5 Boring AI Automations Making Us 7 Figures

Speaker A: Today we're lifting the curtain on how a real business actually uses AI behind the scenes. Not the flashy social media posts that everyone makes, but the boring internal work that saves us hours and brings in lots of revenue. Like our support bot that answers members questions while we sleep, a database that flags which customers are about to cancel and a course that Claude code built in just 45 minutes. I'm joined by Gail Breton, my co founder, authority hacker, where we show you what actually works for business owners and not what looks good on camera.

Speaker B: Let's jump right in and really the goal of this episode, Mark, is just to show people in real life how we use Vibe coding to run a business that is mostly information work that you know, operations that are not necessarily what people do but rather things they can get inspired from that could help them in their daily operations. And you'll see that most of this is non customer facing and yeah, let's just jump on the first one.

Speaker A: Okay, so uh, I'm going to go first. One of the things which we offer in our product in the plus tier is a WhatsApp community and it's literally just a WhatsApp chat group that I've set up. It's none of the, we don't have any of the business infrastructure set up. It's just like me personally on my phone set that up and I need to make sure that the hundred and something people who are trying to join that A are entitled to like they have the right membership access to be able to do that. I need to make sure I have their phone numbers, I need to make sure I add them to WhatsApp and that I've track in our system in circle, uh, our membership software that they have been added so I don't try and do that again or anything. And I know if I need to remove people in future that that's easy to keep track of as well. So it's a really simple process that I was just doing it manually in the beginning because it's like ah, uh, it takes two minutes, like doesn't matter. But it's one of these things that just the more it comes up the more you do it it's like oh, this is a pain, this is a pain. I wish I'd automated this sooner and kind of bought back some of this time. Uh, uh, so I'm just gonna show you the automation in practice. All I do load up cloud code and if you type forward slash WhatsApp onboarding, it's gonna load my WhatsApp onboarding skill it's gonna pull the data from the application sheet which it's set up, and it's gonna give me the phone number of the person I need to or the people I need to add to WhatsApp. I copy that, I go over to WhatsApp and I paste that in and add them. That bit is manual, but the rest, all the logging, all the templating is automated. And I've just now also set that up as a routine, which is a, uh, triggered workflow. So every day at 6am now, this workflow is going to run and it's going to do it. So I don't have to even, like, think of doing it manually. It's just automatically going to happen.

Speaker B: The one thing you need to watch out with routines is the threads don't show up on your chat history on the left. They show up, uh, in that routine and on your phone only. So. So, yeah, and it's good 6am because it starts your limits early as well, which means you get a shorter session on your first session, which means you can hammer cloth code when you wake up.

Speaker A: All right, so let's talk about how this actually works in practice. And this is the unfiltered reality of how we set it up. And I just want to say, now it's not perfect. So, Gail, I know you're Zapier and I'm like shaking your head in disgust at why I still have Zapier, but I'll explain that as I'm going through it. So we have a form which is embedded in our Circle community in a space that only members and max members have access to. And they update their profile field, basically. And when they do that, Zapier is listening out and it triggers an automation which sends me a Slack notification and adds that data to Google Sheet 100%. We can automate this with cloud code. Like, we don't need Zapier for it at all. When I set it up, we have a zapier account. We're using it for some other stuff. I was like, this is the lowest friction way to do this. Right now I just need to need this set up. This was before I built the skill, by the way. This, uh, was just to trigger notifications so it gets added to a Google Sheet. And I used to just have to go in, check that Google Sheet, update it manually, check if they had the right access, do all that stuff. Now, however, when that skill runs, it has access to my Google Drive through the Google Drive cli, which is like a sort of MCP Thing it's basically has full access to read and write everything thing. It was a bit of a pain in the ass to set up, but, you know, CLAUDE code walked me through it. And once you have that set up, you can read and write things. I think my mistake early on was I thought I could use the CLAUDE connector in there to do everything, but apparently that can only read information.

Speaker B: It doesn't have a right to not write to it.

Speaker A: So that was, that was a bit of uh, a. Yeah.

Speaker B: For the members who are listening. I actually made a lesson on connecting this specific CLI in a new course. And so it's like I teach clis in general, but I use that one because I know it's a handy one. So it's like there's all.

Speaker A: It is all a CLI is really is a way for AI in this case. So CLAUDE code to access all of the functions that are possible inside Google Workspace, right? So it can open a sheet, it can read a sheet, it can write to a cell, that kind of thing. And it just makes it very streamlined, very easy and very quick for it to do. So you can also achieve the same thing using the mcp, uh, as well. So next thing is CLAUDE will read the sheet and it looks up the members circle id. So that's the data we get out of Circ Circle and I think out of Zapier as well. It's like this id. It doesn't mean anything to me. I have to use AI to look it up, but it does automatically it figures out what membership tier they're in and if they are in the right membership tier, then it's going to add it to my list of people to add and I just copy it, uh, gives me everything in a copy pasteable format. I add them to WhatsApp, they're in WhatsApp, I type done and then CLAUDE tags everyone correctly so we know who's, who's in there, updates the member fields. Very, very, very simple automation. And as you can see, like CLAUDE does all these like really minor things like you know, checks. There's no duplicates so nobody's added twice. You know, it makes sure they have the right membership tier, make sure like all the data in circles, uh, added and it's like uh, it's very easy for me to copy paste in uh, because it has like the plus symbol for international phone numbers because like sometimes people don't have that. It's like it just fixes all the tiny stuff that takes me like a few seconds or a minute. Here or there. But that shit really adds up over time.

Speaker B: Here's my question. And that. Why don't you try to use the computer use to have it add people on the WhatsApp app if it runs like 6:00am?

