I was working on a project that involved indexing GitHub repos that used really long prompts. Iterating over each section and figuring out which parts of the prompt led to which parts of the output was a quite painful.

As a frontend dev, I kept thinking it would be nice if I could just 'inspect element' on particular sections of the prompt.

So I built this prompt debugger with visual mapping that shows exactly which parts generate which outputs. Now that context windows are getting huge, working with massive prompts should not be as difficult.

Planning to open source this soon, but I'd love ideas on how to improve it: - Should I consider gradient-based attribution or other techniques to make the mapping more accurate? - Would this make more sense as a CLI? - What else can make this actually useful for your workflow?