I have noticed that when asking LLMs to work with existing code, their output reflects the existing codebase: the better the existing code, the better the LLM code.