Speaker A: Yeah, I mean, it's probably the next thing I could do with it. Uh, I mean, the honest answer is when I last used computer use, admittedly this was like four months ago or something. It was pretty crap. But I hear it's gotten a lot better. So.

Speaker B: Yeah, and it's not just that, uh, it's like you have time, right, 6:00am, you're probably not at your computer at the time. So even if it's slow as hell, it can probably just do it. I do more and more browser use. Computer use now, actually. It's not perfect. It's slow and it's clunky, but it will do it eventually if you give it enough time, basically.

Speaker A: So, yeah, I mean, that's like very simple, basic skill. I think anyone can build that in, uh, a very short space of time. You, though, have created a slightly more complex skill and then done something quite different with that. So that's probably like the next level up from just building a simple skill. You want to talk to us about that now? Yeah.

Speaker B: So, I mean, as we said, we're getting gradually more complex through this. Your first one was like a simple skill. The next one I have is actually still kind of a skill. And the output is basically like the lesson notes. So, for example, this is the new course that we just released. Like the cloud code means business one, which is an update to our cloth code course, but it has like new work templates. Works a little bit like OpenCloud and Hermes and stuff. But the point is, the entire course inside Circle, which is the platform that we use, was created by cloud code. Only thing I did was shoot the videos and they were edited through the script. And then I basically gave cloud code the link to the videos on the script. And it did everything. It created the lesson notes, it created the illustrations, it humanized the whole thing. So you won't see a single EM dash here. Uh, and it actually sounds pretty good and sounds like me did all of that. Uh, and it didn't create one lesson, it created all lessons, all notes, everything. When the course is done in one prompt, basically, and I can actually show you the prompt, you can see the prompt that, uh, I actually, uh, gave. It was like, hey, I want to create a new course called Cloud Cosmin Business on Circle, not published yet. Uh, create the lessons video 1 to 9 including 2.5. It's like a lesson I put in between. And I did not even export the videos or the transcripts or anything. I just gave it the link to the folder on descript, which is like the cloud video editing tool that we're using. And I was like, for each video I want you to create beginner friendly, non jargony lesson notes that walk them through the steps and the concept thought, uh, and what to do. Well formatted. When explaining concept, create images using GPT images, which is another skill that we have. Like the image that you just showed for example, just basically did that. And I just told it like whiteboard plus hand drawn marker, which is kind of the style I usually go for. And I kind of keep it consistent. For each lesson, create a, uh, thumbnail with my photo. And so again for all the thumbnails of all the videos were also created in one shot. So that means 10 thumbnails plus about 30 to 40 images, but plus about 10 lesson notes. And I said, when you're done, upload everything on circle. Point me to the folder with the thumbnails because the thumbnails I need to upload with the videos at the end. Ask me any question. If you need more context, you can see the outline of the course here. And I gave it a notion link. Again, I did not export anything or whatever. It's just all connected to everything.

Speaker A: And Godspeed, uh, I like how you wrote Godspeed at the end. Does that think that made any difference?

Speaker B: No, uh, it's just fun for me.

Speaker A: This is like a really complex project. Not a task, it's a project. I mean, when we had a team of people before this, there's multiple people working on this for a day or two days, if not a week to pull something like this together. So it's pretty impressive. You're able to do that with just this one prompt.

Speaker B: Now this was about 40 minutes.

Speaker A: Is that basically because the model is Claude code and the model is so good now, or what's the reason it's able to pull this off?

Speaker B: Yeah, it's just able to kind of stay on track longer. And I think that's kind of what you see in terms of the models getting better. So a lot of people are like, ah, Opus 4.8. Is it really better than Opus 4.5, et cetera? It is, but not necessarily early in the context of small tasks. It's like the way I explain to people is you cannot write emails 10 times better anymore. With AI, there's no way to write better.

Speaker A: Diminishing returns on what you can get. And especially if you watch We're Guilty of this as well. YouTube videos, podcasts, uh, where people are talking about the new models, they come out and they have a few hours or minutes to test it. So they go, I'll generate an image, I'll write an email, I'll do all the stuff they've done before. But there's marginal difference versus a big project like this with thought. Uh, it takes a lot longer to actually produce it. So that's where you're seeing the difference.

Speaker B: Yeah, it's like these outputs are saturated, basically, at this point. They will get marginally better with the new models and you'll find a slight improvement. But the reality is, this is where we've progressed in the last six months is like, you need to be more ambitious with your prompts and give them bigger things. It's a discussion. We have a lot internally. How can we just offload more responsibility? This was literally like, this is going to be student facing and you're building this now. The two things that I had to do that was uploading the videos and the thumbnails, mostly because the MCP doesn't allow, not because Claude couldn't do it.

Speaker A: It's just like, it doesn't have the circle. MCP doesn't allow, uh, uploading videos. Okay, that makes sense.

Speaker B: And then other than that, it was adding some screenshots. So, like taking screenshots for some areas. So, for example, the initial install, and you can see the thumbnails are good, by the way. They're not too bad, basically. And like these screenshots, for example, I did manually because we're not at the point where it's good enough to create the right screenshots at the right place, et cetera. So I created a few screenshots, but you can see otherwise, the explainer images, et cetera, they're good. They're like, simple for beginners. Uh, yeah, it's very scannable. It did the formatting very well. It did all of that. It figured it all out for me, basically. And so in the end, putting this whole course together was like Claude working for 45 minutes and me maybe another 45 minutes of making sure that screenshots are here, uploading the videos and uploading the thumbnails, basically. And there's about 10 lessons. There's 10 more that are coming out actually, uh, this week. So this course will be at 20 lessons. But the idea is like, yeah, I just need to shoot videos now. And then everything else is almost fully automated. And it's quality. I don't think the quality compared to what we used to do previously, I don't think the quality has dropped. Uh, arguably it's better. The explainer, images, et cetera. We didn't do that before. So that shows you as well how you can combine multiple skills. Because there's a skill for creating the lesson nodes, there's a skill for creating images. There is like some skills that explain how to use Circle, for example, and it kind of combines all these things into one megaprompt. And does that something I want to

Speaker A: add here because, you know, a lot of people challenge us on like, hey, when I use AI, it just gives me kind of sloth. It's not as good. But I think the difference here is that you have created original content here in the tutorial.

