Try to minimize amount of text that needs to be generated. Process diffs

[bddap]

Large documents take a long time to process. They cost a lot to refactor. Additionally, longer documents may contribute to "temperature" related mishaps. For long texts, the gpt4 tends to take shortcuts, inserting "..." in places, even when the system prompt explicitly asks it to output the entire text.

Try having the model generate some machine readable diff representing the refactor rather than outputting the entire modified text.