deep research but the web search here is pretty solid like opens web search is on the level of normal complexity search I would say. So this is another thing is like telling it to do web search to find more information. Don't just rely on it with the training data. It knows a lot in the training data, but for stuff that's up to date, it can do customer research. This is actually less of a coding, more of a marketing. I'm learning more about that my target avatar for this build project to make the app better, right? And this is something you can do when writing a landing page, when writing email copy. Use these AI agents to understand the industry more, to understand your users more, to write better copy, to do competitor research. There's so much you can do with tools like Codex. People think it's only for coding because it's in the name Codex, but it's not. You can absolutely use it for marketing, for hiring, for growing your business in all kinds of various ways, for managing your personal productivity. There's literally an endless ways you can use Codex. Anything that you as a human can do on a computer, you can use AI agents like Codex to help you save time or even do it fully autonomously. Okay, so there it is. It spent a solid minute and now it uh created the list. I would say uh I did not tell you to remove any of the things we already had just to add to it. So please do not remove the stuff I told you earlier. Aka different facial changes. Okay. So here Codex did multiple different web searches until it had enough info and then it saved the results into our files that is much more accurate than what I could think you know women want to see on their face. This is based on multiple web searches. So I'm going to trust that a lot more. Okay. Let me show you another cool feature inside of Codex. Let's say you accidentally shut it off or maybe you don't use the computer for a few days and then you want to go back to previous session but you start Codeex and it's a new session, right? Every time it creates a new session. Well, guess what? There's a command slash resume where you can resume a saved chat. And here it is. We can see that we have this one chat here and hit enter. And this is the chat that we were just in with the web search and with uh yeah, with the app idea that we're building right here. So, anytime you want to go to a previous conversation, just type in slash resume. Now I'm going to ask it suggest 10 potential clear names for this app idea. Let's see what Codex cooks up here. Beauty. I like beauty. I like number two. Let's go with that. Okay. Amazing. So it updated all the references to use this new name. And by the way, not every time is it good to resume a previous session. Sometimes if there's a lot of like irrelevant information, a lot of noise, it's better to start with a fresh session, which by the way, you can also do by typing new or clear. Both of these will start a new chat. You don't have to crash the terminal and start Codex from scratch. Type in slash new or slashcle and that will completely clear everything and begin a new Codex session. Now, what I'm going to do is I'm going to show you just how powerful Codex is by telling it to oneshot this whole app. So, ideally, this would probably be a mobile app, but I'm recording this on a computer, so I'm going to show you a web app. So, I'm going to say, let's build this MVP as a web app. Ask me four concise questions to understand my vision for this based on what's missing in and then tagging files you do to add symbol and then you read me or agents IMD I mean these are the only two files we have right now in this project but tagging files is essential because it tells the agent look there look into this file this is essential context and then I'm going to say be concise because um context has a tendency to be very verbose, you know, and like yap a lot. Okay. So, do you want the MVP to use real AI images generation editing? So, I'm going to use voice prompts. Yes, I want the MVP to use real AI image editing. We will use the nano banana 2 model from open router API. In fact, I'm going to give it the exact model slug. So, there we are. Nano banana 2. Here we are. Here is the exact mortal slug. Save it. I think the first version doesn't need to have any accounts. It should just be upload only. Question number three, we should store this in local storage for now. Question number four, what matters for V1 is all three of these things. Realistic results, fast generation speeds, and beautiful gallery style experience. We will get fast generation speeds if you use the exact models like I gave you above. And the realistic results just means that the prompts we include cannot be too aggressive. And the beautiful gallery style experience, we need to copy the interface and the layout of Pinterest which I sent you earlier. Now save my answers into both MD files. Okay, so now I think I've given enough context. So next I'm going to launch it and I'm going to switch to extra high and I'm going to try to have it on oneot the entire app. I'm adding your MVP decisions to both dogs. Real AI editing. Okay. And actually I'm going to show you one more pro tip. This one is super valuable is how to double the speed of codeex. You can literally make it faster and that is with a single command slashf fast. Okay, so it finished typing. So if you type in /fast, you can toggle on fast mode. Now obviously this will burn through your limits sooner. So if you only have um Go the $8 subscription or the $20 subscription plus, you