Speaker B: The video. Yeah, right, exactly, the video.

Speaker A: So it's not like making up facts about how to get Claude code up and running itself. It's using your information and it's just like finding nicer or clear ways to articulate points that you've already made. So it's really like a repurposing task rather than a full, uh, generation task. It does why it sounds.

Speaker B: It has a bit of freedom in the sense that it's allowed to expand a little bit. What I said. So you can go and use web search and it's like, let's say there's like some details, some settings that I forgot to mention, whatever. It will kind of correct me in the lesson notes and then make sure the whole thing is better. But it's a little bit of wiggle room, not start from scratch. And I did the planning. I did all of that. I used AI to help me with the planning. Sure. But it's like I was involved and it wasn't fully automated, whatever. What was fully automated was all the boring work, basically all the rewriting things and making sure the notes are clear. Uh, uh, making these explainer images, briefing them, all of that, the prompting. I didn't have to download them, I didn't have to do anything. It just uploaded them for me. Like all that busy work. And I think that's kind of the thing. It's like people aim to make AI create their content, but I think they should free up their time to create content.

Speaker A: Exactly.

Speaker B: And then just have AI do all the boring stuff, all the uploading, all the even editing, et cetera. And then that's how you create content that people still care about. So this course is quality uh, people love it. So far we've had really good feedback

Speaker A: I think as well people are, uh. So people might be watching this and think, oh, I'm not creating courses or I'm not so much of a content creator. This doesn't apply to me. But let's say, you know, an agency or you're in an internal sort of B2B team, right? You are creating content when you have calls with your clients or sales calls or any kind of call that you have a transcript with. You can run a similar process here to, you know, generate your proposals or um, your project plans or whatever else it might be and it'll like. It's not difficult to build something like that. I mean, you saw Gail's prompt there as well, like three or four paragraphs

Speaker B: because there's a lot of skills behind. But yeah, sure, like once you've built your infrastructure, like you need to kind of build this kind of knowledge base. Same like my workspace has a lot of knowledge about the company, the tone, everything. It has a lot of context to use. And that's why I think it reads well. I think it's like the reading, the writing, for example, is very clear and it doesn't feel like it was super badly written by AI, for example, um, because of that context.

Speaker A: And I'm just going to take a sec to do a shameless plug here, but if you want to learn how to build a workspace exactly like that and all the skills that go behind that, then we're actually running a free webinar for you guys next Tuesday. That's going to be the 16th of June. It's at 5pm UK time, so it's about noon Eastern time. Uh, it's completely free. Gale and I'll be hosting it. And we're going to show you behind the scenes on how we build all this kind of stuff, how we set up our cloud code workspace. And it is for business owners who are non technical. So don't worry if you've never written a line of code in your life. We don't write code. We communicate with cloud code in English. Uh, and you can build all this stuff pretty quickly. Um, and we'll show you how to

Speaker B: do that and more. We're going to show you some better ones. These are the easy things. This is the lowest level, basically, like

Speaker A: day one or week one stuff. Yeah, yeah, this is easy. So authority, hacker.com forward slash webinar. And you can sign up for free there. Hope to see you there. All right, Gail, So we've talked about making simple skills, we talked about making some more advanced skills, but something we've been doing a little bit more of is taking skills and turning them into kind of autonomous agents or bots, or. What would you really call these things?

Speaker B: Yeah, I mean, basically moving them. These things they run on your clothes, code. You're usually typing on your keyboard for these things to happen. It's like, how do you make them happen when you're not at your keyboard? And I think you wrap a bit more code around it and you host it on the cloud, basically. How do you do that? And I think that's kind of the logic progression. You start with making a skill, you're like, am I happy with this? Is AI doing a good job, et cetera. And then for some jobs, for the courses, I wouldn't do that. But for some of the jobs we're going to show you, it makes sense that you should not even be intervening at all. It just happens. And then that's kind of like you build this false day multiplier when your attention is not even needed for the thing to happen. Because that's kind of the limitation that we are starting to feel right now is like, we run all of these things on Claude code, et cetera, but you're kind of running like five threads at once and so on, and there's a lot of things happening.

Speaker A: And then it's like you're still like, that's not the problem. That you're running five physical threads in Claude code. It can handle that. It's like it's your brain. My brain is the bottleneck. Yeah, because like, you do five times as much work and that's great, but then you have to think about five times as many things. And even though you're not physically doing it, it's like you're doing less actions, like actions per minute, but, like, your thoughts per minute is like way, way higher. And that is tiring.

Speaker B: Yep.

Speaker A: Uh, so this kind of gets around that.

