I think that the key to using any LLM for writing code is to not overwhelm the model with too many asks or feeding it too much code at once, even if the context window is large. Baby steps and EXTREMELY clear requirements for each one of those steps are absolutely critical to getting the most out of any of these models.
One hack I've found is to have the LLM (or yourself) comment your code very explicitly. When you feed it back, the improvement is substantial.
I think that the key to using any LLM for writing code is to not overwhelm the model with too many asks or feeding it too much code at once, even if the context window is large. Baby steps and EXTREMELY clear requirements for each one of those steps are absolutely critical to getting the most out of any of these models.
One hack I've found is to have the LLM (or yourself) comment your code very explicitly. When you feed it back, the improvement is substantial.