I've Changed My Opinion On Vibe Coding

Hey guys. So, I've been pretty vocal about vibe coding over the past year or so, and most of what I've said has been pretty skeptical, watching people ship apps without understanding a single line of what was generated, watching beginners get stuck the moment something broke, and not to mention all the influencers pushing the idea that you don't really have to know anything about software development to to create successful apps and and SAS, which I'll never agree with. However, I have changed my tune a little bit on vibe coding in general. All right. So, before we go any further, let's define what vibe coding actually is because there there's a really a spectrum when it comes to coding with AI. In fact, in my coding with AI course, I talk about five levels ranging from oneshot prompts with with platforms like lovable to just using autocomplete and VS Code. And I think the the sweet spot is right in the middle where you're letting the agent write the code, but you're making the architectural decisions, testing, writing spec files, etc. And that's what I teach in my AI course, which I'll have the link for in the description if you're interested. But to me, that's not vibe coding. Vibe coding is where you're barely looking at the code at all. And as the name implies, you're kind of going off the vibes, not the actual syntax, which is something that I've been totally against in the past. However, over the past couple months or so, I've been going allin with AI. I've created a home lab with eight machines that's managed by my Open Claw Travis. I have agents talking to each other, assigning tasks, and been creating all kinds of projects, mostly things that I can use in my daily workflow. So, I've gotten a bit, I guess, more liberal with just letting AI cook, where before I would monitor and spec out every little feature. The main reason that I've been able to kind of change my opinion on this is because of the models. Opus 4.7 with claude code and GPT 5.5 with codecs and even OpenClaw with GPT 5.5 are all amazing if you understand how to direct these models to to get the results that you want. So I I really don't think that we have to babysit the code as much as we did with for instance GPT 5.3. and I'm seeing much less hallucination and for the most part it does what I want on the first try. Now I know that some of that is because I've learned how to communicate with these models, how to manage my context and memory, how to map out my documentation and spec files. So it doesn't mean that just anyone can can pick up any of these models and build a successful SAS. Which brings me to my main point. I don't think vibe coding is okay under any circumstance if you don't have a foundation in software development and architecture. So for instance, my mother who basically just only knows how to use Facebook will never be able to create a successful SAS no matter how amazing the model gets unless she decides to learn software development with which I know she's not. And that's one thing I I won't budge on. And if I do, I want you guys to hold me to it. And I've seen this play out a hundred times. You know, somebody with no real dev background uses an AI tool to build something. In fact, my accountant just did this and told me about it and had this exact experience. They get version one in an afternoon and they're thrilled. Version two is harder. Version three breaks something in version one and they don't know why. And they ask the model to fix it. And the model makes it worse in a way that they can't see. And now they're three layers deep in fixes that don't address the actual problem. and the codebase has architectural choices that nobody actually made on purpose and they can't even describe the bug clearly enough for the model to help. So this is not a tooling problem, it's a knowledge problem. The user can't form the right mental model of what's happening because they never built the the prerequisite intuition. And no matter how good that model is, it can only meet you where you are. So even if it's smart enough to to fix what's wrong, you don't know where to tell it to look. And even if it knows where to look and finds it on its own, you can't understand the solution so that it doesn't happen again. And that's only part of it, right? There's a lot of technical skill that you need outside of just actual coding syntax. Imagine vibe coding if you didn't understand Git or version control or didn't even know what it was. if you didn't know how database tables work, if you don't know how to set an environment variable, you know, you don't even have the vocabulary to vibe code and the agent is it's going to be asking you stuff and asking you to make decisions that you don't understand. And I haven't even mentioned deployment, uh, maintenance, scaling. That's all stuff that that you need to know. So, with that said, you don't have to be a 20-year veteran software developer to use AI. I think that people get frustrated with us sometimes because we say learn the fundamentals, build a foundation, but what does that actually mean? So foundation at least in terms of a in in web development context, it means understanding how the web works, requests, responses, status codes, um what happens between the browser and the server. It means understanding data, how to model it, how relationships work, what an index is. It means that you've you've written enough code by hand that you can read code and see it and not just skim past it. And it means that you've been burned by enough bugs to to recognize bad patterns. You don't need a um a CS degree. You don't need to memorize sorting algorithms or be a genius, but you need to have built things from scratch, debugged them, broken them, fixed them. There's no shortcut to to that pattern recognition and the model can't give that to you. The model can only really amplify what's already there. So with that said, what should beginners do? I wouldn't say stop using these tools, these AI tools. That ship has has sailed and honestly fighting it is is the wrong move in my opinion. But change what you're using them for. Use them to learn faster, not skip learning. AI is actually a really incredible tool for learning how to code. And I don't feel like that's talked about enough. In fact, me and my team are building on uh working on building a platform that merges AI with coding lessons, but you can have the model explain the code that it generates, type it out yourself, you know, break it on purpose and try to figure out why it broke. And the developers who are are are going to do well over the next few years are not the ones who can prompt the best. They're the ones who can prompt well and read code and design systems and debug under pressure. The AI handles the typing. You still have to handle the thinking. And this is mainly towards beginners. Like back to vibe coding. Once you get past that learning stage, vibe away. However, don't make it the the only way that you create because you're going to forget a lot and depend on it too much. So, I'm not against vibe coding anymore. I just think that it should be a tool, not your entire workflow for every project.