Show me the example of that “bad coding” and I’ll show how to mitigate it. If the rule cannot be formalized, it’s not a rule, it’s a personal preference.
Depends on the prompt (20%) and the harness (80%.) If one has a pile of the hand-written code, there is no issue to feed the LLM with it and make it write in that manner. Works for writing, too. It obviously requires the review and nitpicking corrections, but my own code/writing requires it as well.
The harness is the last thing I would allow the LLM to change on the fly. What I do, I rely on static analysis more than on anything else, I instruct the model to not use silly text tools grep/awk/sed and rather work with AST whenever possible (I aim 100%,) I validate each code change in many different ways, and I call LLM if and only if absolutely necessary. Until then I do local. Basically my harness is a thin layer asking LLM to perform small tasks without choices, validatable by the static analysis in local. Although I have a plugin system with hot reload of plugins. It just turns out it’s of no use.
Good for you. I am talking about features, needed by users who vote above, not about you explicitly.
Thanks again for the detailed and thoughtful comments.
(was a CSS issue)