I Don't Prompt AI Anymore. I Do Looping.

0:00In this video, I'm going to be testing out the looping feature that is supposedly going to replace prompting and I do believe this is going to be the future of how you build with AI. So, in this video we're going to cover it all. I'm going to build out a landing page by implementing this looping feature and just see how it goes. We're going to be using a GPT images 2 for the design mock-up and then we're going to be using Claude Code Obelus 4.8 to actually build it out. And I want to make note of this post by Peter Steinberger the other day. He says, "Here's your monthly reminder that you shouldn't be prompt in you shouldn't be prompt coding agents anymore. You should be designing loops that prompt your agents." So, we've always talked about human in the loop, right? Well, essentially this replaces the human in the loop so to speak as

0:45you're going to basically have an agent that checks in and iterates until the job is complete. Also, Boris Cherny, Claude Code creator, says, "I don't prompt Claude anymore. My job is to write loops." So, I also want to get into this post of what actual loop is because I think it's going to be really helpful to understand what the heck a loop actually is so you can get a good idea of like what we're actually going to be doing in this video. And just to go into it briefly so understand what it is is loop engineering is a designing the system that prompts your agents. It finds the work, it hands it out, it checks the results, and then it records what's done. And then it decides on the next thing. So, design once instead of prompting per turn. And like I said, we're going to be using GPT images 2 to

1:31come up with the hero section mock-up of our next app that we're going to be launching and then it's going to be looping until it iterates and gets it as perfect as possible. All right, so I'm going to ask my Claude Code that's in my project folder to give me a prompt to give to an image generation model to be able to at least come up with this mock-up. So, for what you know about our project, give me a prompt that I can hand off to an image model to generate a mock-up of a hero section for our landing page. I want it to be unique. I want it to be modern. I want it to have lots of motion. I want it to actually resemble our actual project to the T. Make no mistakes. I need the prompt to

2:16be very detailed. I want the hero section to completely represent what we are building here with Build OS and ADE. All right, just came back with my initial prompt and I'm just going to literally hand this off to GPT Images 2.0 cuz it's the best for at least coming up with mock-ups. So, it scanned the entire folder and understood what the project is actually about. So, here is the copy-paste ready to go prompt and I'm just basically going to copy that and then paste it right into GPT Images 2 so we can get some nice-looking designs. Now, I just want to help you understand the shift while the image is being generated right now of where we were just a few years ago of basically prompting ChatGPT to be able to go back

3:01and forth to be able to write code for us. And now we're in this evolution in this period of where it's becoming more of an agentic workflow. You're seeing more of a loop type service where basically instead of prompting it every time, you're getting these feedback loops of maybe an orchestration agent and other agents that come back and report the work to the orchestrator to check and verify the work. So, that's what you see here. You've heard human in the loop for a while and this is going to be the way of the future of prompting kind of going away and having to actually go back and forth with the AI or LLM of your choice to now actually having these loops in place to be able to create, you know, beautiful designs in agentic workflows that basically work

3:48autonomously while we work on higher leverage activities. So, I've been building in this space for the last year. I built Easy Flip. I built the Magic Hand. And I had wish I had known about this long before, but you see it blowing up on X right now. You see Peter Steinberger talking about it. Boris, head of Claude Code and creator of Claude Code, talking about it. And we need to understand that this feature, this dynamic, if we're going to be building, you know, into the future. I do believe this is going to be the way builders build. So, we need to understand the concepts of what actually loop engineering is, so we can apply them to our projects and also buy back a lot of our time. So, it just came out with a beautiful photo. I'm going to show it off here, and then we're going to create that looping system to see how good Opus 4.8 is at design. I think for

4:35the first go around, it did a pretty decent job. Just for the sake of this video, I'm going to just go ahead and go with this, and then see how close Opus can get to at least one-shotting it or getting it close on a few iterations. So, not like my preferred, you know, design choice, but for the like I said, for the sake of this video, let's just get it going to see how close it can actually get. Okay, so I downloaded the photo. I'm going ahead and just drag and drop it here. Now, create me a landing page that is exactly the same as the photo that is attached to this prompt. This is going to be for the BuildOS landing page. We're specifically we're working on the hero section. I'm also going to give it another prompt where

5:20basically says, "Run the test suite." I'll show it here in a second. "Run the test test suite. If anything fails, read the error, fix the root cause, and then run the tests again. Keep looping until every test passes, then stop. Do not check in with me between the attempts. This is important to be able to at least have this feedback loop have running in the background." So, this is what's going to initially start it off, and we'll see how it goes. Also, just a quick little trick for people that are new to using Claude Code. If you hold a shift tab, you could cycle between modes. There's auto mode, plan mode, accept edits. I want to talk specifically about auto mode because people keep getting annoyed that they have to keep going back, prompting, going back, prompting, accepting yes,

6:05yes, yes. Triggering auto mode is going to use more tokens without you realizing it, but it does help you work on multiple terminals at once if you're building out multiple features. Just thought I'd point that out. You want to hold shift tab. Super helpful. I really think that this is going to be a step in the future. You've seen the huge announcement inside of Codex, the forward slash goal feature that kind of blew up about a month ago, and this is no different. It's the same concept. You basically define a goal, and you let these AI agents perform the tasks in this type of recursive loop until the task is complete. This is the the next step into this agentic development environment that we've all been waiting

