The bit about strict guardrails helping LLMs write better code matches what we have been seeing. We ran the same task in loose vs strict lint configurations and the output quality difference was noticeable.

What was surprising is that it wasn't just about catching errors after generation. The model seemed to anticipate the constraints and generated cleaner code from the start. My working theory is that strict, typed configs give the model a cleaner context to reason from, almost like telling it what good code looks like before it starts.

The piece I still haven't solved: even with perfect guardrails per file, models frequently lose track of cross-file invariants. You can have every individual component lint-clean and still end up with a codebase that silently breaks when components interact. That seems like the next layer of the problem.