Transcription
AI coding today is powerful, but it keeps us in the loop every step of the way. We're starting to see tools that allow us to hand a task to AI, and what it hands back is a completed product. This is really a glimpse of the future today.
Now, this isn't about replacing interactive coding. We'll always have that. This is about expanding it. You can always pair program in your editor, but now we're going to be able to hand jobs to AI and they hand back finished product. The promise here is simple. AI doesn't just help us write code. It can take on full tasks, running the tests, fixing what breaks, and returning an entirely complete pull request for us to approve.
I'm excited about today's video because I'm going to show you a way to work with AI that actually allows you to scale yourself. All of these efforts, I know that these AI coding systems are making us better and faster and more efficient. But what if there were more of you? That's really what we're going to dive into today. Where are the tools? How good are they? And can you start using them right away? Let's dive in and find out.
Okay, before we dive in too deep, let me show you what we're going to be working on. There's this little application that I've built for the NFL season this year that's just basically a schedule system. We're going to be looking at this in just a second.
The first thing that I want to look at is GitHub itself. You might think of GitHub as a place that you store your code, just kind of a vault in the cloud that you can get your code, other people can see it, that kind of thing. And it really is fantastic at that. Of course, we all know that. But it has a couple other features that I want to call out here just in case you're not aware of them. We're going to get through them very, very quickly. Go research them on your own. They're very well-known and easy to understand.
But the first thing here is it's possible to work on multiple what they call branches, which basically means our code is almost always on the main branch. It's just the standard normal original branch when you create a repository, but you can go in and create a second branch that you can make as many changes as you want and it doesn't change or disrupt whatever's on the main branch. You can think of it as a version. So each one of these branches can be a version of the code and at some point later in the process once you get to your your branch to a place that you feel like it's ready you can create something called a pull request. And a pull request is a way to say my branch is ready take a look at the code verify it make sure that it's right and then merge it in or pull it into main branch. The idea of being that you eventually end up with a strong main branch moving forward.
The third thing that we're going to cover really the last thing that's of import here is issues. Now, issues are a way for you to be able to kind of create tickets on a repository might be one way that people would be able to understand it. It's just a place that you're going to be able to come in and say, I've noticed something about this application. I wish it would do this more. We will use it a little bit different from that. And many people do use this as a full ticket system where you might write a story in one of these issues to say, "Oh, I want the application to start doing this and here's my definition of success." That's actually the way we're going to use this in just a second. So let's dive back to the application so that we can understand what we're going to ask for.
Okay, now the fun bits. So this is the application that we're going to change. As we talked about, it is just a simple application that shows the schedule of the NFL games coming up by week sort of thing. If you click on a tile, you'll see that it has this button menu bar at the top that allows you to select your level of excitement. So I might want to watch this one or I really want to watch this one or I don't know if I care to watch this one at all. So, what we're going to do in this case is we're going to ask the system to change this. And instead of just having two potential states, you'll notice that the outline kind of matches the color of the button. That's one of the details. We want to go and use some circles or kind of bullets that you can select from four or five different colors. That way, we can have you you create your own little definition of of what a watch interest might be. That's what we're going to ask to do.
All right. And before we get to the editors themselves, I wanted to show you very very briefly kind of what we are trying to affect here. We've all worked in this interactive kind of agentic coding experience that we're talking about whether you're using cursor codecs codec cli or or something like clawed code. These are all these agentic patterns where you're a pair programmer with an AI and moving forward with the code. So you can see here it's a human and an AI building some code iterating to the next step and the next step again and again. We we're all familiar with this kind of pattern.
What we're about to introduce is what I'm going to call a bot pattern. Just the easiest way to kind of conceive of what I'm talking about because they're not all exactly the same. But essentially what we're going to do is give a task to an agent alone and let it iterate until it completes the work. Now that sounds great. sounds like what we would like to be doing with something like clawed code more often, but this has a particular use and a particular value.
So, let's talk about the value first, right? The biggest benefit of being able to use these kinds of tools really gets down to asking yourself, am I using my own techniques to help empower me? So, in other words, we're all using AI to write code faster, maybe better, maybe more completely, maybe messier. However you conceive how you're using AI, we're using it in that way just to write code. But that's akin to it basically being our pen. What we want to say is, can I get AI to do some of my code? Can I offload some of my work? Right? That's what we're all really kind of asking. And this pattern actually allows us to start to do that. And it actually works. And I'm going to show you that today.
