303 | From ChatGPT to Codex: How AI Agents Are Transforming Marketing and Business Operations with Dan Sanchez

20:03Isar Meitis:Yeah, to be fair, they just gave everybody until sometime in July, I don't remember, like 50 % off on Claude, which means I downgraded immediately to the$100 a month from the$200 a month plan. but in general, I agree with you. And the general rumor is that OpenAI is going to make their stuff even cheaper just because they're both in like final stretch for their IPOs and they want to show that they're better than the other guys. So it might get even a bigger spread from a cost perspective. Yeah, so chat GPT is more competitive from a cost perspective. So those are the reasons why I chose Codex.

20:39I just find it's a much more intuitive tool for generally non-code work, though I am using it for code. It is very good at this. So if I can, I'd love to talk about some really interesting practical applications that I've done. Yeah.

20:52Isar Meitis:Let's dive in. So I am, in fact, this is an old project. In fact, I'd rather just talk through it, but maybe I can pull a big example of it. While you're looking, two cents about this whole agentic universe versus those of you who are still using chat. And I've done multiple episodes on Cloud Code and Cloud Cowork. But the biggest difference is that it actually works with you rather than you tell it what to do and it gives you answers and then you have to probably do the work yourself. So you brainstorm what you're trying to do. You find the right path with the AI. So the AI will suggest a path on how to do the thing you're trying to do.

21:28Isar Meitis:And then it can actually do the work and you can define which checkpoints you want as a user to review the work. So as an example, I haven't written a single proposal myself for a very long time. The AI does the research on the client. The AI takes the chat from like the transcript from the call. The AI develops the plan for the proposal. It writes the proposal and then it shows it to me to review. Once I review it, it knows how to put the bells and whistles on it, right? It knows how to add my header and footer and table of content and, you know, all these kind of things that need to be in a proposal.

22:02Isar Meitis:And it also knows, because it's an agentic world and it's connected to a lot of other tools, it knows how to put it in my CRM. It knows how to create a draft email. It knows how to put it in the relevant folder or create the relevant folder on my Google Drive. So all these things are agentic capabilities that allow the AI, again, like Dan just said, understand the concepts, understands the context, develop a plan, and then with your approval, create the actual thing that needs to be created. So let me show you the thing. I am working with a company called Trim Healthy Mama, and they are a supplement company, but they put out a lot of recipes for their supplements.

22:41It's a whole thing. So I have this whole site with recipes on it. And you can see they're like interesting recipes. They got nice photos. And I was like, you know what? What can I do here? I have about a thousand recipes on this website. And I know this image generator can do some really cool things. I'm like, I wonder if I can create a process that takes every single recipe, turns it into an infographic. You know, you've seen those infographics of maybe like a flat lay photo with all the ingredients that goes into a recipe labeled. And they're just kind of interesting little infographics. I'm like, huh?

23:16Well, I have the recipe with all the ingredients and it's all well-structured and everything. I got an interesting picture of what the color and the thing looks like. I wonder if I can make a system that will literally take all thousand of these, generate all the images, and then just re-upload them all. With ChatGPT, this was impossible. Shoot, doing this manually is a freaking nightmare. Even if you can do all the images automatically, it's still a nightmare because I have to download them. I got to check them. I have to, we're just re-uploading it times a thousand. Anything times a thousand takes a long time.

23:46Five minutes times a thousand is a very, is a long time, but I'm like, huh, I wonder if, so what I did was actually worked at step-by-step and you can see I have this like whole dialogue. I'm not going to get into the whole dialogue, but let me tell you, I'll show you little pieces of it. The first thing I did was I just took a web URL like this one. And I just, this is how I recommend approaching codecs. If you're like, huh, I know it can write code, which means it can access WordPress. So let's test this one little baby step at a time, one baby step. And all I did was take the URL and I was like, Hey, I posted in here and just gave it instructions being like, Hey, take this URL.

