I appreciate seeing this, but I would like to see more people talk about WHY handwriting code is still valuable. For me, it's not because I'm a purist, but when I tried drinking the coding agent kool-aid, I saw my productivity get worse rather than better. I've thought about sharing more on this, but I've seen a worrying trend where anyone who talks about how they're not seeing amazing productivity gains from AI gets told it's a skill issue and they're doing it wrong.

Anyway, I’ll briefly share why after trying full agentic coding in earnest, I’m back to writing almost all my code by hand (I even turned off AI edit predictions). I will not deny that a lot of AI code is quite good these days, but in my experience it’s nowhere near the point where I’m comfortable not reviewing it. So if I have to review it, then this means I need to read verbose more code than most human developers write and that is just not fun. But beyond just not being fun, as I got farther from writing the code myself, it felt like I was a worse reviewer. Obviously total skill atrophy doesn’t happen overnight, but I felt like even being away from coding for a few weeks or months, I was more and more “out of shape” and whenever I did dig into the code base, even if I had reviewed everything and guided the AI, it all felt really alien. And then there is the part about needing to read tons and tons of annoying AI generated prose and argue with it constantly. Ultimately after a few months of trying this on some personal projects, I realized that because I had been so disinvolved with writing the code, most of it needed to be reworked in pretty substantial ways. Once I started writing the code again, I realized my net ability to steer the code in the right direction was faster when I typed it out than by going through this tiring LLM prompt/review/re-prompt loop while my skills were subtly atrophying each day. Now that I’m back to writing almost all my code by hand, I sort of hold the opinion that hand-coding is an unreasonably effective away of developing software architecture intuition and steering a codebase in the right direction. Also, I realized typing code is not actually that slow - thinking is definitely the slower part and the more you offload your thinking and understanding to robot who has literally zero persistent memory or stake in your project, the less capable you’ll be of doing the thinking when you need it.

I still use AI for brainstorming, reviewing my code and one-off prototyping/benchmarking tasks, but usually even if it writes some decent code snippet in the chat for me, I’ll go and handwrite it myself because then I actually get to learn something! This also let’s me confirm that the AI suggested approach is correct (usually it’s pretty good, but sometimes it’s off). I know some people are saying you have to stop reviewing code if you want to be fast with AI - that’s true and maybe that’s what the future will look like. But if humans still need to be involved in understanding and guiding code, I think handwriting code is quite a good way to stay on top of things.