Coding without AI: a revolutionary new way to work • Minimum Viable Blog

Minimum Viable Blog

Software construction, opinion, and self

For the last 20 years or so I’ve been trying out a new way of programming, one that (surprisingly) doesn’t use any LLMs or AI coding agents at all. It’s just me, a keyboard, and an IDE or text editor, writing code line by line—literally typing things like functions and curly braces.

I realize this sounds eccentric. What am I, some kind of minimalist? Actually, I’m what you might call an “early adopter.” I’m trying something different, something radical, and I know not everyone will catch the vision. Most people think coding and AI are the same thing. But I’m here to tell you that AI-less coding is a real thing, it can be done, and it has some surprising benefits.

I also enjoy eating food prepared by real people, outside of a factory, and I almost exclusively listen to music made by people playing instruments. I have a lot of weird, niche hobbies.

Let’s discuss those benefits. You might even decide that AI-less coding is something you want to try for yourself. (Maybe just for a weekend, in case it becomes the wave of the future.)

1. Increased understanding

When I’m coding manually, I can’t simply tell my agent, Fix this bug that hides the Submit button, or Make sure this asynchronous job updates the database. There is no agent! Instead, I have to carefully read, test, and debug the code to build up a mental model of how it works. This is difficult to do, but I’ve gotten pretty good at it. Along the way, I’ve noticed:

My mental model is correct much more often than AI is. Once I’ve examined the code thoroughly, I can make all kinds of useful predictions about its behavior.

When the code has bugs or doesn’t match the spec, I can often visualize, in my mind, the exact file or method that needs to change. It’s hard to describe how much this has sped up my workflow.

As I read code, I can’t help but discover things I didn’t plan on learning, since code is all so interconnected. Those “bonus learnings” frequently come in handy later.

I often notice ways the code could be improved so it’s more reliable or easier to work with.

When people ask questions about the application, I sometimes know the answer off the top of my head, without consulting any chatbots.

Eventually I reach a point where—for lack of a better term—I’m an expert in the codebase and all the technologies it uses. That expertise is available to me all the time, everywhere, for free. No agents required. It’s just right there in my brain.

You might think this sounds slow, but in fact, the longer I work with a codebase, the faster I get at building on it. AI’s still faster at throwing together quick prototypes, but I’ve found myself building those on my own too, because there’s just no substitute for deep, genuine understanding.

I’ve used that understanding to become extremely valuable at work, earning several promotions and accolades over the years. Don’t tell my coworkers I’m coding without AI, though—they might see it as an unfair advantage.

2. Working software

The world has an enormous appetite for software that doesn’t really work, and I’ll admit that with a little effort, I can even hand-write software that doesn’t work. But most of the time my AI-free code works pretty well after a little testing and tweaking. And due to some mysterious effect, it works better over time. A lot of my code can just be left alone and it keeps working for months or years. I’m not kidding. I feel like I’ve discovered a forbidden secret, having code that doesn’t need to be thrown away and regenerated all the time.

When my code has bugs—and let’s be honest, that’s a constant whether you use AI or not—I’ve started feeling much more confident in my ability to fix them. Instead of going the rounds with a chatbot that makes up several incorrect solutions before finally writing enough code to paint over the problem, I can use debugging techniques that gradually narrow down the source of the bug until I know for sure where it is, and then I can fix it with high confidence. Sometimes the fix is a single line of code, if you can believe it.

I think I finally understand people who sew their own clothes or make their own furniture. It’s just a higher quality result. Imagine having a T-shirt that doesn’t fall apart after a few wears! Sure, it’s more effort up front, but you can’t argue with quality.

3. Simpler code

You may not realize this, but AI often complicates simple tasks. Sometimes it writes you three pages of data structures and algorithms when all you needed was an if/else block. Writing code by hand has taught me to see complexity as a liability, not an asset, and to...