The Startup Ideas Podcast - ChatGPT Codex is like 10 AI software developers (tutorial for beginners)

Episode Date: May 21, 2025

Join me as I chat with Ben Tossell to discuss how non-technical users can leverage OpenAI's Codex to build and modify websites without writing code. The video showcases a practical example of adding a new tab to a personal website through simple text commands, with Codex handling all the coding work. Ben explains GitHub basics, pull requests, and how to manage code changes while emphasizing that this approach makes coding more accessible to beginners. Timestamps: 00:00 - Intro 01:28 - What is Codex 02:38 - Explanation of GitHub and repositories 05:54 - Demonstration of adding a new tab to a website 10:53 - Why use Codex if you are non-technical 15:26 - Explanation of merging and closing pull requests 18:18 - Best practices for non-technical people using Codex 21:54 - Tips for Debugging using ChatGPT 23:47 - More advanced use cases beyond personal websites 25:30 - Cloning a Repo and using Codex 30:08 - Conclusion and Final Thoughts on Codex Key Points: • Codex is an OpenAI terminal product that allows users to type in tasks, generate code, and push it to GitHub without writing code themselves • The workflow involves creating tasks, reviewing generated code, creating pull requests, and merging changes • For non-technical people, Codex offers a more accessible entry point to coding than traditional development • Best practice is to start with simple personal website projects before attempting more complex applications: LCA helps Fortune 500s and fast-growing startups build their future - from Warner Music to Fortnite to Dropbox. We turn 'what if' into reality with AI, apps, and next-gen products https://latecheckout.agency/ BoringMarketing — Vibe Marketing for Sale: http://boringmarketing.com/ Startup Empire - a membership for builders who want to build cash-flowing businesses https://www.startupempire.co FIND ME ON SOCIAL X/Twitter: https://twitter.com/gregisenberg Instagram: https://instagram.com/gregisenberg/ LinkedIn: https://www.linkedin.com/in/gisenberg/ FIND BEN ON SOCIAL X/Twitter: https://x.com/bentossell Ben’s Bites: https://www.bensbites.com

Should you try Sam Altman and OpenAI's new AI engineer in the browser, Codex? Well, in this episode, that's what we unpacked. We do a live tutorial for non-technical people around how to use codex to get the most of it. I brought on Ben Tossel. He's one of the most renowned people for explaining no-code tools for non-technical people. If you want to see more content like this, like and comment so I know to create more free stuff, enjoy the episode. Ben Tossel on the pod, he just went viral for talking about Codex.

So we had to bring him on the Startup Ideas podcast to introduce us to Codex. What are we going to learn today, Ben? Well, we'll find out. I think what I've seen is people talking about Codex as, this is really good for like senior engineers and their workflow, and I don't like people telling me what it's for. I just want to figure it out myself. Can I use it?

Does it work well for me? If so, it's a tool for me. So I'm non-technical. I also have to like disclaim that. I'm more technical than a non-technical person and less technical than a technical person. So somewhere in the middle, and I'm just making these tools work for me.

Cool. So today's going to be an introduction to Codex by a non-technical British man. and let's get into it. Okay, cool. If you don't know what Codex is, it's effectively a way to type in a task. It will code it for you

and then push it to GitHub. That's basically what this is. So here it is connected to my GitHub repo, my personal website, the main branch, and the website looks like this. This is it so far. So if we want to

change something here sell it on I have another tab as like food I like I'll just say add another tab next to investments

tools that is called food I like in the dock put tacos so because I want to

I want this to be true to the non-technical person so you had to set up a GitHub you had to create a website how did you create that website originally with codex yeah so yeah so if we go back this is something I did cover in my video which you need to have

GitHub connected so yeah GitHub is effectively a way for your code to be stored and you can like have, so just like what a profile looks like on GitHub, but not a very good profile, but a profile nonetheless. So you have these repositories,

which is basically like, that's your project. So this one, for example, is my personal site. It effectively stores all the code files that you have, and then any changes you make, you commit them. So they're called commits. They all show up here. So you can see that it looks like I've been very busy. I sort of have, but I've been sending all these two codecs to do, and then you'll see some other

coding agents in here, which we won't talk about in this video. But there's more other tools out there as well that I highly recommend. So that's very briefly what GitHub will do. And then you have to create a new repository. So when you've got a new, like you want to do a personal site, you want to do a marketing site, whatever you want to do, you've got to create a repository for it. If you type in any name,

you can make it public or private, add a read-me, just explains what your project does, and then create it, and then you'll end up with the page that I was just on, something like this. Yep. And that's it.