will uh burn through your limits fast. But if you have either the $100 a month pro plan or the $200 a month pro plan, which is what I have, then just always be on fast mode. There is no reason not to be on fast mode. It's very good. It just doubles your inference speed. So, it's absolutely addictive. And I'm actually going to do /model before I deploy it. I'm going to select to extra high. Actually, no, I think high can handle this. Never mind. I'm going to state on high. I'm going to make sure I'm in fast mode. You can see it also right here at the bottom. I'm going to say now get to work and build this entire app fully and completely like a professional developer would. Again, make it a web app and keep building until the full app has been finished. A very simple prompt in plain English and I'm just going to have Codex get started on this and build a whole app for us hopefully in one shot. Now, while this is going, I'm going to explain to you how to use Codex with your own IDE. IDE stands for integrated development environment. And this is what you need when you're building software. Now, obviously, for simple stuff, for a simple demo, you can just stay in a terminal like this and have the agent do everything. But the more advanced, the more complex your software gets, the more need for a proper ID. You can see the code, you can see your file structure, you can use Git easily, you have integrated terminals. There's many benefits of using an IDE, and there is no reason to be scared of it. It's just like any other app. So, if you're able to install Slack or Microsoft Teams or a video game on your computer, you can absolutely install and set up an IDE. It's that simple. Now, if you don't know what an IDE is, as I said, it stands for integrated development environment. And a simple explanation is that's a software application that combines essential developer tools into a single interface. So, this is a great article from GitHub. You can read it. But basically, IDEs have code editors, debuggers, compilers, version control systems, and AI capabilities. But the beauty of Codex is that it can do most of these things for us, right? So we're not going to be debugging or compiling or doing the version control. Codex can do these things, but you still want an IDE to see the code and to understand it. Now, by far the most popular ID in the world is Visual Studio Code, also known as VS Code. And this is a great IDE because not only is it open source, it's also completely free. And u everything is built on top of it, right? So if you use windsurf, anti-gravity, cursor, all of them are built on top of VS code. So if you start using any of these idees, it's very easy to understand other ones because they are all forks of VS code. In fact, you can literally go to the VS code GitHub and you can see it has 184,000 stars and this is the official repository by Microsoft. So they are the inventors, creators of VS Code, but this is one of the most popular open source projects of all time. And if you have an brilliant idea for a unique AI vibe coding tool, you can just fork it. You can see there is many many forks of VS Code and that's because it's not only really really good but also very easy to use. Now my favorite fork of VS Code is actually cursor. This AI powered IDE that right now they have some features that VS Code doesn't have but uh yeah I've been using Cursor for like two years over two years now and it's very easy to use and you can also get started for free. You don't even need to pay for any of these subscriptions because we will be running codecs inside of Cursor. And they actually complement each other very nicely. So I'm going to click on download for Mac OS. And by the way, just go to cursor.com to download it. And by clicking this button, it will begin downloading the cursor installer. Boom, there it is. So double click on the installer and simply drag in cursor into your applications file. Then you can just type in spotlight search and type in cursor which will launch the cursor app. Now, if you're doing this for the first time, it might want you to log in and create a cursor account. But again, you can do that completely for free on their website. So, just do that. And then you'll see a screen like this where it's telling you to open a project, right? So, I can open project. And actually, I'm going to open this same project right here. Codex beginners course. Uhuh. I'm going to open this. And this is now we can see, you know, on the right. By default, it's on the left. So, by default, you're going to have it like this. Okay. Boom. This is what it's going to look like for you by default. In the left, you have the files. In the middle, you can open any of these files and see the actual contents. And on the right, you have the cursor agent. As you can see, I have Codex here set up, but I'm not going to get too ahead of ourselves. Right. So, this is the layout, but personally, I prefer to have this on the right. So, you can just right click in the header of the sidebar and click move primary sidebar to the right. So don't be confused when the sides are swapped. This is how I prefer to have it. But you can already start to see the benefits of having an IDE. Not only do we see the actual codebase that we're building, we can see the structure of it and what folders we have, but also you can click on any of these files and you can see the code for yourself or maybe see the prompt, right? Agents.md see the prompt and you can change the system prompt easily. And if you do command