Speaker B: Okay, so the next part I want to show you is actually our support bot. So it's actually. We have a get help section. And you can see that. For example, people can ask questions and get help, and we help them. We are still here helping people. However, I am not awake 24 hours a day, and I'm not always at my computer. And sometimes people are stuck on things and it's not great to wait for many hours. And the thing is, I often found that people's problems I could solve, not necessarily by my great knowledge in my head. But rather by prompting AI properly. And so quite often I would just answer to people. A summary of what I would get from an AI chatbot connected to the knowledge base that I have also in my workspace and my notes and all that stuff basically. And then I was like, wait, I'm becoming a copy paste machine again. Uh, it feels like added value, but really it isn't because I wasn't doing much. And so what I did is I built this little Age Helper. Uh, Age Helper is basically that. This is my avatar. This is the Age Helper avatar, which is me, ah, as a robot basically. And the idea is, ah, Helper is always here to help you out. And whenever you post a question in the uh, get help section, only then if it's a technical question, it's going to answer. If it's not a technical question, like if someone's like super mad at us, et cetera, it's actually going to skip answering and it's going to wait for a human to answer. So it's pretty smart in the way it decides what to do. And so initially I built it on my computer. So I would call my computer and my computer would wake up and it's built on codecs and then it was connected to a, uh, whole prompting system. It gives these pretty cool things, these pretty cool instructions and stuff. Sometimes it even gives prompt for people to give to their cloth code so that cloth code solves the problem for them. We build all that stuff and the idea is I get back to them and so how is it built?

Speaker A: Just to be clear though, so you had this system internally that you were using to answer a lot of questions before?

Speaker B: Yeah, I did it manually. I had a skill. And then I was like, I'll make it a bot. It ran on my computer. And then eventually I was like, well, what if I'm traveling? My computer is off, I run on a laptop, I don't have a Mac mini.

Speaker A: So from a skill to a bot to an autonomous bot, essentially.

Speaker B: Yeah, pretty much that's where I'm at.

Speaker A: Right.

Speaker B: And now I'm actually expanding it. So I'm building like DMs into it, for example, so people can DM it and uh, not necessarily share publicly their issues and it can help them because people are liking it quite a lot actually.

Speaker A: Um, and so I want to be very clear as well, because I know a lot of companies will be like, oh, I don't want to automate my support. That's a bad experience. They have horror stories from big, uh, companies like PayPal that you can never get through someone and you just keep getting automated answers, things like that.

Speaker B: Ah, yeah.

Speaker A: Well, we've fully disclosed that this is, uh, AI Helper. Like, if you come out of this, Gail, there's a big image at the top of the get help section that explains how the system works. Like, you'll get your AI Helper reply and then we will come in and reply when we're awake.

Speaker B: And you can see it jumped in, right? It's like I jumped in later and I was like, yeah, I actually agree with what the bot said.

Speaker A: So, yeah, so people are getting the best of both worlds. They're still getting the same answer or similar answer to what we would have given anyway. They're still getting the mark or gale kind of oversight and they're also getting an instant answer as well. And so when you present it like that, then it's a win, win for everybody.

Speaker B: Yeah, I agree. And it's like, for example, Detroit, this guy. And actually this was before I fixed the formatting issue with the new bot. But this guy has this complex question. And then AI Helper gave him this answer, uh, and gave him the prompt to give to Claude code for how to build his strategy and everything. And literally people are like, oh, this is super helpful. I hope to learn how to hire someone like you later. So it's like, it's actually a good thing, right? It's like, uh, I think we get

Speaker A: away with it a little bit more

Speaker B: because it's an AI thing.

Speaker A: Right?

Speaker B: I agree.

Speaker A: You do have to be a little bit careful with that stuff. But I mean, customers getting instant answers that are good.

Speaker B: Yeah. And then I DM'd him.

Speaker A: Very positive thing.

Speaker B: Then I jumped in, I DM'd him a skill that I use for strategy. And I was like, try to use this. And the point is, this guy had to. Anyway, let's go back to how this works. I don't think people care too much about how we run things. And so how does this work? It's like, yeah, member asks. And now what I've done is actually the, um, Codex instance that runs this is on the DigitalOcean server, which means that it's always on and then my computer can be off. And it is just a terminal on a DigitalOcean server. I think we pay $12 per month. I basically operates that and I've worked quite a lot on protecting it from prompt injection and things like that. So, for example, uh, it does not have many tools. It can only do web search and, uh, the context of the membership and all the notes that we gave it for knowing what to say. And the knowled is actually gathered by a third party Python script. So it's kind of like gathered by a script that runs before and then it gives it to the bot as context to use basically. And then the bot can use that or not use that. Then what happens is the bot can only draft an answer and it doesn't even interact with circle. So someone cannot prompt, inject it on like, you know, do some crazy things on circle. And then what happens is just it passes text to a script that uh, then post the answer, which means the script itself cannot be prompt injected and can only do one thing, which is post an answer. It cannot interact in any other way with circle. So the point is, I've thought a lot about security on these things. M if you guess in the membership, you know, I'm a bit stupid on security maybe, but I care a lot about that. And it's not just a random codex

Speaker A: thing because stupid is the right word. I think paranoid is the right word there.

Speaker B: I'm a bit scared. So basically I'm careful. Uh, and so the point is like

Speaker A: stupid would imply you don't care about it.

Speaker B: I care about it a little bit too much sometime. That's kind of the way I built it. And it's quite interesting because it has a filter. So the AI before it answers basically decides if it's going to answer or not. So in the output it says post it or don't post it. And it will help with technical issues, for example, and finding resources on the thing. But if someone's like hey, I want a refund or you guys suck or whatever, um, then it's not going to start replying.

Speaker A: No one's posted that so far though, right?

