I'd like more people to be writing about this, because I find it fascinating. I have my own private project that is heavily LLM-coded, just to learn about what it's like. It's amazing how everything is different yet everything is the same. Big complicated features can start working quickly and give a massive endorphin boost, but then fixing them up and integrating them in properly and polishing the UI? It all feels even more painful having experienced the heady thrill of the initial implementation. And things get to a point where you can feel the inertia set in, the point where things have gotten so hacked up that the LLM can't make any progress without creating an even bigger mess. It's the point where you have to go back and fix up the architecture, or scrap the whole thing and restart with a better plan, or a bit of both (rewind to the "last sane point"). It's like developing with a jetpack -- you can go way faster towards your goal, and you can slam into walls way faster and more painfully too.

I think there are tons of learnings to be shared about how to do this stuff, but it seems like it's all blocked behind arguments over whether AI is the best or worst thing ever, and penis-measuring contents about how to hold the tool. The net benefit is a very open question, and both the doom and gloom perspective and the AI booster perspective are valuable and have a lot of things right. But there's a dearth of information about what things work, what things don't, what happens in the process of using AI, how to adjust one's behavior and which of those adjustments is harmful even if effective.

But then, I'm part of the problem. I keep meaning to write up a series of experience reports, but it's a lot of work. More fun to vibe a new feature into existence, or to finally fix a UI wart...