Being self taught, there are lots of things I never formally learned, rules I know from the rule of thumb, and not the deeper knowledge... So I set out to learn the root of what can be used to measure good robust code... Spent an hour asking lots of questions, learning about LCOM, Halfstead, why circular dependencies are bad, and so on...

The next morning I figured the same LLM could compute that on my code, so I asked it to make an agent to do so, and report issues to me...

And then I ran that agent with next to no changes on a feature that had grew organisally over the last months, that I knew was messy and sometimes difficult to work on, despite being unable to precisely say why... And it did tell me exactly why, and proposed changes to improve stuff, and then implemented them...

Up until that point, I'd felt like the LLMs always produced bad code, that worked for a specific feature but often broke stuff or evolve poorly over time. Then I realized if you had the LLM do code improvements, it could do that fairly well too...