Getting the most out of Codex

Claude Code is a pretty amazing tool. You probably know that. I know it, I'm using it all the time. And a couple of weeks ago, I shared a video with some insights on how Claude Code to get the most out of it. I also created the full Claude Code course. But over those weeks, I've also heavily used Codex, and, uh, for example, this drawing app, videos recently. That was built with Codex and Claude Code and, of course, myself. Now, obviously, Codex is an AI coding tool, so the things I shared in that Claude Code video pretty much also apply to Codex. But Codex also has some extra aspects, which I like a lot, about which I wanna talk today, how I'm using Codex in addition to Claude Code to, well, get the most out of that. And one interesting thing I wanna start with is that on the Codex website already, the first thing you see, at least if you're on a Mac, I guess, is this desktop app, that you can use the Codex desktop app. Now, Claude Code also has a desktop app, but for me, it always felt like kind of an attachment to the normal Claude app, and I never really liked it or used it a lot. I used Claude Code pretty much only in the command line. And Codex, I used there too a lot, but I will say this desktop app is pretty, pretty good. Now, at the point of time where I'm recording A Windows version is planned, I think, and it is an Electron app, as far as I know. So yeah, it should be available on other systems too. I think the reason I read why it's not available on Windows yet is, uh, the sandboxing, which is built in where the agent runs in a sandbox, which uses native platform features. That doesn't seem to work on Windows yet, I don't know. But this app is pretty nice visually, for example. I like the aesthetics, and I like the way it works. It feels really, really fast. It has a clean UI. I also, for example, like that you don't have your file tree in there by default, but if you have pending changes, uncommitted changes, you have this very nice clean overview of which files changed and what changed in each file since the last commit. And you can leave tiny comments here by clicking a plus. Remove that. I don't know. And if you do that, if you add such a comment, that gets added as context to your prompt. That is a really nice way of combining both, essentially. Reviewing the code, diving into the code, I do. You should still review and read the code. Don't fall into the white coding trap. But then, you can directly add your comments there link them to your prompt, your follow-up prompt you're going to send AI. So, I like that a lot, and I think that's one really, really cool feature there. Now, that being said, I'm not using the desktop app for It's just, uh, one way of using Codex, which I do like to use from time to time. I'm, I'm really using it in different ways, using Codex and Claude Code at the same time, also to stay up-to-date what's, uh, about what's changing and how these things behave. By the way, side note, just as I have a complete Claude Code which you find linked below the video, I also created a complete Codex course, where I share all my knowledge about Codex dive way deeper into Codex and how to use its advanced features, also features like git work trees or how to use Codex as a non-coding agent, which can also be super fun. I got all that in the course, and you find a link to too. So, if you wanna learn way more about Codex the, the links are below the video. Now, as I shared in the Claude Code video, I still don't loop. I don't use the Rolf loop thing. Is that still a thing? I don't know. I stay in control. And yeah, planning still matters. And of course, therefore, you also got a plan mode here. You can enable it in the desktop app simply by hitting shift-tab. But it also exists in the command line, so you can use Codex as a command line Code and hit shift-tab there as well to go into the plan mode there. And I'm using that all the time. That hasn't changed. I find that super, super useful. Just as with Claude Code also, explicit over implicit still applies, and I'll get back to the skills part here. That still applies, and you still wanna provide relevant context. Those AI agents and models that are used in these agents, they're really good at finding relevant actually, so you don't need to mention every single But if you do know that a certain file will matter, like here, this file, you wanna point your agent at it. You wanna tell Codex about it. Or here, where I wanna add, um, SEO, metadata, and so on. This is a Next. js application, so it

