Nice! Overall, frameworks that re-prompt an LLM with feedback or failure modes of the original outputs do very well.