A prompt that makes prompts
Meta prompting uses a language model to generate or improve the prompt for a task, rather than writing every word by hand. You describe the goal, and the model drafts an instruction, examples, or a schema you then test.
Common moves
- Draft where the model writes a first prompt from a task description and constraints.
- Critique where it reviews an existing prompt for ambiguity, missing cases, and weak format rules.
- Rewrite where it produces an improved version addressing the critique.
- Expand where it generates more demonstration examples in a consistent style.
Why it helps
The model often surfaces edge cases and clearer phrasings a human skips. It can rephrase a vague instruction into a precise one and propose a tighter output schema, giving you a strong starting draft to refine.
Keep a human gate
A generated prompt is a hypothesis, not a finished artifact. Always evaluate it on real cases before shipping, since the model can write confident instructions that fail in practice or bake in subtle bias.
Key idea
Meta prompting uses a model to draft, critique, and refine prompts, surfacing edge cases and clearer phrasings, but each generated prompt stays a hypothesis to evaluate on real cases before shipping.