Vibe Coding is Broken — Here's the Fix

James Pelton here. This is obviously not my normal setup. I am at the HumanX AI Conference in San Francisco. Vibe coding, what we're gonna talk about. I'm getting tons of DMs. I woke up this morning with four emails of people saying, James, can you fix my app? In some ways, AI coding is amazing. It's like a miracle. All these people who have no coding experience all of a sudden can build their own apps, their own software, and it's great. But the problem is that AI coded software is very flaky. It does not work great. It works, you know, these emails I got this morning, everything was fine until I added this one feature. Now AI can't figure out what's going on. It broke everything else. And I've run into that too. You know, I've built a lot of apps with AI and they work pretty well, but they're full of bugs. It's just, it's not very reliable. It's not as reliable as a human going in and typing all the code, although it's a lot faster. But this is a problem in the industry. Only 37% of developers, a GitLab survey said, trust the code that AI writes. So enterprise customers, you know, enterprise software, software that needs to work is not moving to AI because it's not reliable. It's broken. So let me know in the comments if you've experienced this with your own vibe coding. But the three problems that I see with vibe coding, first off, there's no specs. Nobody wrote down what the app should actually do. If I get into Cloud Code and I say, build me a task manager app, it will get going on it, but there's no spec. I mean, that's very vague and Cloud will make some guesses and it'll say, I think it probably wants to do this, probably wants to do this, and it'll just go on it and it'll go on it at, you know, a million miles an hour. But this needs more detail if what's in my head is gonna actually be translated to what is actually produced. But problem two is there is no testing. So in traditional development, you have unit testing, integration testing, end-to-end testing. And if you go and change something, you kind of know if something's broken somewhere else. AI doesn't really do that. So you might have, you know, good working software, then you add one feature, AI goes back, changes a whole bunch of stuff, and now your whole thing is broken. Or maybe the new feature works, but old things stopped working. There's not testing. Sometimes you don't find out until your users tell you, which is bad. Third problem is we have AI agents stepping on each other's toes. So if you don't have good specs, you don't know what you're building, one agent you start with and it builds what it thinks it should build, then you shut that down, you start a new agent. That's like a new person. And it comes with its own idea. It looks at the code and says, well, actually, I think it should do this. And it changes. And it's just, it's a mess. There's no coordination. And I'm not saying that vibe coding is evil or bad. I think it works great for prototypes, for exploring ideas, for MVPs getting stuff out quick. But for stuff you actually want to put out into production, vibe coding is not going to work. For example, if I say to Cloud Code, build me a task manager app, is this supposed to be a web app, mobile app? What's it going to do? It's going to just kind of guess. And should it have a user login? Should there be drag and drop? Should it be just check boxes? I mean, it's going to make all these guesses. And at the end, it's going to go very quickly. But at the end, it might not be anything close to what I had envisioned. So I found something new. It's called ZenFlow. And this is a code editor, but it's made for AI coding. So you don't see your actual code here. That's not the future. The future of coding is not going to be people sitting down writing words, writing code words in Python or something. The future of coding is going to be talking and working with AI to build the right thing. So what ZenFlow does is it starts with a very intense spec. So you actually, it asks you questions and forces you to slow down and actually write out into a file what it is that you're trying to build. And this is super important because then if another agent pops up, it can go look at, oh, this is what we're trying to build. And there's a good plan there in place. Then at the other end, it does testing. So when you're done, you know the tests match the specs and this works the way that it's supposed to work. And here's how it's supposed to work. We as developers used to spend about 5% of our time planning and then 95% of our time writing code to build it. And we're flipping that on its head. The coding part's easy now. AI can take care of the coding, but we're gonna spend 95% of our time planning and making sure we're building the right thing and that it's documented well. So this is ZenFlow. I'm gonna try this task manager app. So I wanna build task manager. And I'm gonna give it just the same build a task manager to-do list thing. So again, very vague. And you're still using, you know, you can plug in Cloud Code, Codex, Gemini. You're still using the same AI coding. You're just putting a layer on top of it that specs out and builds the thing that you wanna actually build. So I'm gonna use Cloud Code, Cloud Code open. I'm gonna do SONNET 4.6, save on some of those tokens. So you can see immediately it broke it down into requirements, technical specification, planning, and then finally implementation. So normally your vibe coders would just get started on implementation. But this automatically adds these three steps, requirements, technical specification, and planning to make sure we're building what we're looking to build. So again, we're shooting with a machine gun. With AI coding, we wanna aim