After running scenarios, have an agent digest the original prompt and scenario feed back to suggest edits/additions/improvements to the prompt as to fix the failed results. This would allow for a self-building mechanism with simple user input.