I, you can close the AI sidebar to have a full vision. If you do command B, you can close the primary sidebar. But there is many benefits of using an IDE. But one of the main ones in the era of AI coding is the integrated terminal. To start out, just do command J or control J on Windows. And this will open the terminal inside of cursor or VS code. Again, all of this is literally the same for VS code or wind surf or anti-gravity. All of these things will be the same. So if you learn the basics of VS code, it's so easy to switch between these different IDEs. Now here what we can do is actually we can type in codeex and launch codex in this project right and then here if you drag in on the node and drag it to the top we can have it in the main window so we can run multiple codexes inside of cursor and this is both better for AI because again anytime you launch the integrated terminal it's already in the correct folder you don't have to do cd and ls and navigate into the right folder it's already in the correct folder so any agent we launch, it's going to be working on this project. It knows that we're working on this project. But another benefit is that for you as a human, you see everything and you can easily manage agents and you don't have to, you know, wonder what they're creating and what files you have and don't have. It's much better for everybody involved basically. So, let's switch to the terminal and let's see what's happening. It wants to allow installing. Okay. So, I'm going to do yes. In fact, I'm going to put you on more of a tip, more of a game. How to launch Codex so you don't have to do these constant approvals. And that is by putting it in YOLO mode. So, if you type in Codex, you can do space dash yolo. This starts it with full permissions. It's like dangerously skip permissions for cloud code, right? If you're familiar with that command, cloud code, dangerously skip permissions. This is the same command but for Codex. So, I'm going to drag it up and we're going to be using Codex here. I can kill this process and let's see what's happening in this terminal. The app builds verification pass two issues current tool. Okay, so the llinter found some issues. So it's fixing them by itself. So it seems like it built a full app in one shot. Look at this. Hundreds and hundreds of lines of code without me having to intervene or guide it or anything. This is the power of Codex. It is the most persistent AI agent that can run for longest and it makes the least mistakes. So I'm not joking when I say people who don't use codeex are missing out because it really is essential for building software. Okay. So now it's fixing some of these issues and we should be able to start the app. So I'm going to say study the code base and tell me how I can start this app. Again with AI you can be as stupid sounding as possible. No prompt is too dumb. The worst thing you can do is not send a prompt and not ask. Right? So it's going to analyze the codebase. It's going to see what text tag we have and see okay this is the next JS project. So we need to do npm rundef. Can the app start even without the envir let's ask that question. Yes the app can start. Okay let's see if the terminal is running here. It build the beauty mirror like it ran for like 8 minutes or something. The app is okay. Is the front end actually what I'm going to do is I'm going to kill the terminal here. I'm going to resume the same codeex but inside of cursor inside of the IDE. So again press command J to launch the integrated terminal and I'm just going to do arrow up. This is by the way a pro tip. If you do arrow up it goes through the latest the terminal commands that you already typed in. So you don't have to type in the same stuff over and over. I'm going to do d- yolo again. Boom. Drag it up again all the way to the left. And actually for your convenience you can rename it. So you can do like rename codex 01. Boom. Let's go here. You can right click, rename codex 02. And here I can do / resume to resume the chat that we did earlier. Here we go. And now we have our own chat which we were running in the Mac OS dedicated terminal, you know, like this. But now we're running it inside of the IDE. So it's uh all in one place. It's way way more convenient, way better. So, I'm going to say uh is the front end running? Ask a super blunt question checking whether it's running. It gave us these steps, but it should be able to do all of these steps by itself. No, nothing is listening. Okay. So, I'm going to say start locally. I'm going to copy this whole thing. There we go. I'm going to copy that. I'm going to say please do these steps yourself. Steps. Boom. Steps. I'm going to wrap it up with the XML tags. Most of these it should be able to do itself. U except for the open router API key. We're going to do that in a moment. But I wanted to launch the app to see what it looks like. Okay, so it created an inv key and let's see if I can start the server. I'm actually going to start myself npm rundev go local 3000. There we go. This is not bad at all. Wow. Okay, so in the next section we're going to add open router and we're going to actually start testing this. Okay, so the app is running and we can upload a file. This works nicely. However, before I add the opener API key, I want to show you one more thing and that is another way to use codeex inside of an IDE and that is the IDE extension. So, inside of VS Code and most other IDEs, you will see this button right here. This is the extensions button. When you go here and type