But then you're like, okay, well, what do you mean? Where's the project? Where's the site then? And I'll have linked mine here. So what I did actually to start this whole thing off was I had my site built on Card, which is a no-code build that I've loved for years, and it was doing fine.

My site was up and live for years, and I just didn't do anything with it. And then I actually went, and I just did right-click. I think I did V-page source, which shows you the whole code that serves the page. This is the code that is on my personal page. So everyone can go and look at this. So on the no code builder, I had this. I just did all copy, paste into a coding. I actually used a different coding agent factory to then say,

put this all into code and put it on this GitHub repo. And that's where this then started off. I've also connected, you can connect it to then GitHub pages, which then lets you have your URL live and you can give that to people and have custom domains there. I use Vesel. I'm just trying to be a bit fancier, and that's another way to deploy stuff.

And you don't really need to worry about that too much. If it's like your first foray into sort of coding stuff. Yeah. Cool. Okay. So websites live. We want to add something. We want to change something.

We add a task on codex. Yeah. So add another tab next to investment tools that is called Food Alike in the archiput tacos. So if I just, you can hit ask, which is just it doesn't generate any code. It'll just tell you stuff about your code. We're going to hit code here and then it'll come up as a task here. And you can see these are all past ones where they show how many lines of code it added, how many have been removed, whether I actually merged it, closed it. And we can talk about merging, closing, all that kind of stuff. In a sex, we'll see it first hand. So when we go

into this task. You can see it's doing a bunch of stuff that looks a bit like hackery, that you see in the movies. It's just using the terminal to find different files, understand what the code actually says. You can see there, this is the code. It's my command palette modal. So I got codex to add this. When you do command K comes up with like shortcuts modal. So then I can type it, I can just do I or T and it will navigate the tabs. just in my keyboard. So this is now, so I'm interested to see if it adds a new tab

and if it adds that to the command line as well, if Codex are smart to have to like piece those pieces together. So here, it's saying these are all the things that I'm doing. I've looked at each file. I'll check out what the documents look like. I'll create a new document for the new section with the title, Food I Like and Tacos. matching the style.

I'll create the, da, da, and update the main site with a new tab. And it has, said, I've spotted a key event listness for I&T, but no changes need there. I'll add a food tab. Okay, so that doesn't look like it's going to add its own. So the I&T is my keyboard shortcuts

for navigating those tabs. I thought it would do that, but maybe because I didn't ask it to. So now, it's still working on it, almost done. We'll see this in a second. So you can see that it's added two lines of code on the food document and more on the other one. This is what it looks like when it's done it successfully, it seems, anyway. So I'll show you the files that's changed. This is the food file. This is the

main website file. So it's added another button, food I like, and then added the food piece here in the tabs. It did its own test. So Codex is like a use case they talk about a lot is testing. And we can get into that a bit later on because I think it's a bit more technical than you probably need for this, like for this workflow anyway. But as we're here, we're like, okay, great, it's all worked. So if I go to my sites, refresh it, it's not there. So I haven't actually pushed any of those changes.

So this button is what you want to do next. Create new PR. So PR stands for pull request. And effectively, code is like one thread, just one line, which is the main branch. They call it branches. So it's your main branch. Every time you go and work on something else, it's a fairly standard practice to create a branch. So you basically take that code, you copy it over to like another version over on the side.

you work on that feature, as long as that feature is successful and there's no conflicts with the main code base, then you can merge it back in and no issues should be there. So that's what I'm hoping for today as well. So if you click view the pull request, so on my repo here, pull request, add food tab, summary, gives you a quick summary of what it did. This is my, because I'm connected to Vassal for deploying, I can do that and see the testing stuff there. So that's something we can get into later. All checks have passed no conflicts with the base branch. So you can merge it.