first or else we're gonna hit something that we don't intend to. So, and I love this, it asks first for questions. What platform should this be built for? So we wanna be a web app, mobile app. That's an important question. Claude would have just guessed. Let's make it a web app, a standalone app. What features do we want? Let's make it just personal. Is this something that we're gonna be sharing with other people? Let's make it just personal. And then it takes my answers to the questions and it puts together a document. I also can review with another agent. When you spin up a new agent, it's kind of like a new person. So you can have a second agent go look back at what you did in this step that the first agent did and see if there's anything you need to look at, any changes that you need to make. It's just a button right here, super cool. So we're gonna call requirements good. Let's go on to technical specification. So this is like the actual tech that you wanna use to build this. So if you're not a developer, you might not care as much. You might just use whatever suggestions they have, but you should have, this should be documented. So future agents know, hey, we're using this database. We're gonna write this in React. We're gonna use this JavaScript. We're gonna use this server. So ZenFlow goes through and writes the tech specs for this. All right, so it made some decisions. It's gonna use Next.js, TypeScript, Tailwind CSS. If you cared, you could go in and change that and say, no, I wanna use this. And say, I wanna use PHP to build this or something like that. But finished tech specs, let's go on to planning. So again, haven't written any code yet. We're in the aim, aim, aim, three series of aiming before we're gonna actually go start to implement things. So then after planning, it did the project setup. It actually created the files based on the specs that we put together. It started to put together the dependencies. And then now we can actually just get going on these different tasks. And we can start them in parallel. Since we have the plan all there, we can spin up different agents that are gonna go just work on different parts of this all at once. And you can see as it's going, it's even creating the tests to make sure that it matches with the spec that we put at the beginning. And you'd see the planning step put in a bunch of other to-dos that we need to do. And we're just gonna have the kind of these six implementation steps. We're just gonna have agents working on them all at once to save time. And while those agents are going, question I get a lot is, I'm starting a lot of side hustles, a lot of businesses. When do you start an LLC? When do you start putting these businesses in an LLC? And how I generally handle it is, before it makes revenue, it's just a hobby. I don't need an LLC if there's no money changing hands. Once I start generating revenue, then I wanna quickly get that into an LLC. And so I found Doola is an easy way. Some people go find like a local lawyer or something like that. It's way too pricey. I prefer doing things online. But I found Doola, where you can start an LLC very, very quickly. You can build your business and let Doola kind of handle the legal side of things. But it'll walk you through. Do you need an LLC? Do you need a C-corp? If you do an LLC, do you want it to be an S-corp? Or it'll explain kind of in plain English. It'll go do a search to make sure your company name is available. And then it'll take care of all the paperwork for you. So it'll get your EIN. It'll set you up with a bank account. Then as you grow, it can even handle your taxes. It can handle a lot of your accounting. You can even connect it with like Gusto and let it handle your payroll. So it's a great way to kind of start a business. But at the beginning, you might just need the LLC. And then as you grow, you can add on these other features. So very affordable. I have promo code JamesPelton10 for 10% off. But again, then you don't have to spend your time worrying about the legal structure. You can just go build and build your business. So all these features are done. All the tests that went along with them pass. If I want to, I can have another agent go review it. But now that everything is done, I'm gonna kind of do this last step to just a final verification to make sure it's working the way we want it to work. So after it went through all those tasks, it is up and running. So Localhost 3000, here it is, my task manager. But I can keep talking to it. I can make a new task to add a new feature or, oh, hey, I have a bug task. And for each task, it's gonna kind of do the same thing. Spec it first, make sure, hey, this is what we're gonna be doing. Make sure the tests still pass at the end. But I think this is the future of AI coding. It's not gonna just be hopping in and start putting prompts to AI. We need a layer on top of that to make sure the code going out is good. And I think ZenCoder and ZenFlow is the answer to that. Like, I believe it so much, I joined their team. I'm gonna be helping them with marketing. I've been a dev for 20 plus years, and I think this is the answer to making sure we get good code out there. So I'll leave a link to ZenCoder. You can go download it, try it yourself. I'll leave a link to Doola, go check that out if you need the legal structure. Let me know in the comments, what's your favorite, what have you vibe coded that you're proud of? I have a friend, shout out to Aaron. He built a wordle with lifelines. And there are all sorts of cool stuff that you can build, but let me know in the comments, would love to check it out. Appreciate you guys very much. Hello from San Francisco. I will be back home soon. God bless you all, and have a great rest of your day.