24:18I want you to scrape the recipe ingredients. Take a look at the image, download the image too. And I want you to turn this into an infographic of a flat lay. I described the picture that I wanted and this was the result as like, can it, can it generate the image? Yes. Look, it can generate them. This is a different one. This The photo that I got was probably from a side profile, but it did a top lay and it did a really good job of generating all the ingredients. If you can't see, it's like a top lay, every ingredient in a separate little dish as if you were expecting it from a cooking show. And everything's got like handwritten or text showing you like, oh, this one's apple cider vinegar.

24:53This is mineral salt. And it looks really clean and professional.

24:56Isar Meitis:It looks like a design company actually made this. It's got the right logo at the top. It's got the right font. It's got a little call to action. It's going to be floating around printers. Get the full recipe at trimhealthymama.com. So it has like a call to action in case the recipe gets disconnected from the URL somewhere. So this would be, this is hard to make if you were doing this manually, but with a chat GPT image generator baked into codex, this is easy. Okay, so I'm like, okay, I can make the image. Ha, I'm like, cool. Now I'm like, codex, what does it take to like pull all these recipes and create like a checklist for yourself?

25:29It's like, oh, well, it would be awesome if I could get like your index file. I give it the index file. It creates a checklist. And I'm like, cool. What would it take for you to make poll one, do it and then upload it? It's like, well, I would need access to your WordPress API. You're like, great. What do you need to do that? And you're like, and it's like, sure. You log into WordPress here, go to this page, go to your user and then generate the API key. This is what I need. Tell me what it is. And I'm like, okay, here it is. It's like, great. Yep. Accessing your WordPress. Oh, give me this one more thing.

25:55And it's like helping you troubleshoot little issues it runs into. You're like, okay, got it. It goes and it's like, wait, I'm now searching your API. guy, it looks like I have access to do this. Let me do a little test. Okay. Yeah, we can do this. Do you want to get started? You're like, absolutely do it. And I'm like, now that you have the checklist of all thousand recipes, I want you to create a project where we're going to go through this. And I just want you to do five. Okay. So it manually goes and pulls five. And I had to talk through the steps. I didn't like just say, make the process.

26:21I'm like, here's the steps you need to follow. I want you to go and check it, pull the image, do this. And when you create the image, I have very specific instructions. Make sure to re-reference the logo file, re-reference the image file, and then pull the ingredients and make sure that here's the call to action for the bottom. So I had very specific instructions on what these images to look like. So I wanted there to be uniformity across a thousand different infographics. And we created that. And of course, it didn't just have that in the chat. It created an MD file. We created a whole project folder for this.

26:49It created an MD file with the step-by-step instructions. It created a different MD file with the checklist of all the recipes. So we didn't redo the same recipe, right? And I said, great. Now walk through the checklist, do the first five. And it did it systematically. First, it would go and access the WordPress site through the index image, pull the ingredients, check the image to make sure everything was right. So I had a process for checking to make sure the logo was right and that it didn't screw up anywhere. And then it would take it and then actually upload it to WordPress and then check to make sure it got uploaded to the right WordPress recipe, right?

27:17So all that would take a long time. And then I didn't have to automate it to the point where it was like thousands and thousands. I could do a batch of a hundred at a time and spot check the hundreds. But it's a huge project that would take a very long time to do. But with Codex, it took me about probably 30 minutes to get the process and kind of like check it and set it up. And then probably another hour to walk through a batch. And the cool thing is once you're walking through a batch of 100, you can set that thing, let it go, and then come back. Let it go to sleep. Leave your laptop on. It keeps working for you.

27:47Come back in the morning. 100 are done. Now, I did let it go for 100. And somewhere along the line, it had screwed up. So 100, we're gone. It had screwed up 100 times. I'm like, okay, delete all those from the post. Go fix that all. and we're going to have to re-render these. But it's like, I didn't have to do anything. It could go and do the work of fixing all the things and you re-render it and fix the one little bug. As like, one of my favorite things about working with codecs is that every time you're like, hey, we didn't do the thing right, update the docs. This is what you need to do instead.