So if we just merge that, that will then, you'll see that here. This pull request is coming in. So it's just pending. Then I'll take it a couple of seconds. And then once we've done that, it'll deploy. And we'll be able to see it on our site. So that loop is a bit long currently. And this is where we can get into the sort of testing piece

and running tests, environment, previews, like all of that is stuff we can dive into because this is the simplest version, I think, which is just like write a task, get code, push it to your code base, see the change. Okay, quick, before you get into that, why am I even using codex?

When I look at codex as a non-technical person, it feels 10 times more overwhelming to me than some of the prototype being first tools like both and lovable and stuff like that. So why should people even, you know, besides playing with the tools is fun and it's good to learn, why should a non-technical person even do this? Maybe they shouldn't, is my answer.

Great. I think, I mean, it's part of my job to just like test everything, but more so I like to see where things feel like they might be good. So seeing Codex, the UI of Codex is very different to the UI of everything else. I mean, you could argue that like a bolt would look kind of similar where it's got like the sidebar. And then, I mean, I don't know if I've got any other builds here. So this is one of the while.

So it's got like the sidebar chat. And then you've got your files, your code. You can see the preview, all that kind of stuff. You can deploy it. You can integrate with GitHub. Like all of the same pieces are there. But I think what I find interesting about code is you don't look at the code or think about the code at all.

I think you're obviously going to see it here, but I very rarely look into any of these files. I'm only doing it to sort of demo it. But generally it would be like, okay, add this. Okay, it just says it says it's open. That's what I'll talk about that in a second. So it says it's open. And I'm like, okay, that worked.

Poor request. Did it. Pass the checks. Fine. And this you can now do on your mobile as well through chaturit. I think that you can do it all quite easily. So you can see a will where you're sort of setting off mini tasks to an agent to go and do that. I think bolts, you can do it, I guess, is on your mobile.

But it's not going to be great. And I just think generally, text-to-app builders are overbuilding everything. or completely setting everything up as an overbuilt tool, but actually all the features are underbuilt. So if you ask for a Spotify, a Spotify clone with Bolt, it'll spit out like UI perfect, looks amazing, but all the pieces don't actually work. And when it's those pieces that you actually want to get working

in order for it to be something that you can deploy and do whatever you want to do with it. So you focus on the end piece and then you're going to work your way back through all the pieces that need to work which is actually hard if you ever have tried to do that whereas in this one I'm very iteratively adding onto something and every time I'm adding a new thing technically there are tests and checks and everything should be all fine

so there should be a lot less time spent doing work on things that are not going to work like things that I'm going to then spend hours debugging. I'll either just give Codex another task or try a different coding agent. So my workflow is kind of changing in that sense. I get the chatting back and forth when I'm trying to make something. And I think it's really helpful to have the like the ask version of these coding agents versus the coding ones, same as in cursor. You like have one chat and you just ask about what's happening in the code base,

blah, blah, blah. but something about this feels a lot different and it feels a lot more like writing or like you said this is your to-do list but you're not doing the to-do you're not doing any of the to-do's something that's way way more advanced than you is doing it in a way that feels delegated and off your plate so I just think you start here and you can think well at some point the like the PR stuff that's going to be automated like I'm sure you can get bots now that if there's a successful pull request, it'll just merge it to the main branch.

So you never have to go on to GitHub. And then you can do, I think, I think the limits around a task can last about 30 minutes or something until it starts breaking. This is early preview, obviously, so it's all a bit on the fritz. Can you explain a little more about what we're seeing around merged and closed

and what's happening? Yeah, but let me just show you, did this work, Okay, it did. The keyboard shortcut does not work. But food alike, that all works fine. Tools. Yeah, great. So that change all happened. So then this is merged. So that means it's been merged to my main code base, my main branch. It's all part of the main code base now. So if there's another branch, it'll have that code already in it. So when you create a task, it starts off as open.

you don't see any open ones here because I've either closed them or merged them. This is an example of one. I said, make a light and dark mode icon button switcher. And this is for a different test that I was doing. It made one. I looked at it as like when I was testing it and I was like, no, that's awful. So I just close the pull request. You can just deny it basically.

