AI coding is easy, actually

So I recently came across this article by Peter Steinberger. Peter is a fantastic guy who's using AI very heavily to build all sorts of applications. And while this is a very fascinating read, I kind of just got obsessed with this image over here because I kind of see my past self in this image. I went through this entire little bell curve and arrived at the end over here. And I kind of want to talk about it. And as far as I've seen from my own experience and also talking to other software developers, this is the kind of bell curve that most software engineers go through when they are using AI agents to program. So if you look at 99% of developers, they're still stuck in this "Please fix this" era of software engineering with AI. And to be honest, this is where I myself started as well. Back in the early days of 2023, when these AI tools were coming out, like Microsoft Copilot and others. Back in these days, we didn't really know all that much about prompt engineering, and we kind of just told the AI to do something and it failed because the models were not good enough at the time. And for most people, their perception of AI just hasn't changed since then, right? Even though we think of software developers as being cutting edge and closer to the future, most of them are not like that. For most developers, their perception of AI is still the 2023 Microsoft Copilot that can't really do shit, right? And so for these people, they have no idea how much better current models have become, how much better the tool calling has become, and how much progress has been made since then, right? They don't know how to write good prompts. They haven't taken some time to learn good prompt engineering practices, and so they just get horrible results from the AI, right? And these people, and I'm tired of saying this, but it's just so true, these people are going to get left behind. The technology is here, the genie is out of the bag, and it's not going to go back in. If you are taking six hours to write some code and another person is taking, let's say, 10 minutes because he's using AI, that is a huge time saving over there. And yeah, you can say that there's so much more to software engineering than just writing code, and to that I say this other person, this other AI-using person over here, this guy is going to have so much more time to actually focus on those other things of software engineering, like architecture and system decisions and all sorts of other large things, right? Whereas this person is not going to have that kind of time. So who's really winning over here, right? You don't get any bonus points for taking longer to write code, right? If you can write code quickly and then you can think about it and improve it and, you know, work on other important things on the product, great. But if you're taking more than the necessary amount of time to write code, that's not exactly a flex, and you're only making your job harder by doing that. Anyway, next we have the what I call the optimizers and the agent maxers, and these people have a ton of shit going on over here in their agent workflows all the time. They have multiple agents running at the same time, multiple agents running overnight for hours and hours building weird features. They've got sub-agent workflows, they've got slash commands and all sorts of other things. They're just going really, really deep into the agentic programming world. And I'll be honest, even I haven't gone that deep into this because quite frankly, you don't need to. Honestly, I would say that the original author of this article, Peter Steinberger, is the only one who actually going that deep into agentic programming. I see a ton of people on Tech Twitter that are doing the same thing. They're running Ralph Wiggum loops, they're chaining multiple agents together, all sorts of weird agentic shit, right? The problem with these agent maxer types is that they're orchestrating all of these complicated agent workflows together, but they're not getting shit done. And that is the crucial part over here. Chaining together all of these different agents and having all of these super complicated practices when using agentic programming, it's not doing you any favors. It's not helping you build a better product. It is literally just giving you a false sense of productivity and just another cheap dopamine hit. It is literally all it is. The only reason someone has such complicated workflows is if they want to one-shot a large app for a cheap dopamine hit. There is literally no other reason to do this. Now lastly, we have what I call the enlightened group because they've gone through this entire cycle and they've arrived at this place, and this is kind of where I am as well. I'm going to tell the AI, "Hey, I want to do this thing. I want to build this feature or fix this bug. Here are the files that reference the relevant code. You can analyze them, and here is the other context and instructions that you need. Make sure to follow them, and then just go do it," right? I just let the AI do its thing after that. So people in this stage, they've already been through the Ralph Wiggum loops, the custom sub-agent workflows, the agentic tooling and orchestration, all of that nonsense, right? And they understand that the AI models of today are so good that you can literally just explain what you want in plain English with clear instructions, and you can get amazing output just from that. You don't need all of these different complicated workflows anymore because you can literally just explain it in clear, plain, simple English and get very good results just from that. And that's my advice to you as well. If you haven't gone through this entire bell curve of agentic programming, let me give you some advice. You can skip the entire thing and you can arrive straight over here because this is the best place to be, all right? You don't need to overcomplicate your entire setup with these sub-agent workflows and Ralph Wiggum loops and all sorts of other bullshit, right? Just explain to the model in plain English what you want to build. Give it the right context. Give it the, tell it where to look. Tell it which files to analyze. Give it specific instructions, and then just let the model do its thing. And you're going to find 99 times out of 100, this is going to be so much better than this. So go ahead and pick any of the frontier models right now. At the time of recording, that's Opus 4.5 from Claude, or you can use GPT-5.2 Codex or Gemini 3. Gemini 3 used to be a very good model, but right now, these days, it's just hallucinating so much for me that I'm just defaulting to Opus 4.5 for everything I do. So pick one of the frontier models, whatever the latest model is at the time of you watching this video. Explain the feature you want to build in plain, simple English. Give it the right context and the specific instructions, and then just let the model do its thing. And leave all of this complex sub-agent workflow Ralph Wiggum shit to the unemployed Tech Twitter bros who aren't actually shipping anything real. They're just trying to impress other vibe coders with their vibe coding skills. You don't need to be one of them. Let them chase the dopamine of building some weird slop app while you can actually go and build something that real people are going to use.