asking for edits re-renders the entire content
I notice when I ask for an edit to the generated content, its re-generating it through chat gpt, or whatever model.
I use Aider a lot as a personal code assistant. And their implementation and approach is to to generate a unified diff (by generating only the changes that are needed in udiff format) and apply that to the original work rather than regenerating the original work with the edits. I cannot overstate how exceptionally well this works. It does require some fine-tuning of the unified diff prompt and a error correction agent to fix when things go wrong. Which is in frequent, but it does happen.
In my opinion the experience would be greatly improved by generating a unified diff and applying that diff against the original source.
Nice, love this idea. Right now I'm basically starting a new chat session each time. If we include a diff it would definitely provide more context. I'll experiment