in Codex, you'll go to the marketplace and you will find the OpenAI codeex extension. It should have over 3 million installs and make sure to install it, right? So it's completely free and it works the same way. You can use the same account and this is another way to use Codex. Now you might be thinking okay David but why do we need this? Why are we like why are you showing me the extension when we already have the CLI? Well inside of cursor if you press command I or control I on uh Windows you'll see the primary agent sidebar or the secondary sidebar whatever you should see button at the top once you install the extension. Make sure to click enable. Right? If you don't see it, you should click these three dots or these three dots and click open C codeex sidebar. So this is this right now. Here we can use CEX just the same way but with a nicer user interface. So if you prefer to use CI GBD and you kind of don't like the CLI, you know, which a lot of people don't like the terminal interface because, you know, it's not as beginner friendly, then the extension will probably be more for you. In fact, the extension is my favorite way to use Codex. And again, there's four different ways to use this. And I'm going to show you the other two. So far, I've only seen two ways to use Codex. The CLI command line interface and the extension. But I'm going to show you more two more later in the video. So here, you can use Codex the same, right? We have the models right here. Select the model, select the reasoning effort, and you can type in /fast to turn off or turn on fast mode. Obviously, I do want to turn it on. So this is how we can use it with uh two agents at once, you know, one running in the CLI and one running in the extension. So let's add the open router API key. That way um the app actually works. So what remains just adding the open router API key. Yes, we're local. Okay, there we go. All right, so we need to do that. So, let's go to open router.ai. Click on top right. Make sure to create an account if you don't have one. Again, you can get started completely for free. Then, when it comes to credits, charge up couple dollars. You don't need $300 or like I have. Just $5, $10 is enough. And the beauty of open router is that you can use any model you want, right? So, let's say a new model drops. Okay, we have entropic opus 41.6 fast. That is crazy. I didn't know that's, you know, open router. Well, we can use this or we can use GLM 5.1 which is one of the best open source models right now or maybe the latest version of Gemini 3.1 Pro or if you want GPD models, GBD 5.4. The point is Open Router has all of the models. So, you only need a single API key for your app to work with everything. Now, we specifically want to use Nano Banana 2. This is uh Google's fastest image editing model and it's much cheaper than Nanobana Pro and much faster than Nanobana Pro. So this is the official name Gemini 3.1 flash image preview but it's also popularly known as Nanobana 2. So again, create an account. Go to top top right on the settings. Click on settings. On the left, click on API keys. And then click this blue create button in the top right. You can name it something like a beauty mirror API key. You know, it's the name of our app. Let's put some limit like $20. This is good practice. Always put a limit. If you use something more, then just put like a weekly limit, you know, if you don't want it to run out. But definitely put a limit on every API key. It's good practice. then create and never share your API keys with anyone. Okay, it's like passwords. Keep them private. I'm going to rotate my API key before uploading this video. But for now, I'm going to copy switch back into cursor and we need to put this in the env. So, uh, env. Okay, I'm just going to send it to agent. This is not the best practice. Here's the API key. Create the env file. This is definitely not best security practice, but it is the simplest. I'm just going to send it to agent and he's going to replace it in the env.lo, which we should be able to find if we close some of these things. Yeah, there it is. Yeah. So, this is why you also should u use an IDE because you can actually access these files. Also, I'm also going to say create a detailed git ignore for this project because we are getting to the point where this makes sense. uh to throw this up on GitHub and we should actually initialize a G repository even before uploading this to GitHub which we can use our extension for start a new Git repo in this project and do the first commit. Boom. The agents already know what Git is. They know how to work with it and um yeah they know they know what to do and you can use Git even without GitHub. So even when you're building a project locally and you're like, okay, this, you know, I don't know if this is going to continue. I don't know if it's worth throwing up on GitHub, you should still use Git, right? So Git and GitHub are not the same thing. They're not the from the same company. Git is a open source thing, open source project that is built for version tracking. It's created by LOS, the same guy who created Linux, and you can absolutely use it locally. It's for version control. You should use it in any project. In fact, I should have created the Git repo way sooner. There's no point to wait. Even if you don't plan uploading it and pushing it to GitHub, you should absolutely use Git for every project. Anyways, uh let's I guess reload and let's see if this works. So, what I'm going to do because I don't want to disrespect