Like, hey, I want to make this change. And you're on that GitHub page. You're like, well, none of the checks passed. it didn't work or whatever, you can just close it and then just not use that. So none of that is affected. It was on a branch. Your main code's fine. So you basically like think about it, I guess, in the support sense.

You open a ticket, tickets being worked on, and then a ticket's closed. Sort of problem solved. That's kind of how I think about this anyway. So yeah, we've opened the task. We've merged it. We've closed it. It works on my site. I did no coding.

and I did maybe like four or five button pushes. And the numbers that I see next to merge and close, I see like a positive number and a negative number. What are those numbers? Do you even know what those numbers are, Ben? I do. They are the number of lines of code changed. So plus 12 minus zero.

We didn't take away any code, but we added 12 lines of code across your whole code base. It's not just like in one file. Generally, that's, yeah, that's what it is. So even on my personal sites, I have this, which I haven't worked on yet, which is, I basically try to copy the Codex UI, and that's what I use for this, inspiration. But yeah, I've got these numbers here.

They don't actually work. But this one actually pulls in the live GitHub repo. So any changes I make are reflected in there every 15 minutes. I don't know if we did. Did we just do 12 lines? We did, yeah. So that actually is an updated one. So if I click on that then, it'll show me the commit of what I just did. Just, again, a little thing I just said to Codex, try and do that, and I did it. Like, no problems. Before we go, you know, you've been playing with Codex for a few days now. What sort of best practices do you have for non-technical people who are playing with it? I'm worried about what people will do and then call me out for. But I would just start really simple.

Like use it on your personal site or create a new personal site if you're precious about yours. One of the first things I did like I mentioned was like, okay, I've been on this no-code tool builder for God knows how long. It'd be nice for me to just have the site in code so I can make any changes in any weird and wonderful extra pieces that I'd like. So just do it line by line and like piece by piece. You can see all of these merged pieces like add a food tab. I could do that in the code myself. Remove this text animation. Add keyboard shortcuts.

Add a dark light mode switcher. So like all of those things, they're all like nice to have fun, interactive things that you see on other people, like other developers' websites that maybe you just don't have on your own. But I'll just start with make sure you have a GitHub repo. and it has at least a read-me, so you can just enable that when you're creating one.

And then just go from there, say, okay, make a website with my name as the header, create an about section, and then create some social links, and go from there. And then I think it's just playing around with everything and just merge stuff, close stuff. The thing about GitHub that is good,

but it's not that easy, I don't think anyway, is if you have merged something that then your site completely fucks up. Like I did it before where the site just wasn't loading anything. I was like, oh shit, what have we done? You can always go back to a previous version that did work. So if that food tab thing broke a bunch of stuff on the site, I'd be like, okay, I'll go back to this.

So I can go back to this and all of these are like just disregarded. The code is never there. Don't worry about it. So there's no real danger of if it's like a personal site that isn't client-facing, money-making, whatever. It's something that you're just working on. You can iterate on and add, like, weird, wonderful, funky things. There's no real risk of, like, anything happening other than your site changing.

And then when you get stuck, ask ChatTBT. Like, hey, how do I roll back this GitHub stuff? How do I, like, I want to... Because the thing you can't do in Codex, so there's some limitations here, obviously. Like, you can't upload any files. you can't upload any images, you can't say make the UI look like this. So me trying to do the UI piece was a challenge.

But then you can go to other tools like V0, generate some design stuff, bring that code in here, or you can deploy that and GitHub and point that to here. Like there's loads of ways to do it. Yeah, I just start small and iterate even smaller. And it's just getting a feel for like, could I see myself coding like this in a future?

