The fact really is better prompting can lead to better outcomes. A major part of it comes down to prompting and telling the LLM how to think.
Follow my guide on larger codebases.
I’ll be following up with a more advanced version in due time that shows how I use yolo mode to allow Claude to pass off context to other LLMs and then retrieve from them with a python script that I give to Claude in Cursor as a tool for it to use.
You end up getting results like above, which shows Claude going through a ~2400 line Python script (that it wrote in entirety mind you) and making edits through all of it accurately.
You can see in the second image how much longer I went on with that same session with no problems (scrollbar on the right side shows the length of this convo)