anybody, I'm going to take a AI girl New York. Okay. So, I'm going to zoom in a lot. I'm going to screenshot her face because that's what our app does. Save it into downloads. Save to documents. There we go. And then let's go back to our app. Let's reload this. And let's choose a selfie here. There we go. So the image is on the left. And uh we can select. Okay, these are already all selected. Okay, let's click on generate beauty mirror. Let's see how this look with fuller lips. Smile line softening. Okay, generation failed. We're getting a bunch of failures. So, I'm going to screenshot this. Actually, great. I'm glad it didn't work on the first try. I'm going to screenshot this. Copy to clipboard. Back to cursor. And I'm going to paste this in. And by the way, you can also paste in into the uh extension. It looks better. But, uh I'm going to paste it into the same codeex because um it has all the context. Or wait, was it this one? Which one was the builder one? Okay. Yeah, this one. 01. That's the builder one. I'm going to paste this in and say investigate why I'm getting this error and fix it. So obviously we can also open the console with F12 or command option I and see if there's any errors here. There are not. We can close that. Uh um maybe we can look into the terminal. I don't know. The terminal is running in this guy I think. Okay. Yeah. One background terminal running. So I should have probably sent it to the same uh same agent because this one is seeing the outputs of terminal. Do you see any errors in the terminal logs? But right now I'm not debugging in the best way because I'm running checking the current dead server directly. If there any runtime errors they should show there what I see. H maybe we can reload. No generation failed. Everything is getting generation failed. Okay. I'm going to let this agent run. Okay. Actually this is good. So to debug this, I would use perplexity deep research. Research the proper way to use nano banana 2 model inside of open router. And then I'm going to give it the exact slug model slug. Right? So nano banana 2. Here we go. This exact model. Boom. And how to properly attach images to it and send text prompts. Check the official documentation and give me a detailed report as a single markdown file. So probably what happened is um Codex messed up the open router formatting. So I'm going to do a deep research and you can use CH GBD research. You can use Gemini DB research doesn't matter. I think perplexity has the right balance of depth and u quickness. So I'm going to use this to figure out what's happening. This Codex is running. It's doing some web search itself. But again web search is not deep research, right? This is good for a quick info but not a docs. So actually I'm going to say do we have a good place to store different docs or not? I would probably create a /doc folder. Okay. So I'm going to say I'm going to switch to model and medium. Create slashdocs folder on the root level. And in there create open router nano banana to syntax.mmd. Keep the file empty for now. Just add header. And I'm going to put in the output of the deep research in there. And I'm switching to medium because u most changes don't require high. And now what we can do is do we can do command P or control P on Windows to look up this specific file. Right? So I can do nano. Boom. Here we go. Open our run two. And we can open it directly. This is the beauty of an ID. You can open these files directly and paste stuff in. Right? So we don't have to interact with the agent every time because perplexity here is going to give me the results of a deep research. And I just want to paste the results into the file, not have the agent rewrite them. Okay. So it's creating the markdown file. Here it is. We can click on download. Download as markdown. Boom. Actually open that. Double click. It's going to open in some text editor. It chose anti-gravity for me. It doesn't matter. Can even use notepad. But I'm going to copy the contents. I'm going to replace this here. Now I'm going to say read. I'm going to tag the file. Open it. Now there it is. I'm going say and analyze our codebase to see what is different and fix it. So this contains 520 lines of official documentation from open router on how to use this specific model and how to attach images and uh yeah now I'm just going to dispatch codeex to fix it right so this is one of the essential prompt engineering techniques is providing enough context another thing about prompt engineering especially for AI agents like codex is clarity you need to be clear in what you want a lot of you are being too vague now you don't like the mid mid with curve right okay this meme I want to show you this meme because it's exactly applicable to prompt engineering the people here the complete beginners complete dummies they are too vague their prompts are too short and they don't tell the AI agent enough the agent doesn't know what the user aka you is trying to do agent is confused in the middle the people around 100 IQ they are like spending five minutes on every single prompt they use like some advanced templates prompting templates, you know, they have everything prepared. They're using like 10 different things. They're copying prompts. They're purchasing prompts from others, doing crazy stuff, right? The people at 100 for the IQ, the best people at prompting, they write concise prompts, but they're super clear. They're very clear in terms of what they want, what they want the agent to do. And they don't spend 10 minutes writin