I've used Aider <https://aider.chat https://aider.chat> a fair amount, and I have read about Paul's experiments with different "diff producing" methods. I think

I've used Aider <https://aider.chat https://aider.chat> a fair amount, and I have read about Paul's experiments with different "diff producing" methods. I think a good avenue to experiment is two-pass. First, ask the LLM to output only the changed code, then (in a separate conversion) as the LLM just to update the relevant lines with the change the first LLM provided. I suspect this would clear up a lot of the "LLM failed to transcribe old code" and "diff isn't well-formed" errors that this tool has.