28:13Best, best things ever about working with these tools is that it can remember stuff. You can tell it and be like, hey, every time we run this process, make sure to add this note or you did it, you did the wrong thing again. Make sure to double check this every single time. Emphasize it. Make it important. And this is how we delegate to humans. This is why it's such a big difference.

28:33Isar Meitis:I want to touch on a few things that I think are very, very important for people to understand, both on the concepts that can be applied, not just for marketing in this case, but for everything, and also on the small little details. And I'll start with the small little details because they're connected directly to the last thing you said. Those MD files, those markdown files, are people hear about agents and skills, and they think there's something magical. It's like unicorns in the forests with spirits that make the agents work. The whole thing, the whole agent universe is these markdown files.

29:03Isar Meitis:And what a markdown is mostly an SOP, a standard operating procedure that's telling the AI how to do the thing. And if you think about it, what is the ultimate employee? The ultimate employee has a set of things they need to do, and they will follow the instructions to the T every single time. and if they get stuck, they will try to figure out a way on how to get unstuck and if they are still stuck or they think something's wrong, they will stop and go and ask questions. This is like the best employee you can define and hire. This is exactly what AI does through this agentic universe. So these MD files are just SOPs.

29:42Isar Meitis:The other thing they are, which I really like what you said as far as lessons learned, I have a skill built, which you can all build for yourself, that writes lessons learned across projects. So I have one that runs one, write lessons learned for the project, but I also have one that says, if you see anything that can be helpful for other projects, as an example, when you generate images, never do this or always do that. When you connect to N8N, always connect in this way. When you are using API keys from third party, like anything that is not very, very specific to that project, it writes to a folder that all the other tools have access to as well.

30:16Isar Meitis:And they all have to read it when they get started. And so then you get It's a learning organization, not just this skill versus that skill, this project versus that project. You're getting an organization that learns from any mistake that is made in any project that's done. The other two things from the best practices on things that you mentioned that are very, very important. One is you work with the AI to figure out the right SOP. Just build the process. Test it in a regular chat inside of Codex. Like, don't build the automation first. Just try to get one right. Then go to Codex and say, okay, awesome.

30:47Isar Meitis:The way we have it now is what I want in the end. Please now write the instructions, ignore all the stuff that didn't work, write the instructions that did work and build it into a process in an MD file that we can follow every single time. Then build another thing that Dan said, build a test loop. Now that it's there, go and verify that it's the right file. It's in the right place. It's loaded to the right folder. It follows the guidelines. Like you can build all of these things into the process itself. And then run a test. And the test is not one. The test needs to be a bunch. you can decide what a bunch is, right?

31:18Isar Meitis:Could be a five, could be 50, could be, but two or three is just not enough because two or three doesn't have edge cases. And edge cases is what's going to break your stuff. So you need to go through enough to go into situations where let's go and take an example and follow Dan's example because it's fresh in everybody's minds. Let's say that one of those folders that the AI goes to, for whatever reason, does not have the description of the ingredients, the actual original webpage. So now the AI goes and it doesn't know which ingredients to include because they're not included on the webpage. It will make it up.

31:50Isar Meitis:So now you have something that is a glass of water, but it has 16 ingredients in the image afterwards because the AI knows it has to make up ingredients. So these are the edge cases you're going to run into. And there could be technical edge cases. Okay, I can only run a batch of 50 and you tried 100 and it's going to get... Like there's many different things that could happen. And so you got to think through these edge cases. You got to test enough to figure them out and then build the right scaffolding in order to deal with them. And then you can do really, really amazing stuff. One of the things I love about the way Codex laid out its UI is that one, we have all our projects on the left.

