Glossary
What is meta-prompting?
In short
Meta-prompting is a task for a neural network in which you ask it not to solve a problem but to write or improve a prompt for another task or another model.
Prompts · meta prompt
MMOOZZGG.com editorsChecked: 1 min read
For a child
Sometimes a teacher does not make up a task alone but asks an older student to write one for the little kids: clear, interesting and not too hard. The older student thinks not about the answer but about how to ask well. You can do the same with a smart program. You ask it to write a good request, which you then give to the program itself or to another one.
In plain words
Writing good tasks for a neural network is hard: you must set a role, a format, examples and limits. You can hand this work to the model itself. You describe the goal in your own words, for example: I need an instruction for a bot that answers questions about delivery. Then you ask the model to turn it into a detailed task. The model returns a structured text with sections, and you check and edit it. The same trick helps improve requests: you paste your weak request and ask the model to find its weak spots. It is like asking an experienced colleague to rewrite your specification before you give it to a vendor.
For business
For a company, this technique saves hours on writing instructions for AI assistants, bots and internal helpers: a draft is ready in minutes, and an expert spends time only on editing. It also helps staff with no experience: they describe the task in plain words and get an instruction that the model understands more precisely. The cost is small: one or two extra model calls. The risk is that a generated instruction looks convincing but may contain extra rules or contradictions. So the result is tested on a dozen real cases, not taken on trust. Managers and teams learn to write and test such instructions in corporate training.
For a specialist
Meta-prompting is the name for two related approaches. In the first, a model generates or optimizes a prompt for a task. The Automatic Prompt Engineer method (APE, Zhou et al., 2022) asks a language model to propose a set of candidate instructions, scores each by how well another model solves the task with it without examples (zero-shot), and picks the instruction with the highest score. On 24 language tasks, these instructions matched or beat human-written ones in 19 cases. Suzgun and Kalai (2024) described the second approach: one model works as a conductor, splits a complex task into subtasks and sends them to experts, that is, copies of the same model with their own instructions, then combines and checks their answers. The approach does not depend on the task and needs no examples. With a Python interpreter, this variant beat a standard prompt by 17.1% on average across three tasks. Related ideas: iterative refinement, prompt templates, multi-agent systems.
Example
A head of support writes to the model: Write an instruction for a bot that answers about delivery times and does not promise discounts. A minute later they have a detailed draft with a role, rules and examples, which they only need to test on past requests.
Related terms
Questions and answers
Sources
Where companies use it
Not sure whether your company needs this?
AI solutions map