Speaker B: So far it hasn't happened. But who knows, that might happen one day. And if it does, in principle it should not respond, basically. And so that's the idea. Now one thing that's really kind of cool with this bot is as I said, it runs on codec cli, which runs on the subscription, which means this basically runs like this whole thing runs on 20 bucks per month subscription. This runs on GPT 5.5, so it's a good model. It's expensive as well. And that's why the answers are so good. But the point is, if we paid API prices, we'd probably pay hundreds of dollars. And because I actually built it on a server that can run a terminal, I can connect it to a basically subsidized subscription from OpenAI and therefore reduce our costs significantly.

Speaker A: Is that kind of kosher or is that a bit.

Speaker B: Actually OpenAI is kind of cool with it. So if it was entropic, they would not be cool with it. But Sam Altman came out and if you pay for subscription, you can use it however you want and they're fine with programmatic use or whatever. So it's like so far so good. It might change in the future, especially as need to turn more of a profit, but right now it's worth doing for us. And that's kind of a tip for those of you who are, who care about that, is that if you run on um, something like a Digital Ocean server, you could run the terminal, you can run a subscription and you can save a lot of money. Like a $200 subscription would give us so much. Eventually I could imagine us expanding this, paying $200 a month and get basically unlimited usage for all these bots, for all our needs, our uh, automated needs, basically. That's kind of how I see it. And then eventually if they stop it,

Speaker A: does that have to be a uh, full time 247 server? Would that work on like a Cloudflare workshop?

Speaker B: No, it wouldn't. I don't think it would work. Maybe it would. I mean I'm not a programmer, so I'm like a crazy way to make it work but like out of the box you kind of need like a terminal. Uh, and so that this runs on Linux and it's a terminal and everything, but you can run it on very cheap server. I think the server we run is $12 per month. So.

Speaker A: Yeah, and just to be clear as well, like we really don't know the ins and outs of all of the like, you know, Cloudflare and Digitalocean and things like that. It's like Claude had or Codex has built it for us.

Speaker B: This is a proper vibe coding podcast like we're talking about. Like it's like even a lot of these logics on. Like I was like, let's just make it secure, let's brainstorm ways to make it secure. And it was like, okay, well we need prepare for prompt injection, so we're going to gather the context outside. We also won't make it take any action. It's going to write text basically, so it cannot do anything.

Speaker A: You're not starting by saying, hey, here are the five steps I want you to follow. What you're doing is you're saying, I want it to. This is the outcome and I want you to think about security.

Speaker B: Yeah, I'm scared of prompt injection. How do we fix that? Uh, update the plan and then basically we'll do some research and do it right.

Speaker A: It's almost like a good way to think about it is like you're the like, uh, non tech savvy CEO talking to the CTO who has to then go away and build it. And that CTO is now Claude or Kodak.

Speaker B: Yeah, exactly. You come with your questions and your problems and then you have it walk on solutions. And now is this going to be like enterprise grade solution? No, but for a bot that just answers on a private community, it's fine basically.

Speaker A: And something that you said to me last week actually was like, try and push the envelope out a little bit more. Because I think a lot of people, they have an idea of what they think I can do. Especially if you're kind of a bit of a perfectionist and you're quite particular about how things are done. Sure. Um, a few people can relate to that. Then you kind of like tell it what it needs to do. But if you just give it the goal and the objective, you might be quite surprised at like how capable it really is now.

Speaker B: And what I like to make it do as well is I'm like, oh, instead of trying to come up with things, just do some research in GitHub repositories and how people are addressing this problem. Right. So it goes on a bit of a web search or it sends a sub agent that does a bunch of web search and that comes back with useful context that then guides its technical decisions. And so you're kind of piggyback riding technical people that posted stuff on the Internet, basically.

Speaker A: So I'm not using Codex very much at all right now, but I'm um, building most of my stuff in Claude code. Sometimes when it does that, it comes back with like, you know, quite technical programmer Y type language. And I get a little bit like, uh, I don't know what this means.

Speaker B: Just tell him that.

Speaker A: Quite often. Yeah, uh, quite often inside my chats I just stop it and I write Eli 10 what this means. Like explain this like I'm 10 years old what this means. And it makes a very nice analogy and it's just like, oh, okay, right, I get that now.

Speaker B: Exactly. So a lot of people do that, right. They get an answer from AI that they don't understand and they're like, oh, that's it, I'm in technical land, I can't do this. And it's like they give up. But Actually, it's just like, hey, I have no clue what you just said. Explain again. And quite often as well ask me questions. And it's like I have to make decisions. It's like, oh, do you prefer using, uh, this, uh, crazy, uh, branch of this thing, or do you want to create a custom cli about the API? Crazy technical terms. I'm like, I have no idea what you're saying. Just rephrase this please, in a simpler way.

Speaker A: Another good one is like, all right, now we need to set up the server and do this. And that would be step one. Step two would do this. And the answer is do it yourself or just do it for me.

Speaker B: It does that. It gives you work and you shouldn't do it. Ah, it's like I'm like, no, no, no, I don't do anything. I stay on the trap. And uh, yeah, I think that's one thing we talked about as well. It's like you install the terminal tool and then you just have it operate the terminal tool and you don't leave your chatbot. The truth is, the most of this customer support bot I built on my phone, I built on the Codex app by just talking to it like a walkie talkie while walking my dog. And most of it, actually a lot of my Codex building time is on my phone now and just kind of talking into it at the gym, outside, whatever, or even on my sofa end of the day when I'm super tired. And I just get ideas for features. So for example, last Sunday I was like, ah, would uh, be great if people could dm, um, and kind of talk about more sensitive topics with the bot because it's pretty helpful. And then it's like, I already have a feature that's set up now. It's not live, but it works. And then it's like I need to test it more. But like, uh, eventually.

Speaker A: Are you reading the response it gets or do you have it like text to voice into an AirPod or something?