32:24We're very used to doing that in ChatGPT. We have the chat interface in the middle. And then when you click on these MD files, they open up on the right side, which is very much how Canvas used to work. It's changed now, but like it used to open up on the right side as a doc. The only thing I don't like about it is that I can't edit it natively in this right side. That would be an improvement. But now I could just open it up. It has a little button where you can just open it up. but mine opens up in a sublime text, which is kind of a coding app, but it's fine. And I can just make changes here.

32:49And that's cool. But like, it's just little tiny things like that. I find add up over time. It's just more intuitive and easy to deal with. Have you seen, this is kind of a bonus feature of Codex. It's completely unnecessary, but I still find I like it. It's the pets file. You can wake up. Have you not seen these?

33:06Isar Meitis:No. There's these little characters you can make. And I custom made this one. This is based on a character I developed for my own show, The Edge of Remarker. If you look at the cover art, it has this character I call 404. Actually, he named himself. But these fun little things are these little characters. The cool thing about them is they stay on top of the window. So if I switch to doing something while it's doing some long work, it'll keep me updated on the work that's happening. And then as soon as it's done with something that maybe takes 20 minutes, because not everything takes overnight, right?

33:32A lot of things you're doing, it works for like 10, 15 minutes. But you go and answer some emails on Gmail, and then it's giving you updates. Because no matter what window I go to, it's kind of coming with me, right? And then you just click on and it brings back the codex thing that it was telling you updates about. It's a funny little tiny UI thing that I think is kind of fun because it's this little funny character hanging around with you. But it's called the pet and it's just a fun little nice addition they did. They didn't have to make it fun, but they did.

33:58Isar Meitis:Very cool. It just dances on the screen. One last thing is you may not, I'm doing this for a lot of knowledge work. I'm doing all the kind of stuff I used to do in ChatGPT. I now do in Claude for a lot of this work stuff. In codex. But yes, in codex. but I'm also using it to code and it's so much fun. Now, I think where this really gets fun is for marketers and other business leaders who don't code a lot is to build the little integrations you wish existed that didn't exist before. For example, it's very common for marketers to create social content in something like ChatGPT. But what happens to it?

34:31You might even put it in a canvas, but you're going to have to copy and paste it and move it over to its final destination. It's kind of a pain in the butt, right? So I use HighLevel as my kind of marketing platform. It does all the email stuff and CRM. It does a bunch of different things. But one of the tools it has in it is a social media management and scheduling tool. I was like, great. I'm like, Codex. Every time I say post this to social, I want you to do it through high level. Help me figure out how to make that happen. Of course, it's like, okay, well, we can work through the API. Here's, and we're going to have to code some code files and just make it so, and write the MD file to say like, every time you say this or make a skill for, I probably wouldn't do it as a skill because different clients would need different tie-ins.

35:11I've actually done this for myself and a couple of clients. So that when I'm working in their folders, it only accesses their high-level account. But every time I say post this to social, it works through the API because of the code it's developed. It had to build its own little, not a lot, but like a few code files in order to access the code, tap into the API, schedule it within the tool or post it immediately depending on what I specify. So now every time I see something, I'm like, great, send it or post it to LinkedIn. It just posts it. It's just done. I don't even have to do anything. I don't have to copy and paste it.

35:41Shoot, you could put that on a loop so that every time I run through a process, like maybe after I walk through a podcast, after I finished a podcast and it's got the transcript, I'm like, great, run the process. And the process includes turning into social posts, turn it into a blog post, make sure to upload it to YouTube. It's got all the files and then it just runs through the process. I don't have to tell it to do all the individual stuff. Again, all the process is written down in an MD file. So when I say run the process, it's like, oh, he means this MD file. Let me check. Okay, MD file says access high level.

36:08oh, okay, do this. Okay, oh, for the blog post, I'm gonna have to access WordPress posts. So I could just give the command and it goes and accesses the code we've already developed. So code is part of the repertoire here, but you don't have to be a developer to do it. And these are like lightweight. I'm not building an application. You don't need applications anymore. Codex is the application. It's just develop using code now to extend the functionality. So I'd like, I know there's plugins. I probably need to get into plugins sometime, but I don't really care because I can just build my own plugins.

