cuddlyogre
· Level 11
random
A side effect of using an LLM to help you build is it's easy to forget to commit your changes as you go, leaving you with hours/days of changes with a commit message like "work". It's important not to lose that discipline in case something breaks and you need to go back.
Comments
You can automate that in your claude md and I have that for several projects. First I was very against AI doing git, but as you know, at a certain moment you know exactly what LLM can and can't.
But good for mentioning, I will automate this for devplace as well.
Part of my problem is I didn't expect to use the LLM as much as I did. Small tweaks turned into giant features the more I trusted it and I just forgot to commit between changes. I'm getting back in the habit, but now the project's history is forever a mess.
Yeah, happens much with LLM development. It requires full-time management and taking maintenance into account when doing stuff. This is my first quite big project that is developed for so long using an LLM. Since it's public-facing and open source, I'm extra motivated to do well.
But indeed, it starts with a little bit fix here and there, then you gain trust and it goes: prrrrr, failure, prrrrr, prrrr, prrrrr - success, hopefully.
But when an LLM did implement something bad, it has a hard time to recover from it because it always autocompletes. I had something weird in the chat function of this site, and whatever I did, it kept giving bad result - while a chat is not so hard to make for that thing.