like in six months time, am I going to just think this is so obviously how everyone should be writing any code? I can't write a line of code myself, bar like the minimum stuff. But I can ship all sorts of stuff with tools like this. And especially if being mobile native just feels like I'm going to be shipping all the time. When you say, ask chat GPT if you don't understand. something. Do you mean go to chatchipt.com or do you mean go to chatubtt.com slash codex and hitting the ask button? I generally go to chatubtt.com and then I'll change between the models to see what I what's going wrong. And the good thing about using chat chitb is you can still say

like look at this repo or you can try and some other tricks as well like you change the first letter of GitHub to U, so Uit hub. You'll see this, which is, just looks like a lot of text. But this is all the code that is involved in my site. So I can basically copy all of that, and it's going to be 9,000 tokens. I can copy that. Go to chat.com. And then that's, yeah, and just say,

my personal sites are there any bugs? You just do like that kind of thing. So I just tend to do that. I haven't really used the ask. I think I did it here. So I asked, how do I ask you to push changes to my staging environment? How do I check the staging from a UI perspective? And it can do it. But it's more connected to, it's not connected to the internet. So it's like, I'm just looking at your repo and answering based on what it has, what knowledge it has to give you that response. Whereas I prefer it to be based on my code and also various sites that I can go look up how I can help you fix this thing.

So generally that's what I've been doing. Yeah. Good tip. I also like your tip around try this for, you know, well, you kind of said like there's levels to this. level one is build a personal site. What are the next levels? After you've built a personal site,

what's level two, three, and four as you're learning about codex? Yeah, I mean, I think I haven't seen anyone, I saw someone tweets about this. Like, has anyone used codex for like working with databases or authentication? Because they're like, there are a few steps up to deal with those. Then there's things like MCP,

which are making a lot of this very, very, very, easy in things like cursors. So it's, again, maybe I'll just be patient and wait for MCP stuff to be in Codex. Or maybe I don't use Codex anymore. And Claude have just released their SDK and there's a version on that. Like, there's all different ways that this is probably going to go anyway. Yeah, I think the first step is like it's just getting you to get you down the rabbit hole. Like, oh, I can use this thing to code onto a site that I own and can see. cool what would I want to add and even on your own site you start going it would be great if an integration with GitHub could pull in the latest

information like even that is I would say a next mini level up it really depends then on what what kind of things you're into and want to build but I think there's there's obvious things like having user sign in sign up there's like people saving or bookmarking items in your product or whatever you're building there's the day database layer, there's like the whole AI piece is another thing. But the best thing about this whole space is there's so many open source like repos on GitHub. So you are going to be spending time on GitHub that just, I mean, I think there's a, there's a guy who works at

Together AI, Hassan, who always releases all of these open source. Yeah, all of these open source, like mini AI tools. So this is like a summarized PDFs. Okay, great. And then they're always open source, and this might even look scary, but really it's clone the repo, which is clone the code base by using that.

And for that today, you'd have to use some of the other coding tools. I think it's like, it's going to be fairly soon where you can just, I mean, I might be able to do it now. We can try. So if I fork it and then just do that. So that, I've now got that code.

I wouldn't just change the name and say it's my code, but that's what it is. So then if I go to codex, I'll need to go to my environments to create the environment, basically connect the right repo. What was that called? PDF or something. It might just, there we are, smart PDFs.

So if I just do that, I think there you are I used to be. So then if I click into that, how do I get this code up and running? And dumb question, like why did you need to create a separate environment for smart PDFs? It's just, so they called it environments and it's just like a different connection. So it's a different place that it's talking to and writing code to. Because you can see here, these are all. It's like a folder. Yeah, it's effectively that.

So I think even the fact you have to come into environments, create an environment doesn't feel intuitive. No. It just feels like you should be able to, I don't know, an ad button here that is just like connected new repo. But the environment's piece is much more advanced than that, which is why it's probably done this way. So here they have this piece, the code execution piece,

which I haven't actually touched yet. But often if you have code like what Hassan's got, so I assume that's why I ask that question, because I think it's going to be helpful to show a bit of that. Where is the code? So yeah, I'll see here, when you're running, to run this app, you need to run P.N. MPM install.