But what we're talking about is being able to do this in multitude. You cannot scale when it's you helping an AI build something. You're the one that's going to be the bottleneck. If we start employing more and more of this bot pattern, if you will, this asynchronous agentic pattern where we move out out of this loop until the very end, then we can have a lot more software built all at once. And that really is our scalar value.
Okay. So, what do we know? What we're going to do is we're going to take this screenshot here of this tile so that we can give this to an AI and tell it the problem that we're trying to solve. Come back over to GitHub itself in the NFL repository that we're dealing with. And we're going to create a new issue here. So I want to create a new issue. I'll come down into the body. I will paste that image. That's this screenshot here. So now the screenshot's in here. And the other thing that I'm going to paste in is a big PRD. So where did I get this PRD from? This PRD is me saying to chat GPT, here are the changes that I want. In fact, I think I even gave it a version of the screenshot as well and said, here's all this screenshot information, and what I'm looking for is a set of buttons. All I really want is those little dots up at the top. And Chat GPT came back and said, "Yep, here's a reasonable, very succinct, cuz I wanted it very short PRD that you can give to something." So now we have kind of a new issue on this repo. But what are we going to do about that? nothing is going to take off and do any work for us here. So, we really want an AI to be able to do work here.
Okay, I'm going to start in Claude Code. The ones that we're going to go through are Cloud Code and Gemini CLI because they have actual slash commands that we can then set up um the integration with uh GitHub. Here you see install GitHub app. Once you move through this and and hit the button to say yes, I want to use this repository. What it's going to do is open up a website that is going to configure your repository with Claude itself, with your account against Claude. And that's where it will share keys behind the scenes. And this is what is really important for you to understand. It's sharing an API key for you. And so everything we're about to do, big call out here, is not free. These are going to be things, one of them is free. We'll get to that in a second. These are going to be things that you're going to have to pay for as they work. So, it is not a solution to your coding woes yet, unless you want to get back into paying the rating cost for your API usage. But to to kick the tires here, it's not terribly expensive to kind of show this off and really has some value. Just hit the configure button and then that will drive you back. You can come back into Claude Code and once you update with the the workflow information, it's going to save things into your local project. Those things need to be pushed up to GitHub for all of this to work.
So I'll show you this in Gemini CLI and then we'll take a look at what was added to the repository really quickly so that you kind of have orientation of what's going on. Okay, same sort of thing here in Gemini. If you hit slash, you'll find setup GitHub. Selecting that will push you through a set of sequences that are similar though different. And I will point out that if you're doing Gemini, you're going to have to set up some keys here yourself. that isn't done for you like Claude was just doing for us. So, you'll need to go over to the the workspace in AI Studio in Gemini itself to get a key. I have a free account and I'm able to use this. So, this does work. Now, I will give you another clue at the end of this that uh Gemini does not quite make the code change that we're expecting. So, it's not quite working though. This is the intent of this whole process. So, it should work. I just wanted to show it to you just in case it's working for you and it just happens to not be working for me. But this is how you do it and it is free or relatively free. I have a free account and it works. If you come back to Gemini, you can also see that they give you a URL here to take a look and understand how their system works and what it's supposed to be used for, those kinds of things. But let's take a look at how we're supposed to use them now that we have these set up. Really, it was that easy. Now we have coding agents, bots that work in the cloud.
All right, here we are back in our repository. Of course, very simple. This is the issue. This is what we want changed. If I scroll down to the bottom, I didn't show you the first time. I had already set this up. So, this happened already for us. Um, the Gemini CLI had already kind of come and taken a look at things and poked out its head and said, "Oh, I'll take a look at that for you." It does not do the work like it's supposed to. Whole different issue. So, what we're going to use here is claude. They both work the same way. You should be able to do at gemini cli and tell it to to do this issue and that will work. And if you just submit this comment again against this issue, this is kind of the key phrase that it's looking for is just its own like at mention. But what we're going to do is at Claude here, same kind of thing. And we will say add a comment. Let Claude take over. You see that it turns into a known name here. Claude puts eyeballs on it. They have very nice ergonomics. Again, kind of really paid attention to how they interact with you. And you'll see in a second what it's going to give us back is the task list that it has decided on to actually execute on. So, it's going to go and figure out what it needs to do and it's going to give it back to us here. So, you can see it's telling us it's working. We can go watch that job if we want. Looking into these run jobs, things like that. I'll show you in a second are up here under this actions tab. No need to dive into that right now. We'll just keep this simple and look at it from this standpoint.
