How do I use "Metaprompting" to make an LLM write better prompts for me?

All answers to this question.

Metaprompting is incredibly effective for breaking out of "creator's block." To make it work, your master prompt should tell the AI: "You are an expert Prompt Engineer. Your goal is to help me write a system prompt for [Task]. Before you give me a prompt, you must ask me at least 5 clarifying questions about the target audience, the desired tone, the specific constraints, and the expected output format." By forcing the AI to gather requirements first, the final prompt it generates is much more aligned with your actual needs than a generic one-off request.

   Answered 2025-02-02 by Mary Johnson

Does this technique work better with larger models like GPT-4, or can smaller open-source models handle the meta-reasoning required?

   Answered 2025-02-04 by William Ross

• William, in my experience, this is one area where the larger models really shine. To effectively "metaprompt," the model needs a deep understanding of its own architecture and how different phrasing affects its output. While a model like Llama 3 (70B) can do a decent job, GPT-4 or Claude 3.5 Sonnet are much better at identifying the nuance needed for a high-performing prompt. Smaller models tend to give very generic advice, whereas the larger ones can suggest advanced techniques like variable placeholders.

     Commented 2025-02-05 by George Adams

It's basically recursive improvement. I use a metaprompt to generate my initial draft, then I ask the AI to "critique" that draft and improve it.

   Answered 2025-02-06 by Donna Campbell

• That iterative approach is key, Donna. Having the AI act as both the "writer" and the "editor" of its own prompts yields much higher quality results.

     Commented 2025-02-07 by Mary Johnson