Speaker B: I like to read like I'm a reader, but everyone to their own. Like if you wanted, you could do that. Uh, but yeah, anyway, that's the support bot. It's kind of handy. And again, that could be answering customers emails. That could be, uh, helping with technical questions. That could be anything. Like, eventually I want to make it almost like an API and I want to plug it into all our, uh, customer support channels basically. So that's a vibe coded project that went all the way from a skill to like a proper bot that runs nice.

Speaker A: I want to move on now and talk about something which I first heard about when back in the day when I had a real job in a large enterprise company. It's called data warehousing. And it's one of those terms that sounds like kind of corporate bullshit. Bingo. That I just like, oh, send me to sleep now, kill me now.

Speaker B: It sounds boring.

Speaker A: It sounds boring, but it's super cool when you don't have to do any of the work behind it and you just kind of give it the guidance to do it. So do you want to explain what the problem is we were actually trying to solve with this and how we built this warehouse of all our data

Speaker B: and I actually have this little image that was prepared for that. And basically it's the idea of creating one database that has every single event that happens in your company. So if you look at what we do, for example, we collect payments on Stripe. When people check out, it comes there. People use our product on Circle, which is like a community platform. So whether they watched a video or they logged in or they liked something or whatever, that data lives there. Then we do email marketing with a tool called Bento. Again, they have all our open rates, click rate when people subscribe. We also track like, you know, when subscribers visit certain pages on the website, for example, that kind of stuff. It's pretty handy for like signal of whether people are going to buy or not. Uh, we have all our support questions on Help Scout, which is another support platform. And then we usually collect kind of like forms through an app called you Form. Right. And then there's even more.

Speaker A: The reason we use you form, by the way, is because we got this amazing lifetime deal on it many years ago and it's been brilliant.

Speaker B: The truth is, you could vibe code a U form in about 30 minutes I think now.

Speaker A: But it's fine, 30 seconds now probably. But, um, honestly, it's so well built and so nice and it integrates with everything. It's just like cool, we'll just use that.

Speaker B: But the truth is they even have

Speaker A: an AI agent that I now get prompts to copy paste into that agent and it just builds the form for me as well. So.

Speaker B: Okay, okay, not bad. But yeah, still, you could just vibe code.

Speaker A: I'll see you form, by the way.

Speaker B: Yeah, yeah. So anyway, we have all that business data living in different places. But the thing is like the story of a customer lives across all these platforms. So it's like they might have some support tickets. Then we have some activity data here. We know what they paid on Stripe. Et cetera. It's kind of difficult to tell, for example, is a customer about to cancel, for example, which to tell that you need to be able to know when their payment plan ends, for example, their level of activity. Are, uh, they subscribed or not to the list? Did they complain about anything on Help Scout or is there any activity from any form that we have? And then if you combine all that data, you can start telling the story of where this customer is at with your product. But individually each of these things doesn't really help you tell the story. So what we Vibe coded was basically just a simple way to send all the data of all these platforms into one big company database that we host on Supabase. You can start for free. We pay $25 per month for that, uh, right now, mostly because we actually process quite a lot of data. Uh, but it's fine, we will be able.

Speaker A: It's crazy compared to the hundreds of thousands of companies pay for something Oracle or one of these uh, ENTERPR Enterprise database systems.

Speaker B: And that's kind of the thing with small businesses, right? We've always had all that data in all these tools, but we never really made use of it because it just was too much time, too much effort and too difficult technically to do it right.

Speaker A: And because it's in different places, yeah, it's difficult to do.

Speaker B: It was difficult like connecting APIs, all of that, it just breaks and you have to update it, you know what I mean? And so now we can actually go on cloud. And for example I said yousuperbase, which is the database, and tell me how many member subscriptions are expiring this week. And I can tell exactly that. There are 14 memberships expiring this week and I can tell one that is expiring. So they already canceled. There is four that are renewing for sure and these are actually unknown so I would need to dig deeper. And then I could say like, okay, which of these members are at risk? For example? And what it would do is it would get the data from Circle and it would say which ones did not log in for example in the last 14 days or something. And it would just kind of highlight them to me and ah, I could say then please email them and propose them to have a free catch up call with me for example. And then all of a sudden I can reactivate these people and they can keep paying me. And that was literally three prompts on claude code. Now it's connected on the database. And so that's what this database does is it Allows us to get visibility and take action. And not only us take action, but the AI takes action on um, people basically. And we can just optimize our revenue just by being helpful to the right people at the right time, basically.

Speaker A: Yeah. And because that now Claude code has access to that anytime we do something that touches members. So a good example would be the member survey we did a couple weeks ago. So we got about 250 people to reply, fill in like eight or nine questions, multiple choice. And there was usually we used to do just multiple choice questions on surveys. But now because we have AI analyze it, we had like an open text field for, you know, like what could we improve, you know, the questions like that. And it was able to not only analyze that but cross that data against with what was in the SUPABASE database. You know, what type of member are they, how active are they? And you know, we can see like what are the active people doing?

Speaker B: What are they lifetime value as well.

Speaker A: Spenders.

Speaker B: Yeah, exactly. So it's like you don't just take the survey output, you just actually can weight it for how much revenue that is for you and which segment you should optimize for that kind of stuff. And that changes everything. Basically you don't just read the data.

Speaker A: Uh, instead of Gale and I, which we were actually doing, having quasi arguments about like, no, our members need this, this is what they want, this is what they hate, this is the problems. We're like, why are we, let's just ask them and then have AI figure it out. And it did. It was good.

Speaker B: Yeah.