Okay. Excellent. So, Claude did come back and updated that last message and it's saying, "Hey, we're working on enhancing this game tile selection." But what it did is something that you, if you're a Cloud Code user, you're used to. It gives you that to-do list that you can see in its thinking pattern that it keeps updating regularly. You can tell where it is. It's doing that here. And in fact, it will come off and check these off. You can kind of just come back to this job here and watch progress happen against this comment. Very, very excellent pattern. I really like this. Some of the others are not quite as ergonomic. Not surprisingly. We find that a lot with Claude.
So, while this is cooking, cuz it's going to take a minute, let's find another way to do this that really for you maybe I'm going to do air quotes free. Okay, Chat GPT. You're a Chat GPT subscriber, right? You go to OpenAI for 20 bucks a month. Is that something you're already doing? If you are, you'll want to see this. If not, maybe this is the reason that you move over to using it. So, let's take a look. You can't do anything here with connectors. It feels like you should be able to connect here. If you go into this, you'll see there's a GitHub connection. That feels like something that I should be able to start doing here and working with the connectors. The way that they've done this here is if you open your sidebar, you'll see something called codec. And to them, codeex is a separate product very similar to Sora. You can see Sora is also broken out over here. So, if we go to Codeex, you'll see something like this. And at this point, you're in a place very much like the codec cli, which is the command line system. They don't have GitHub integration like the other two systems do. They're doing it this way. So here you get to select what repo you want to deal with. You can just click manage environments. That'll take you over to GitHub. It'll say do you want to trust this product blah blah blah. You come back, you can select the repo that you want to deal with. And you notice that we have branches here. This is the conversation we were having before that you can select different branches to do your work against if you like. But here is where I have to do the same thing that I did in our issue. So let me set that up.
Okay. I took the exact same PRD, dropped it in here. I gave it the screenshot just like we did before and we're going to ask it to code. So now it's going to take all of this information and go off and do its own coding effort with codec cli and GPT5 instead of cloud code.
All right. So what is all of this about? I will say I'll share I get a cute little notification which I do like from codeex that you can go into and select that notification. It will take you into the job in codeex that you can watch here on your mobile device or we can watch if I give it my face and what we're seeing on the screen you can watch mobile. This is really the clue. So this is kind of an interesting aspect doing this asynchronously. We saw that we can scale right I'll be able to set up multiples of these. Now admittedly right now we're doing two of them. We sent one to claude code. We've also sent one to codeex. with any luck, Gemini CLI would also be doing its work, but it doesn't quite work for us right now. Uh, so I'm doing three versions of the same thing. Not what I would need, but it would be remedial for you to imagine. Oh, if I put three different issues out there and said, "Hey Claude, do this. Hey Claude, do this." Then I have three different issues. That's our scaling potential. That's really a valuable potential here.
But the other potential is going to sleep. I am not going to interact with this stuff until it's complete. I did put in the PRD. Make sure that you write new tests, update any previous tests, and verify all tests pass before you come back and claim complete. That is kind of a a needed pattern here. A lot of these prompts already expect that. So if your system does have tests, it will be a much better output than if it does not. And this is a very important clue to the way that we're working with agents these days. Start expecting to put tests in. You don't have to do them yourself, especially if you are in a green field project that you're just starting with AI or been doing a lot of work with AI. Start laying down as many tests as you can that kind of create a ground truth because when you start wanting to do this and asking AI to do work for you, it needs something to compare against since you're trying to unplug. Normally, it would use you. You're trying to get away. This is the way to get away is if I have tests in place that can kind of prove the system works it's supposed to work the way it's supposed if I have tests that can prove that the system is still working the way it was supposed to work. I think I got it that time. Uh, then I can make I can step away and make sure that the agent tests those before it completes and it'll go through a loop cycle to solve that problem. So that's really the answer is you need tests in this case.
The other second clue here that I'll give you is as we talked about just moments ago, this costs not in all cases. Gemini CLI should be relatively free essentially today. Maybe it won't be in the near future, but Google has been so good at giving provisions of of use of this stuff to people that are not subscribed. It doesn't give you all use forever, but it does give you a fair enough use that you'd be able to kick the tires and see it. That is essentially free. Doesn't work. Essentially free. Codeex CLI. If you're a $20 subscriber, it's free. You know, my perspective of free is a little bit different apparently, but it's essentially included in your subscription. Claude Code, weirdly, does not also include it in your subscription. You can use Claude Code with part of your subscription. now, but when you go and do these disconnected agents, these bots like this, it doesn't use your subscription. I see that as kind of a mistake. I think they need to tie that in as well because I believe this is a major pattern in the future for how we write software. It won't be the only pattern. We will definitely still be involved. And in fact, we'll see in a moment, there is a fair step at the end where we are definitely involved. So, humans are still going to be needed in this loop, but not at every step. And that's where we really want to start unlocking ourselves. Figure out what you can do asynchronously. Send them off into the trees. Let them become birds. I don't think that's how ideas work. You get my point. Let's go see what these agents have done.
