Just feed the output through a quality system (with retry if the quality is too bad), scaffold it a bit, then run it back into the LLM. Should work(tm)