Prompt expansion of input to extra categories makes sense if your embedding isn’t working well. But on its own, why use the LLM at all? I think you could have demonstrated the original step first and then shown that it’s useful.