All right, back down in the repo. Let's take a look at what Claude did. This is Claude's job. It of course came back and started checking off the things, and it gives us all of the implementation details and the key features that it gave us back. really very cool write up. This is a a neat traceability. So if you were just watching this feature on the bus, on the train, in bed, wherever you might be looking at this, this is just GitHub. You can just pull it up on your phone. That really is one of these unlocks. Um, you can also see that it created a branch, those branches that we were talking about.
Okay, I want to take a look at the action that ran this. So to show you how I would get to that, I go back up to the actions. I take a look at the claude code action that succeeded here. And this gets us to the action that concluded. And over here you'll see the whole history of everything that occurred. Very cool that you kind of get to come in and look at all of the things that happened. And at the very bottom is what I'm trying to show you. And it is pretty long. All the way down at the bottom, we get to see the cost. So I was build a dollar for this change. And it took about five minutes to do. So it's worth seeing. They are showing us the cost, which is nice, but it does cost. is not part of the subscription, which is a little bit weird. In any case, that's all here for posterity, which is very, very cool to me that I get kind of a physical record of what occurred.
Okay, here we are back in the application. I've updated to Claude's new branch. And if I click on the tile, you'll see our menu at the top. And if we select the item, we get the black border, we get a blue border, we get a white border, and we get a yellow border. And does it turn back off? It does. So, this passes, right? This was awesome. The agent went away, did its little work. I will say mileage may vary. Make sure your purity is very explicit, exactly what you want, and you'll be much happier with the results after paying for it, especially.
But one last check, what about Codeex? How did it do? So, Codeex task is here. This is the task that it built. We could have watched it building, so it's kind of exciting to watch it build. It's got some neat little animations in here showing you what's going on as well. Here are all the files that it changed. It tells us kind of the information about how it tested it, those sorts of things in here. I am going to create a PR and this PR will push up to GitHub and let's take a look at what that shows us. All right, this opens the PR. Remember, we had this pull request concept here. This is also in a new branch and the changes are inside of here. So, you can look at all the different files that have changed. Very similar to what we were just looking at. But what we really want to know is if I change to this, what happens to our application?
All right, here we are back in it. I've pulled down that codeex branch and if we select this. Excellent. Those all look pretty good. We have a black, we have a blue, we have a yellow, and we have a none. And I'm sure I could go in and add more. I was not very clear on how many things to put in here. Once again, be explicit. You'll get a better result. But really, frankly, this is like a full win to me. This is more complex than it seems. It's moving more things behind the scenes than you think. This was not just put some circles up and allow you to click on them. There's a whole database that existed before that it needed to migrate the previous values from kind of a a a token system that we were using to keep track of which level of excitement somebody had into this kind of stuff. So, it really did a lot more than I expected in one shot successfully, but I think that's because we had solid tests and things like that going for us. Okay, this was a great experience.
Okay, so to me that's fantastic. I hope you've heard the one thing that is the takeaway for me here is if you can hear that moving to using agents in places where it makes sense can be the thing that starts to scale you both from I can do a lot more work at one time asynchronously but also I can do work when I'm not at my desk or I don't have to tie up my desk while I'm doing this work. it doesn't have to steal kind of my mental cycles while I'm watching something update and I'm working with it. So, there's a lot of value to being able to push something to the background and say, "Sure, here's a change. I'm going to describe it. I'm going to put it out there."
However, the one thing that I was showing you there at the end, we got lucky that we got to pull up the code and it kind of worked. Now, we need to go back, move that PR up into that main branch. So, we will need to do a merge up into the main branch. Those kinds of things still need to happen. All of this is manual. I had to open the the system, start a server, pull down the new branch, run everything, check it, make sure that it worked. Then I need to go back to GitHub and approve it. So there are steps for you still to take here. If I had kicked off 20 different requests doing 20 different things, I might be able to get them done in a much shorter period of time, but there is still an effort for me on the other end. I think that story will get better and better as we start moving forward and people start seeing this pattern. But really, just getting 20 things done in a very short period of time is already a big enhancement. So, I'm not going to be mad about that.
All right. Well, I hope you saw something in this that was interesting or at least I showed that it was a lot easier than it might seem to start using these agents and start using AI for our benefit to take work off our plate and be successful. It really can happen for you out there already. And it's just going to get better from here, right? Thanks for coming along for the ride on this one and I'll see you in the next one.