6:50for. We are all not no professional prompt engineers here. AI is going to be able to prompt itself much better than you can, but if you're able to at least start off with this foundation of defining a clear goal for your idea or your product, this is going to really help you build up more products in a shorter amount of time. I see it as like prompting is kind of going away. You know, prompting as far as like prompt engineering, learning how to prompt specific prompts, prompt experience, learning how to navigate and operate with AI as your sidekick or your builder or your business advisor. This is going to be able to help you now visualize a goal, an end goal, and have these recursive loops happening 24/7. That's why I'm fired up about testing it and

7:35excited to just show you how close it can get, especially with Opus 48. Could you you how it would be with using Fable. I know Fable got taken away. You know, it was a gift to mankind for a few days there and then taken away, which is unfortunate. But, let's see how Opus 4.8 do. Obviously, you could do the same thing inside of Codex, you know, get Codex CLI installed and have GPT Images 2.0 is already built in. You don't need to do go to separate like Chat GPT to generate a dashboard. And you can do it in there, you know, on GPT 5.5 extra high and then pull it into Cloud Code if you like. But, yeah, I'm excited because I really do think that this is going to be a glimpse into the future of how builders build. Okay, so you could see here it's running tests and fixing until

8:21green. It's basically running the loop right before our eyes. I would say it's been running for at least 5 minutes. And it basically just went into my project folder and obviously I got a prompt on, you know, what the project was about. So, it understood the context. We went into GPT Images 2 to please create a mock-up for the hero section. And now it's just ran this loop um for about 5 minutes and it corrected errors, test 10 on 10, 10 out of 10 pass, type check pass, linting pass, build pass. What else? Uh exit codes didn't render. So, basically it's going to just create that loop until it gets it 100% correct to whatever your goal is. It doesn't even I'm just doing a landing page just because I needed to get a landing page

9:06built. But, I wanted to also test and just show you for some of the most simplest goals, have it run it on a loop. For something that is more complex, that's going to take several hours, have it run on a loop. There's no reason why you shouldn't be setting this up. Like I said, either in Codex or Cloud Code. Now is the time to take advantage of this. And I'm going to be using it more and more just to be able to kind of iterate, get some feedback and not have to keep going back, back and forth, prompting, figuring out what prompt to do, and just check and see what's going on. So, it just came back with a few notes. That's a faithful match to your mock-up. Same composition, content, the eyebrow, two-line headline, drive it by voice in accent blue, and the call to action. So, the test is coming up. It It still says run test,

9:52type check, lint loop until green. So, it's still doing that feedback loop going back and forth. Like I said, this hasn't been more than just a few minutes. So, we're at 12 minutes right now, and it says it's done. Everything is built, verified in green. What I built. Says the build OS landing page hero matches that to the mock-up living in a second by entry in this repo, and it is separate from the Atari desktop app. So, it renders the actual product to the T. The eyebrow, run the swarm of coding agents. We're going to test this out and just check. So, let's see. Um two small notes we're servicing. I added five dev dependencies and to satisfy to satisfy the run test. So, yes, let's see it. Can you fire it up on local host so I can check it out for

10:37before we push this repo? All right, so we're just going to fire it up locally just to see what it looks like. Whoa, yes. Look at that. Beautiful. That is with no skills installed, no front-end design skilled, and it did this feedback loop. Oh my goodness. It has that gravity effect on the buttons. Guys, this is the future of building whether you like it or not. Maybe you've spent already a couple months learning how to prompt engineer. I really feel like this feedback loop is so powerful that building has never been easier. Look, it highlighted all of the buttons. You can click into them. The buttons work. You have a little bit of motion. Obviously, this is just a version 1.0. You're going to try to make this much better, but I

11:23just for the sake of this video wanted to show you the power of how close it can get to the actual product. Let's go back to look at the photo to see how it actually looked versus comparing it to what it built. All right, so you can see here that it did a really good job as far as like accuracy. Like it did it almost to the T, which I'm just super impressed guys. Like if you're not using this loop engineering framework to be able to build these agentic loops 24/7 with whatever goal you're trying to do, it doesn't matter now. I'm going to be showing this off every single day on the channel. We'll be building out loops. We'll be coming up with goals to build for our ADE environment to be able to actually take advantage of this main

12:08milestone. This is a huge paradigm shift and you're witnessing it right here live. Like I said, Peter Steinberger has endorsed it and Claude code creator Boris Cherney also endorsed it. And this is how the top builders are building right now. And look, I mean, the proof is right there. In 12 minutes, it ran a recursive feedback loop. Without this loop, I would have gotten some generic AI slop fest, which we do not want. With building being easier than ever before, the last thing you need to be slacking on is your design. I feel like design is the wrapper and it makes things go, right? If you lack design, not saying you need to spend hours and hours on the design, but if you can make it more polished than the next competitor, then

12:55you're going to have that competitive advantage. So, putting in that loop framework is going to really help you in the design aspect. So, it looks amazing. It looks beautiful. Comment down below, how are you using agent loops? Are they helpful for you? Are they helping you build faster? And if you need help being in a community with like-minded people, we do have our Shipping School community, where we build out apps just like this. We have about 140 members and we do four live calls a week and we're also in the middle of a 24-week course syllabus where teaching people how to build these loops in Codex, how to set up Hermes agent, open claw, how to build with Claude code. If that sounds like a fun idea, then check it out. I'll put a link down below. We'll see you in the next one. Have a blessed night.