Speaker A: Being able to get that level of insights about your business and your customers and what they need and where you should take your product is invaluable. Every business needs to do that.

Speaker B: I mean consultants pay like charge literally tens of thousands of dollars to do that. And all we did was like a

Speaker A: Google millions of dollars, uh, millions of dollars for like McKinsey to go into an enterprise.

Speaker B: Most of our listeners don't hire McKinsey, so probably the level of consultants they will hire will be tens of thousands.

Speaker A: Uh, I'm just saying like the same way, um, building a lot of infrastructure with code was out of reach for small business because code used to be expensive. Now things like having a, uh, it's not really a consultant, but like this level of data analysis, it was completely out of reach before because it was so expensive. But now everyone can do it. It's a great equalizer.

Speaker B: Yep. And just to explain quickly the technical way, how it works is basically every time something Happens. There's something called a webhook in these apps. So, for example, Stripe sends a webh, and it's like it sends data to a cloudflare worker. Um, and that cloudflare worker basically queues it to update to Supabase, and then every 15 minutes, depending on which APIs, it will actually query the API to see if it missed any event, basically, and just kind of update itself. And the idea is the data is constantly updated every time someone likes a post on Circle, every time someone logs in, every time someone clicks on an email, every time someone opens. And we can just reconcile that into member profiles and we get like, full data, basically. It's very, very cool. You should do that if you have any customer data. And one thing about the survey that you didn't mention as well is the survey was only like, seven or eight questions. It was very small, which was easy to do for people, but at the same time, because we could cross it over with the data, we got an insane amount of insights, actually. And so it goes well beyond what you imagine when you do that. And you don't even need to know what you're doing. What we did is we're like, here's the survey output. Use the database and figure out what people want, basically, and focus on and help us segment people, uh, by revenue activity, et cetera, and do all these segments. And that's it. And we got it. So, yeah, that's pretty powerful. So, yeah, you can ask all these questions like, who might cancel soon, who applied to our, uh, plus level was that person's full history. And, yeah, cloudflare, we pay $5 per month. Supabase, we pay $25 per month. And these APIs are usually included in whatever tool you're paying for. So it's not expensive, basically.

Speaker A: And the interesting thing about building on Supabase is you can actually kind of build software within there. Let me sort of talk you through an example here. So they have something called Edge functions, which is, as far as I understand, a little bit like a Cloudflare worker, but it's kind of built directly into Supabase.

Speaker B: Yeah, that's correct.

Speaker A: Yeah. Am I right there?

Speaker B: That's correct. That's correct. It's waiting to be considered. You don't even know what we built. It's fine. It's true.

Speaker A: Yeah. This is the point. Like, I don't, but it's working, right? But I still have this work. Like, I built this yesterday in, like, couple hours. So we now have a cancellation or refund form on Our website again, shout out to Euform. AI, uh, gave me all the prompts to build this. So when somebody fills that in, uh, and they want to cancel their subscription, we get some data, we ask them, like, why they're canceling, and, you know, what we could have done to convince them to stay. Very, very useful information to have. So what happens is they'll fill in that form. Um, we have a quick check to make sure it's not a bot. And you know, that form's not getting spammed. Um, but then it sends that data into Supabase. But the edge function within Supabase is essentially like this agent that spawns up and starts doing a few different tasks here. So it's going to look up all the data about that member. So, for example, which courses have they taken, which lessons have they been through, how much have they paid, how long have they been a member, like, those types of things. And it's going to use all of that data to then generate a response to their cancellation request. So, you know, if they just want to cancel, like, hey, this isn't for me, I'm not interested, Fine, I'm not going to try and stop anyone. But if they're like, oh, it was too difficult and I couldn't figure out how to do this and this, then that's an opportunity for us to save that sale and to offer them a solution to that that may be directing them to the relevant content that might be, you know, some kind of discount or promotion or some kind of offer that we can give to people that way. And being able to do that in a highly tailored and personalized way is really, really valuable because, you know, we've all been through these cancellation processes before, and it's like you get a generic, hey, you know, you're trying to cancel, here's a free month or something like that. And it's like, you don't really care about me. This is just your automation. But in this way, we're giving people a more custom kind of experience of this and kind of really helping them to solve their actual problems in there. So how that looks at the end is when someone fills in this cancellation form, it has this decision logic in here. So, you know, are they eligible for refund? What's their, the kind of monthly billing, uh, cadence, you know, those types of things, what's set and what's not? And then it's not going to send them, it's going to draft them a reply. And so this is a super, super basic reply, by the way. I just built this yesterday, but it shows you what you can do with it. And I can just go in there and edit that as I want or just send it as it is and we're kind of good to go from there. The edge function itself is receiving the data from the Uform. So from the ticket it's using GPT 5.5 to draft a reply, using all of that data and then sending that draft into Help Scout, our customer support software. Uh, and so all of that, which would have taken me, you know, five, ten minutes to do and I wouldn't even have been able to do all of the stuff in there because there's no way I'm going through all of their data and all of their, you know, view history and lesson history every single time. It's like it's not feasible to do that. It does all that in literally like seconds. Uh, because the uh, edge function sits right inside Supabase so it can access the data really, really quickly. There's like zero delay there basically. So yeah, this is something that I built on top of the database that Gael set up there.