36:35I can build my own integrations. Any tool I have that has an API is fair game now. It's the easiest way to integrate with anything and everything in the way that you want them to be done. And now you just tell Codex to do it and it can get it done for you.

36:47Isar Meitis:Yeah, amazing. Two cents on that. There are obviously benefits in connecting third-party plugins because they're plug and play and then the AI can use them. There are disadvantages in doing that. One of them is that many of them has a gazillion things you don't need. And then it just fills up your context window with a lot of instructions and a lot of pieces of code that you don't actually use. And that's one reason not to. And I'm not saying don't. I'm saying there's reasons to and to not do this. The other reason, if you are going to get plugins from third parties you don't know, then you're running the risk of these third party plugins doing things you didn't plan for them to do, such as copy your data to a third party server and use it whenever they want to use it or whatever other malicious thing that they may want to do with it, which I assume is not the case for probably 80 % of plugins and connectors in the world.

37:35Isar Meitis:but it's very hard to know one from the other. And so that's another very good reason. If you just need a simple connector again to post something, to do something. And as I mentioned, expand this beyond the marketing world. Think about your ERP system. Think about your HR platform. Think about whatever other software you're using in the company. And I'm not saying go ahead and now develop this, especially if you're in an enterprise. There needs to be guardrails and governance and things you should and shouldn't connect to and change and so on. I'm putting that aside for a second. But from having the ability to do it, It's at your fingertips.

38:07Isar Meitis:I don't write code. I've never written code, not even HTML like Dan did. I just don't know. Not CSS, not nothing. JavaScript, any of this, I don't have a clue. That doesn't matter anymore. Now, to be fair, people who know how to write code can do more faster. Yes. But people who didn't write code and don't understand code can still do incredible things and all the stuff you don't know. And Dan just mentions that before. Okay, tell me how to do this. Oh, go here, go to your calendar, go to this, go to the menu, go to this. It will have a place to create a generate an API key, generate the key. Don't give it to me.

38:40Isar Meitis:Go and place it in this folder, put it in your key chain, and then I can use it. I'm like, oh, okay, that I can do. A monkey can do that. I'll just follow the instructions. And so you can really do more or less any knowledge work you can do on a computer. Codex, cloud code, cloud co-work. Now there's, as of this week, finally, Copilot Cowork, which is basically Claude Cowork inside of Copilot. Gemini is coming up with something very similar. All of these is going to have all of these functionalities built in. And you pick the one that works better for you, that you already have access to, that you trust more.

39:14Isar Meitis:It doesn't really matter. They all can do these things, but they're all a whole different universe than just writing, chat, and going back and forth and copying stuff and moving it around on your own. And so whichever platform you choose, go and upgrade from the basic chat to the next agentic level because it will change your world and your view of what AI can do dramatically. This will be the default way we all use AI a year from now. Until the next time. I'd bet money on it. Even if they have to change ChatGPT, whatever we're all using in the future, it will be like what we're doing here with Codex right now.

39:49Just like, remember, it was only a year ago that we had ChatGPT 03, which was the reasoning model attached to internet search. Think about how much of a game changer. That was only a year ago. And now it's default. Everybody does it. It's just part of how AI generally works now. You don't even think about the old way where it couldn't check itself on the internet. That's just default. That was brand new a year ago.

40:10Isar Meitis:Yeah, crazy. Insane. Dan, this was fantastic. Great concepts and great real practical examples. If people want to follow you, work with you, learn from you. What are the best ways to do that? Best way to do that is to find me on AI Driver Marketer on any podcast platform and linkedin.com slash IEN slash digital marketing day and connect with me there. Those are my two most active platforms. Awesome. Thank you so much. This was absolutely fantastic.

40:48Thank you.

41:18Thank you.

41:49Thank you.