In my prompt I ask the LLM to write a short summary of how it solved the problem, run multiple instances of LLM concurrently, compare their summaries, and use the output of whichever LLM seems to have interpreted instructions the best, or arrived at the best solution.