Speaker B: Yeah, and that's the thing. We can now start building all this stuff and do that basically. It's really cool. Just wanted to say one thing about the model selection because people were like, oh, why are you using GPT 5.5? And the reason is because actually GPT 5.5 is actually a very, uh, efficient model. It's very cheap if you run it in low reasoning and medium reasoning and I actually have some data to prove it. So this is artificial analysis and a lot of People run on 4.6 Sonet, which is kind of like considered like a cost efficient way to do it, which is not much cheaper than OPUS anymore by the way. OPUS is almost the same price was 4200. But if you run GPT 5.5 in medium or low, you can see DPT 5.5 medium did the same test for $1199 and low did it for $500, which is pretty good. Uh, and it's actually like almost on par. It's almost on par with Gemini 3.5 flash, actually, if I put it in perspective. So If I put 3.5 flash in there, actually 3.5 flash high is here. Yeah, it's actually higher. 3.5 flash is $1,500 and GPT 5.5 medium is 1,200. So actually 5.5 if you use it as like not maximum reasoning is actually a pretty good cost efficient model. That writes pretty well and is pretty smart, basically. So my recommendation right now, I know I used to recommend Flash, but Flash got expensive, is to actually use, uh, 5.5 in low or medium reasoning for bang for your buck for API, actually. So, yeah, it's like that's the kind of vibe coding projects, I guess that was the last one, right?

Speaker A: Yeah, I mean, that's everything. We've been through quite a lot there, but just wanted to kind of give everyone a, uh, practical behind the scenes look at the stuff we're actually building, hopefully to give you guys some similar ideas really.

Speaker B: And it's the stuff that's not flashy. It's kind of internal mostly, but that's what, uh, I mean, it's allowed us to run this whole thing, almost just us, because literally the product is getting a lot of assistance, the support is getting a lot of assistance, and then in the end we still have to kind of manage a lot. And we're in a process of automating and scheduling a lot more things now. But in general, that's the kind of vibe coding projects that you should undertake in your company that will make a difference. It doesn't have beautiful front end, it doesn't have any of that, but it actually will save you a ton of time and you just need to learn a few things, like learning Supabase, learning Cloudflare workers, for example, and then playing a little bit with APIs.

Speaker A: Just to be clear, you don't need to learn that. You need to learn they exist.

Speaker B: You need to learn what it does.

Speaker A: Uh, yeah, what it does so that you can kind of talk to AI and it will do it for you.

Speaker B: Yeah. So most of these things I could see when you presented the Supabase stuff, you basically didn't even know.

Speaker A: I don't really know what it is.

Speaker B: It's probably worse. Like, you know all these tools, both Claude and Codecs, they have, uh, security review plugins. It's probably worth running your code through that if you're vibe coding these things. It will help. There, uh, are special tools as well. There's tools like ColdRabbit, et cetera, that help catching issues. But the truth is the best models, like, don't use a cheap model, but the best models, they have a lot less security flows now than they used to have. Like, it's kind of like nightmares of vibe coding, et cetera. It's not as bad as it used to be. Doesn't mean it can't exist, doesn't mean you shouldn't check, but it's less much, much less likely than it used to be even a year ago.

Speaker A: And you can see we're still quite cautious around things. Ah, it's not sending messages and support, it's drafting them. It's not, it has no access to payments or, you know, anything like that. Uh, so, you know, there's, there's some hard coded like walls that it just can't do and can't get access to. But we're generally quite cautious about this stuff. We're not vibe coding our shopping cart software using like, you know, established companies to do that. So yeah, I mean, I think sometimes what pisses me off is on social media, even on YouTube, a lot of the examples that people share are like, here's how I create content or do some like flashy visual thing using AI. But the truth is, 90 of businesses out there, they're not content businesses, they're not social media influencers. They have a product or a service and they kind of like deliver that. And it's, it's kind of a bit boring sometimes. And it's that boring stuff that you need to work on automating, because that is the core of your business.

Speaker B: You automate this stuff and then you make content still involving yourself because I think that's the hardest thing. Like, sure, you can do AI slope, you can make like a baby podcast and then you get lots of views on TikTok. We won't make much sales. If you want to make sales and if you want to sell stuff, it actually is worth, uh, doing the human stuff that people connect with. AI assistants, but not necessarily fully automated. I don't make a clone of myself on heygen and post videos about that, for example, and then just basically take everything else away. With AI, that part is almost invisible to people. And so you're not dropping your quality, you're not discounting your brand or anything like that. And things like the database. Yeah, if anything it makes the experience better for people. So that's what you should focus on. I know we get a lot of people like, uh, how do I make content with AI? I reluctantly do a little bit of it, but I'm not a big fan because I actually think that's not what they should focus on. Uh, what they should focus on is automating what they sell and all the service layer around what they sell. And then the content part is now you have time to do it and then you actually have a chance. Because most people who fully automate their content with AI point to me, apart from a few YouTube channels or TikTokers or whatever. There's not a lot, but even then

Speaker A: they have the high views. But are they really making a lot of money and selling a lot of product off the product back that everyone

Speaker B: dreams of doing that? I mean, sure, with SEO, sure, I'm sure you can point at some of it, but for anything else, I'm sorry, but show me a real business channel that makes a lot of money with the fully AI generated content. However, show me businesses that vibe code these kind of projects. We show many of them actually. So yeah, that's the final words of wisdom, I guess.

Speaker A: Okay, brilliant. Well, thanks for tuning in to this episode of the podcast. We're only really able to go in depth depth with this stuff in a long format piece of content like a podcast like this. So thanks. It's an hour long. Appreciate you staying to the end. Make sure you're subscribed because we're going to be doing a lot more episodes like this. If you enjoyed it, head on over to our YouTube channel. Um, just search for Authority hacker, find the episode there and leave us a, uh, comment. Tell us, do you like this type of format? Do you want us to show us more quote unquote boring automations, uh, from our business and things that you could do as well in your business business like this? We'd love to hear your feedback on that. So thanks for listening and we'll see you again next week for another episode.