Stop using Claude. Start using Codex [ part 2]

Stop Using Claude. Start Using Codex – Part 2

In Part 2, we go beyond the hype and explore how Codex is transforming the way developers, founders, and AI enthusiasts build software. Discover practical workflows, automation strategies, and real-world use cases that show why many are making the switch.

Whether you're writing code, creating AI agents, or streamlining your development process, this episode covers the tools, techniques, and insights you need to get the most out of Codex. If you're serious about building faster and smarter with AI, this episode is for you.

🎧 Tune in now and learn how Codex can take your productivity to the next level.

Transcript

SPEAKER_01 0:00

You know, we have like 30 channels as a company, and I can just ask AI to summarize everything that pertains to me. And because Codex keeps memories of me and I've given it a document that I've told what I think is important, it can find the things that are important. It'll surface them to me. You can also connect it to your email so it can send and uh receive emails, or it can like tell you who emailed you if something's really important. Um and you can connect it up to Notion, you can connect it to Sheets, and you know, pretty soon you're gonna be able to do it for all of them. The Expo one is pretty useful. It allows you to kind of view a mobile app uh preview directly inside Codex, which is really cool. Um and then obviously remotion and Canva. You can just connect it to your existing tools. And if you don't find the tool that you want to use in plugins, you can create your own skill. And you do that simply by creating a new chat. And you say, I want to create a new skill that does whatever blank. Um and then you can also say use the slash skill. Wait, I think it's at skill. Oh, maybe they don't have it anymore. To reference a skill, you use a slash. To reference plugins, you use the at mention, which I think is weird. I think they should combine all of them, right? So I could say use computer use, at browser use. Um to use a skill, you do a slash. And you can see all the slashes here. I don't see they used to have a skill builder skill, but they don't have it. I think it's just smart enough to just say create a skill and it'll just create it. After you create a skill, it'll show up here in this tab. And here are all the skills that I've created. Um I've created It's so confusing, right?

SPEAKER_00 1:54

Because like you create a skill and then it puts it into the plugins tab.

SPEAKER_01 2:00

Well, no, no, a skill is in the skills tab. Oh yeah, yeah, yeah, you're right. It's in the plugins tab as plugins and skills. And so I think this is super messy, right? That's why I think they should just combine it all into skills or pick one and just stick with it because it is super confusing. And the fact that they have the different, like you have an at for um plugins, and then you have a slash for skills, I find to be like kind of annoying. Um but I will say I've tried all of these tools. This is the simplest version that exists yet.

SPEAKER_00 2:36

Okay, and in in the simplest way possible, what is the difference between a skill and a plugin?

SPEAKER_01 2:44

Honestly, I think plugins are just like a deeper uh integration. I I would be lying to you if I knew the technical thing. I just know that like if you're a company and you want to create an official plugin, you go to OpenAI and you you create it with them. It's similar to, you remember the app's SDK on ChatGPT? It's more of an official process, right? There's a reason why you know we can't just we couldn't spend the rest of this video making the Greg Eisenberg plugin and then get it approved right after this video. You'd have to reach out to OpenAI. You know, I'm sure they have thousands of companies trying to create plugins, and you can only imagine the value of showing up here on the plugins tab, right? So this is kind of their more official integration. Um, and then skills are like what you create. Um, and uh these are just, again, all of these skills are mostly just like a folder that is stored somewhere in here, and then there's a skill.md file in that folder that has specific instructions. And AI is really good at creating skills. So all you have to do is just say, hey, any repetitive task that you have in codecs, say, hey, I want to create a skill so we can do this faster next time. And then the AI will be like, okay, cool. Like it may ask you some follow-up questions, and then it'll just create the skill, and then it'll show up here.

SPEAKER_00 3:59

Yeah, one thing you might want to do is you know, if you're downloading codecs for the first time, maybe you connect it to your Google Calendar and you basically say, like, here's what I do on a given week. And you know, help me create a set of skills that can help me do my job better.

SPEAKER_01 4:18

Absolutely. Yeah, I think I think one of the best ways to do it is just to like literally speak into your phone. I just I tell everyone to do this, just yap about everything that you do on a day-to-day basis. Like go through every task, don't skip any details, take 20 minutes to just speak about everything, get it in a document, and then just look at it and be like, okay, what parts could I automate with something? Anytime there's an automatable process and you can convert it into instructions that could be completed on a computer, that is a skill that you could create on Codecs. Uh one really cool thing is um right, so we just created this report using my YouTube researcher skill. And it looked at the Startup Ideas Podcast negative only content report.