makes sense to do a quick search, find the relevant docs article, copy that article, and then maybe expand the prompt like this that article in there. Now, I won't do that for every single article every single piece of documentation the model should know about. You can, and should, also use other features or concepts like the Context Seven MCP, which is a MCP server you can add to your AI agent, which is, in the end, a search tool for documentation. And when using Codex, adding a MCP is, uh, very simple if you're using the desktop app. There, under Settings, you have your MCP servers... and you can just click Add Server here and fill in your MCP details there. But, of course, you don't need the desktop app. You can also just follow the instructions you find, uh, on the MCP page. So here in the case of Context 7, under All Clients, there is OpenAI Codex, and then you can just tweak your Codex configuration, the global one or the project-specific one, depending on where you wanna add that MCP, include this essentially. You don't need, uh, an API key with, uh, Context 7 for basic usage. So you can do that. But if you know that something will matter, be explicit about it. It speeds things up. It can save you some tokens, them. It may be a subscription, but still, you got a certain amount of maximum usage in there. So yeah, if you know about something, be explicit. It's a good idea. What I also still like a lot are skills. In case you don't know, agent skills is kind of an official thing. There is a standard, and the idea is that you have your skill folders, and you can add them, for example, in your. codex folder in a project, have your skills folder in there, on your system if you have skills that should be shared across all projects. But then a skill is really just a folder with a Skill MD file in there, and then you got a name, you got a description, and that's the important part because agent when to activate this skill, and then you have your actual skill information in there. Now, the idea behind skills, in case you don't know, is discovered step-by-step. They're lazily loaded, you could say. It's not the case that this entire Skill MD file will be loaded into the context window every time you send a prompt to the agent. Instead, only the metadata is loaded into the context window when you start a new agent session, along with the system prompt and so on. And it's then this description in the end, and the skill name, but mostly the description, that gives the agent a hint about when it might want to read more about that skill, when it wants to activate that skill. So that description is really important. And then, for example, here, when working with TypeScript, decide to read the full skill, and then in there, I- I can have my additional instructions on how to efficiently work with TypeScript. And there also is a skills registry, skills. sh, and here, you find all kinds of skills, uh, for example, also on working with BetterOff, so that you can teach the agent, so to say, give it some extra context, some extra knowledge on how to efficiently use BetterOff in this case here. And in the end, it is always all about constructing the right context and giving the agent the right information. So skills, just as with Claude Code, still matter a lot when working with Codex. And I found Codex to be very efficient at, when it comes to activating the right skills at the right time. Uh, I- I was really happy with its behavior And then there are these points here, which I also already mentioned in the Claude Code video, and they have not changed, but they're very important to me. You should absolutely embrace these agentic engineering tools. They are the future of development. I'm convinced about that. And you can be super efficient when using them, when using all the features they offer. It is important. But you wanna always verify the output you're getting out of them. You wanna review the code, and one feature Codex also has there, which I like a lot, too, is its code review feature. Claude Code has a similar feature, uh, and I find these features very, very helpful. You can just run the /review command. Then you can choose what to review against. So if you wanna review the current branch against the base branch, which is kind of what you would do with a pull request, changes, and you can start that and let it do its thing, and that often gives you an additional set eyes, if you wanna call it like this, that looks over your code. But you still also always wanna do that yourself. You're responsible for your code. It's not the AI, it's you, so you wanna carefully verify it and do code reviews and be clear about what you want and

check if you got what you wanted. And I mentioned this also in the Claude Code video, and it's very important to me, you're still allowed to write code. You're still a developer. Of course it's shifting. Of course it's more about planning and controlling and steering the AI and doing code reviews. Our role is evolving and shifting. But you're still allowed to write code. (laughs) You can still do that. If you know how to build something, if you wanna have 10 pixels of margin in a certain place, just do it. You don't have to ask the AI a- about it. You should also, by the way, always make sure code base, that you don't detach from it. It's a trap it's super easy to fall into. become fully detached, and then you're essentially forced to do white coding, and you don't wanna get there. You wanna always understand your code base, and you always wanna be able to go in there, make changes on your own. You will do way less of that than in the past. You will write less code, at least I do. But you're still allowed to write some code. And then there is one last thing I also wanna mention, which also applies to Claude Code and essentially all these coding agents, but also to Codex. You have this concept of a Agents. md file, right? And the idea here really is that you have this central file in your project where you have some basic instructions the agent should always follow. This file, unlike the skills, is always loaded into the context window of every session. Therefore, you should keep this file concise. You should not bloat it, because it will take up space in every windows, in every context window of every session, be super large, it would not just take up a lot of also, the more irrelevant context you have in share along with every prompt, the worse the results will get. So you don't wanna put a lot of irrelevant context into your prompts. And of course, if you had a super large agents. md file, high that there would be so much information there that often wouldn't matter, that you might get worse results. So this file should really be focused and short and concise. You, for example, wanna describe your basic project stack, some key commands that can be executed, some general core code styles or rules you want the agent to always follow, and what I like to do, clearly encode how it should self-verify its work, because that's one of the most important things when using, uh, coding agents. No matter if it's Cloud Code, Codex, anything else, you wanna give these tools ways of verifying their own work. You wanna have automated tests. You wanna allow them to use the browser Playwright MZP, for example. You wanna use linting, type checking, all that matters, and that's something I also like to put in So use that agents. md file to your advantage. Keep it concise, put the important information in there. And whenever you see the agent make a mistake or repeat a mistake, I should say, put the fix for that in that file too. If the agent consistently misbehaves and does something in a way you don't like, tell it about it here in this file, that will be loaded for every new session. And therefore, using this agents. md file efficiently like is often overlooked. It's an after-sight for many developers, and it's a crucial file you should use to your advantage, but you should ensure that you use it correctly. Yeah, and that's effectively how I use these agents. Much of the things I mentioned in the There are some nice extra features, some additions which I didn't mention in video, and yeah, I hope with that, you can get the most out of these agents out of these tools too. They are the future if we like it or not, I think.