Well when you’re coding on your own you can get into the zone and just “flow”. With an LLM you’re waiting for the result, you see it has changed things it shouldn’t have changed and while the over all result is a step in the right direction, you have to go back and fix a lot of the LLMs “corrections” which is super tedious.

I asked Claude to help me out with an issue I was having with a small renderer I was working on, it fixed the issue but also added a memory leak, the leak was easy enough to fix because I fully understood what was going on but if you’re vibe coding and don’t have the skills to debug yourself you’re going to have a bad time.

In the end I don’t like using LLMs for coding but they are good at solving isolated problems when I get stuck. I prefer it when they review my code instead of writing it for me.

I’ve tried the paid models through GitHub copilot and I just can’t find any of them actually useful for anything more than generating tests.

They can generate stuff, but generally I spend so long fixing it manually that it makes the time savings zero or negative.

Only thing I have found useful is the PR review bot. That thing is genuinely incredible at spotting tiny mistakes in massive PRs that make your eyes glaze over.