SPEAKER_00 5:08

So I uh I love this.

SPEAKER_01 5:10

Let's let's show us drifting away from I I I I I meant to do it as a joke. I'm not actually trying to nitpick your channel. But I will say that as uncomfortable as it is to get AI to critique you, it is a useful activity. I've done it before. You know, like sometimes you can't turn a blind eye, you know, and it can and AI will actually find um it can search other channels and see what they're doing right and then compare it to what you're doing wrong. It's like this channel is doing this right. That's why they're getting more views than you. And these exercises are good. It puts you should spend more time on what you're doing wrong than what you're doing right, in my opinion, and just focus on improving it every time. Now, let's say that this was the most useful report you've ever seen, right? Let's say we go through it and we're like, oh my God, like we need to we need to do this more often. One thing you can do is you'll notice here we have this automations tab. I want you to do this every single Friday. Um actually every yeah, let's do every single Friday for the videos of that week. I want to know what I'm doing wrong. Uh please create an automation. Oops, I don't know why I didn't pick up on that. Oh, it's because I have my AirPods in. Um create automation. Uh 9 a.m. on Fridays. And so it really is that easy to create an automation. You just type in that you want this to happen. This is how I normally set up automations. People sometimes want to create an automation before they do it. It's just like do it once and then tell the agent to do that every time on this day. And that's a kind of a good workflow. And you can see here that it shows created weekly startup ideas podcast Fridays at 9 a.m. If you click on it, it opens up right here. And it shows you the status, which is active, which is good. The next run is on May 1st, 2026. The last run is nothing because it hasn't run before. Um, and there you go. It tells you what project. And then you'll notice here there's now a two on automations, and it shows up right here. And if you want to test it, all you have to do is come up here and hit run now, and it will run that automation. And that's how you can kind of test to see how good it is. And then if you notice that something's wrong, you can press edit and you can edit the automation. And I recommend doing that a few times, like whenever you're setting up an automation, you want to make sure it works. And that's how you set up these recurring tasks. And some recurring tasks that I'm seeing a lot of people have in the AI space, or just people in business, is they just have AI analyze their Slack and their email like twice a day, and it sends them an email summary. So they just they get it very concise, and they all have links to relevant uh chats or emails. So it's really easy to navigate. I found that to be a really useful one for me as well.

SPEAKER_00 8:11

And you know, you one-shotted that negative report. So if I, let's say I'm not happy with the output for whatever reason, to refine that, how do I do that?

SPEAKER_01 8:22

Yeah, I mean, say like, hey, by the way, uh the report you gave me was too long. I want more concise and make it a slash, and by the way, this document is a skill by default that that is called Word docs. So I just say make it a Word docs instead of markdown. Um and then you can see I'll use the docs skill here, I'll do two things, convert the markdown into concise docs, and update the Friday automation. You see, I was gonna have it recreate the doc, and then I would say, okay, now update the automation, but it decided that it was just gonna do it for me, um, which is cool, which is actually not how I would want to do it. Um and one thing is in skills, you can always include references with skills. So um, or in in your automation. So you could actually include the document that it creates uh inside the automation. You can say, hey, it should look like this document, right? And you can give it an example. I I always tell people that really good instructions are good, but giving it a really good example is great. You give an AI one good example, it's amazing. You give it five good examples, it's just gonna do a great job every time. Um that's kind of how I think about it.

SPEAKER_00 9:50

Aaron Powell So is there the concept of an eval on output? Meaning, you know, an eval basically in evaluation, short for short for evaluation. Basically, this is a nine on ten output, ten on ten output, and it keeps going until it gets good output. Is there that concept within codecs?

SPEAKER_01 10:09

Um I mean you could set it up yourself. The the you know, this is GPT 5.5. It'll run for as long as you want it to run. And so you could just tell it, say, like, hey, I want you to evaluate it. The problem is, um I think in that case it would probably be useful to have examples of really good outputs so it knows what you want. You know, the AI they struggle with these uh subjective tasks, right? Coding is very um you know what a good output is because the app looks good and it fully works. So it's easy to verify whether it's good or not. But a report, you know, my my startup ideas podcast. Um this is another problem that Codex has. Oops, that's kind of annoying, is like the bullets are all weird. Like so this is one thing. This is another example of something you can fix. Like, fix the bullets with their weird formatting. Um what was I saying? Before the weird formatting?

SPEAKER_00 11:13

Before the bullets completely threw you off. You were saying uh those bullets are I don't blame you, those bullets. Those bullets will throw anyone off.

SPEAKER_01 11:24

Yeah. Uh yeah, honestly. But yeah, you could just create these, these uh I don't remember what I was saying.