So that's a terminal command. You've got to run this and you're going to run this, which feels a bit crazy and like I don't know what you're saying if you're an untechnical person. But if you're in cursor or some sort of actual coding environment, there are terminal command to basically spin up the servers

and the things that you need in order for the app to work. But with Codex, for it to be able to interact with your code and see it properly, you need to run those commands. So here is where you say, it gives you examples there,

but it's like PM, PM, whatever it was, install, and then whatever the other ones. So you have to put them here. This is like proper development environments. When you're building something that is a bit more, it's got frameworks like XGS, Python, whatever, like it's got things in it that's more than

HTML, CSS, and JavaScript. which is more what I talked about today in the personal site context. But yeah, I want to do this. I want to test this out. And yeah, there might be a follow of video to say, I was wrong. Although it was really good.

So, yeah. Cool. Let's leave it at that. I want to get more technical and more down the rabbit hole for this. Ben, thank you for being generous. It's your time and your sauce. I will include where to follow.

follow Ben in the show notes. Highly encourage people to go check him out and his newsletter, Ben's Bites. Anything you want to leave people with, Ben? I think all of this looks scary or feels scary and there's too many things launched all the time.

So that's just the world that we're in, unfortunately. And I think just like getting to the point we feel like maybe just test this one thing and on this one simple site idea and it's one simple feature, you're just going to start there. And I think the whole point of what I do with my newsletter and what I tweet about and record videos on every so often

is trying to make it less scary for a non-technical person. Like today, it wasn't a planned, like tutorial style. This is how you can do these things. But my video might be a bit more, this is like a really more structured way, like showing you actually how to set other projects and stuff. but it's not that scary. Like, I know it's easy to say from the other side of the fence,

but I've been non-technical forever and been talking about no code forever and all those other things. So I come from that world of I can't code, code is scary, what is GitHub, what am I supposed to do with it? And I'm just getting curious enough to go past those barriers to them be like, oh, I've got a site live, and now I've got 50 live, and I've got all sorts of different apps of different complexities and like when you go to the

rabbit hole you're going to start wanting to do more once you've figured out that so don't set yourself up for failure with let me do a Spotify clone and then deal with all the bugs in the middle because it'll just be a nightmare you'll never want to do it again like start with one simple thing try and do that if you fuck it up then ping me I guess um yeah amazing then um yeah my you know just my takeaway from you showing this is it does feel overwhelming, like just the semantics, the terminology, GitHub just itself feels like a behemot. That being said, I think that playing with codex is almost like a lightweight way to learn how to code in some ways.

Like it introduces you to a lot of computer science and development, terminology, in a lightweight way and a bit more of a fun way. So whether you end up using codex or you end up using Bolt, for example, you know, it does, it's just, this is going to make you better at building with AI, period. And the lovables and bolts and curses of the world are probably going to introduce a to-do list task-based UI in the future

because it just kind of, it makes a lot of sense. So my take is, will I be using Codex every day, Ben? No, I'm not going to be using every day. Will I be using Codex this weekend as a learning project to play with my personal website? Absolutely. Yeah, yeah, exactly. And I think it is that terminology that you've got to get over a bit. I did a write-up actually.

It was on our substack that's like GitHub for non-technical people trying to dispel some of that scary stuff. But yeah, I mean, I think the same. And it's more that like you'd rather. do one thing correctly than the big thing and then try like this introduces coding to you like it drip feeds it I think

with you've got to set up GitHub and then you can start coding that's fine with the text to app builders you can just type a prompt and you start building great but it's not the same thing you still end up at the same place at the end of it you still have to deploy it it lives somewhere it's probably going to live on GitHub

like it's going to end up in the same place, it's just how you enter. And I think what you mentioned with, is this going to be the place where someone learns about coding or like, this is where they write their first line of code or AI writes it for them. I kind of feel like this is a better place to start than a text startup builder because like I said, it's going to build a big fancy thing. And what you're going to end up in is like seeing a bunch of terminology, code being written, tests failing. They're all to do with now you have to help the AI solve these bugs. That's a way deeper, like, place to start swimming done here, I think.

So that's what I think is interesting about this. It really doesn't feel like coding, but it is. And it's come from the, like, almost top-down, bottom-up, like, analogy you could think about here, maybe. But I don't think it's built like that. So it's obviously meant for the whole stack. So it's, yeah, we'll see. Then toss, everyone.

Thanks for the time. Hope you learned something. Cheers.