Feb 24, 2026

One good way of thinking about AI codegen is that it removes friction from coding. Friction is what keeps your pants on. Without friction, people don’t need to think hard whether the thing they want to build should be built in the first place. LLM happily takes any task, adds its own implicit assumption and spits out loads of code. Since the process was mostly frictionless until this point, the engineer is faced with a wall of code. The code they never wrote, and are completely unfamiliar with. It takes a great deal of discipline to go through it and internalize completely. What I found though experience is that if I apply the same vigor I do for my own code, most of this slop has to be rewritten. Before I can show these to my colleages, I have to do a lot of grunt work. Usually, polishing the code that took 10 minutes to produce takes at least a day.

The experience is extremely frustrating, and many times I wonder if I should have just write it from scratch myself. But at this point sunken cost fallacy kicks in and I force myself through cleaning up the mess.

I’ve done plenty of cleanups through my career, and it’s never fun. But this is the only way to keep my pants on.