SPEAKER_00 11:33

You were saying I don't I don't remember. The bullets really took us for a left turn.

SPEAKER_01 11:41

We're talking about skill.

SPEAKER_00 11:44

I think you're talking about the you know, the right outputs, examples.

SPEAKER_01 11:49

Yeah, yeah, yeah. Yeah. The number one thing that you can do uh in a company right now, and what we're we're talking to a lot of companies who are trying to implement AI, is just collect as many good examples as possible. Um, this allows your agents to have examples that they can reference, store them in a knowledge base that's easy to find, right? If you store them in like a Notion database, for example, AI is really good at going to a Notion database, scouring the Notion database, finding the right document to use as the example. And in that way, you can just kind of give it all of your documents as long as you keep it organized. It's it's gonna be able to use it pretty well. Um and so collecting examples is a really useful thing. And I've noticed a lot more companies are even recording their screen a lot more in order to capture workflows because I've heard rumors that the future versions of OpenAI, uh, their new their newest model and version of Codex, you're gonna be able to upload videos to it. And so it can watch the video, see exactly how you do it, and then it will be able to control the computer just like you. And that's gonna be a whole new world. And so I'm kind of preparing for that world because most of my workflows as a creator, or like I'm in I'm in marketing, a lot of my workflows require using an interface or using a tool like CapCut, like Canva, and AI will pretty soon get to a point where it can do that.

SPEAKER_00 13:14

And and just a side note, you know, if that's the case, that uh you know these super apps are gonna allow recorded videos to be up uploaded to them, um you gotta think that the value of recorded content is going to go up in the AI age. So just from a startup idea perspective, uh there's a lot there.

SPEAKER_01 13:36

Yeah, I mean, and why do you think you hear rumors of Meta and Microsoft employees now required to record their screen and do certain tasks? And it's kind of like, you know, it's become a meme in the, you know, a lot of people who are anti-AI where like, oh, when they start recording your screen, like it's over, like they're trying to replace your workflows. And yeah, it is kind of scary in a corporate setting. But like if you run a small business and you want to accomplish more with like a small group of people, like you want to be able to offload some tasks that you wouldn't normally do to an AI that can just kind of handle these tasks. Um fix. There you go. So yeah, you can create any sort of document in here. Um recording your screen will be big.

SPEAKER_00 14:31

So the number four negative attack if you go back to that doc, one of the things that I get wrong in this podcast, apparently, according to Codex, is my interviews are way too soft. So it says my guests get a friendly runway. There's not enough pressure on failure modes, costs, security, churns, customer acquisition, who should not use the product. So I'd like to take this opportunity to take the Canadian out of me and not be so friendly and ask you the question with respect to codecs. Um, you know, who should use the product, costs, securities, churn, customer acquisition, like you know, what are the hard questions I'm not asking that I should be asking about codecs?

SPEAKER_01 15:15

You know what I respect about you is you got feedback and you are applying that feedback 10 seconds later, which I admire. That's great. Um, who should use the product? Um I think that if you're constantly I think content creators should use this platform. You know, like I'm a content creator first. That's how my brain thinks. You know, I'm constantly creating lead magnets uh which require vibe coding. And if um and yeah, I find that to be super useful, and I'm also creating documents. So if you're creating websites and documents, you might as well do it in the same platform. Because ultimately, like the websites that I create have to do with the documents that I create. I don't want them to be in separate platforms. I want to be able to create a document, do research, outline my thinking, and then convert it into a landing page seamlessly. And that's what you can do inside codecs. So so as much as I want to help you give good feedback, I think there's a reason why uh these companies are racing to create the super app is because they think that everyone in knowledge work will use these platforms. And I truly believe that that's where we're going. I'm not saying you need to use codecs. I have no vested interest in codecs. I'm not an investor in OpenAI, I've never been paid by them. I just like their platform. And I think it's the best right now, and I see a clear path to it being the best over time. So that's why I'm trying to prepare now, because it's it's only 10% as good as it will be, maybe once these AI gets really good at controlling a browser. That's what I'm really excited about. So for all those reasons, you should use it. Um but again, don't use it if you don't want to.

SPEAKER_00 16:59

Anything else on this platform that's just a must-know in terms of getting the most out of it? Honestly, you know what?

