Reflections on vibecoding ticket.el
3 points by jmmv
3 points by jmmv
I've also found myself, when facing small-to-medium problems I would otherwise maaaaaybe write down on a list somewhere and never get around to dealing with, aiming LLMs at them and seeing what I can quickly obtain. Sometimes these are even problems I feel more than one step removed from being qualified to solve. And when stuff starts working, it's kinda cool, and maybe "enough" for a moment... but like you, I don't feel good about it. Specifically, I immediately dread the thought of having to come back later and trying to change anything. (If I stop here, the code is likely to get thrown away or never changed again.)
But that feeling is so familiar that I bet I know what to do about it! And it's true: on the occasions when I take the time to understand, tweak, and trust the tests, I feel way better. Now I could change this code if I had to, and an agent is way more likely to be able to do it for me to my satisfaction. (This can be a relatively quick followup step, so I'm relatively likely to do it.)
On the occasions when I go one step further and take the time to refactor the code to my liking, I feel even better, and this feeling is familiar too. If it works right, makes sense to me, and seems likely safe and cheap to change, then the feeling I'm having is -- I think, anyway -- the exact same "art"-ful feeling as if I'd wrestled with generating all the code myself to begin with. (This can be a more involved step, and also feels like something I can afford to leave for later, so I'm less likely to do it before setting the code down.)
Right. I was trying to explain this from the perspective of "let's work on this with full vibecode-mode enabled" because that's what some people are trying to push as "the future of coding", so I needed to experience what it feels like.
For another project, I've been using agents to build features after I built the foundations and the testing infrastructure. Basically, I'm telling them to fill in the boring gaps: things I know how to do and are just repetitive and not interesting. And in those situations, they have done great! I've actually felt engaged in the process of prompting and then refining the code (be it by hand or via extra prompts or via TDD). And with more experience in how to write the prompts and how to pre-encode rules in an AGENTS.md file, I'm actually getting "one-shot" results that are extremely close to what I would have written and require no follow-ups. I think this lines up more closely to what you are describing, right?