In plain terms
Models are good at writing clear instructions, and a prompt is a set of instructions. So you describe the task to a model and ask it to draft the prompt. Then you run that prompt, show the model where the results fell short, and ask it to revise. You move from author to editor.
Why it matters
Prompt quality is a bottleneck in most AI projects, and few people enjoy writing long, careful prompts. Meta-prompting produces a solid first draft in a minute and speeds up iteration. It does not remove the need for judgment: someone still has to define what good looks like and test whether the new prompt achieves it.
Example
A product manager needs a prompt for summarising customer interviews. She gives a model the goal, two sample transcripts and one summary she likes, and asks for a prompt. The draft covers things she had not thought to specify, such as how to treat contradictory statements. She tests it on ten transcripts and feeds two weak results back for a second revision.
Most often confused with
Meta-prompting vs. Prompt Engineering
Meta-prompting is one technique within prompt engineering, a way of producing and improving the text. The rest of the discipline still applies: a clear task definition, test cases, measurement and review. A model can write the prompt; it cannot decide for you what the task should achieve.
Under the hood
Common forms: prompt generators that expand a short task description into a structured prompt with role, steps, format and placeholders; critique loops in which a model reviews outputs against criteria and proposes edits; and automated optimisation, where candidate prompts are generated, scored on a test set and selected over several rounds. Frameworks such as DSPy treat prompts as parameters to be optimised against a metric. Risks: prompts that overfit the test examples, prompts that grow long without improving, and unreviewed changes. Keep a held-out set of cases and a human reading the diffs.