SPEAKER_01 17:04

I I I I maybe should have started this. But so yesterday in my video, I posted this video, I wanted to test browser use. And so I had it make this chess game, and then I said, I said, okay, uh use at browser use and play yourself in chess uh one move at a time like earlier. So what you're gonna see here, hopefully, if codex operates properly, and by the way, we're using 5.5 on medium intelligence, which uh I will talk about that right after. But let's see if this AI plays itself using the browser. And I think this is just like a good way to kind of get an idea of like how the browser use works. I recommend just having fun with it at first. I think people try to do like re Oh, you see, you can see the mouse, you see that? So it's literally playing itself in chess as both teams. And look at how fast that's like working, right? Like you know, a few months ago this was way slower than this. I don't know if you tested any of the browser use stuff, but this is like pretty seamlessly playing.

SPEAKER_00 18:23

Yeah. It felt like uh like with Manus, for example, it was really cool uh that browser use was even possible, but it really felt like it was on a dial-out modem. And this this feels like it's broadband.

SPEAKER_01 18:38

100%. Yeah, I just checkmated uh black. So white just checkmated black, white wins by checkmate. I played it through the in-app browser one move at a time, alternating white and black, final result, white wins by checkmate. And here we can see the full sequence. And so in my video, it did this in one prompt. I said, build a chess game, then use browser use to test the game and play yourself until one team wins. And that's what it did in one shot. Which, I mean, that's pretty insane. You know, like it we're we're desensitized to it now, but like that's insane. If you if you think about where we were even a year ago, um, and then you think about where we will be a year from now, you know, um, it may just create its own Canva a year from now and then use that. It might find it might be like, oh, it'd be way faster for me to just create my own tool because Canva doesn't have this little feature and it can't add it. Like it might as well just make its own feature and then add it and then control it.

SPEAKER_00 19:34

And that's so are you not using claud code at all anymore? This is like your daily driver.

SPEAKER_01 19:40

Okay, two things I forgot to cover. That's one of them. Command J opens the terminal, type Claude. I have it set up to automatically run in uh dangerously skip permissions mode. Anytime, like I can say, redesign the chess app to make it the anthropic vibe. Boom. Congratulations, you're using Claude code inside Codex. And you're using it with the Claude subscription. So you are getting, you're benefiting from that same, the same savings, right? Another reason why you want to use these models in these platforms directly, like in Claude or in Codex, is you just save so much money because they're subsidizing, like you get thousands of dollars of tokens for your $100 per month or $200 per month plan because you're using their platforms. So, like for me, there's no really good argument on not using these platforms, if I'm being honest, because you get so many tokens uh as a as a user. And you can also just use claude code directly inside Codecs and it'll work. Um that's one thing. Also, another thing is uh, you know, Theo and uh some other people have been talking about how they like to use 5.5 on low and medium effort for most tasks, right? If you ask it to make a simple change and you use extra high, it'll often like really try and put a ton of effort into it and it might go off and like do unintended things. And so that's one thing to look out for. I'm not very good at discipline enough at doing that. I end up using, I end up just on extra high, just running it over and over again. Um but that's what I've been hearing other people say. And you can also, if it's feels like it's going too slow, you can turn it to fast. Just know that this will be um more expensive. You'll run out of credits faster.

SPEAKER_00 21:31

I also think it's worth mentioning that 5.5 is an expensive model, right?

SPEAKER_01 21:36

Yes. It's if you use it via their API. Um let's see here. Look at that. So cloud code just changed the chessboard, right? And then you can just throw it away. The only thing is if you use if you use codex in the app, you get that like spinny animation and you can see when it's done. That does not happen when you use it in the browser. They're kind of like separate things. So just keep that in mind. But like look at how easy that was. Now, people are like, oh, but Claude's way better at design. So it's like just make the whole app and then tell Cloud to design it. Like that's what I say. It's like it's that easy. Um what did you just ask right before this?

SPEAKER_00 22:11

Um, I I just I was just saying that I you know you should should be noted that 5.5 is a more expensive model.

SPEAKER_01 22:20

Yeah, if you use it via the API, it's twice as expensive as uh the recent model, which is uh GPT 5.4. It's twice as expensive. But what you'll find is that the model is actually more efficient. So it'll get the job done and get the job done better with less tokens. And I think that's where I think the benchmark thing will be a year from now. It's not gonna be how token efficient is it, it's going to be how much money and how much time does it cost to do a specific task. You know, the same way you would hire a contractor or a human employee, you're not like, you don't think about it in terms of token efficiency. You're like, you know, you just think about how expensive is it to get a task done. And so, like, if you were what this model does is it really thinks about your intent more. And so it'll go more directly towards what it is that you want. And so you may find that it actually gets things done better and faster. Um, but yeah, it definitely can cost more money when using it on the API. I haven't noticed my credits going down.