nourspace on Nostr: How to stay sane while using LLMs? Company and manager are pushing us hard to use it. ...

How to stay sane while using LLMs?

Company and manager are pushing us hard to use it. I'm now reviewing ~5-10 vibe coded pull requests a day, some of which are okay, and the rest are nicely organized word sequences that are redundant and in some case totally wrong.

I created multiple skills for the team to use, so this is helping a bit, but in general the cognitive load is ever increasing. Especially when I myself fall into the comfort trap and ask Claude to do trivial things like renaming a file.

The worst thing is that I'm running a constant mental loop where I'm looking for the perfect prompt to solve some upcoming task—mentally! Not actually, because when I settle on something and go start a session I spend way more time than I imagined and I end up frustrated fighting with the LLM stupidity to get it to do what I really want which I kind of knew exactly so should I have started that on my own without agents I would have done the task faster. The issue is that starting a task from scratch now feels pretty hard compared to initiating a session.

Often times I feel I have to correct the LLM hoping that next time it doesn't make the same mistake. For that I also created a skill that I load on every session and during the session I ask the model to summerize the learned conventions. Truth is it helps, but also it easily ignores all the conventions and says that it did that on purpose.

Code quality across repos is degrading so quickly, quicker than the amount of comments, and commits, being added.
In my opinion if every two lines of code there are 10 lines of comments that explain them, things are rotting and they smell.

Using chat interface is like a disease. Worse is using it in the terminal. It was much nicer when I used to copy, ask, paste. Now I have to wait for Claude to finish Bazooling... and the company spends a $1 after every session I start.