Your bookmarks

Loading your bookmarks

Latest updates

New courses, skills, book chapters, and site improvements will appear here.

Create free accountSign upDashboardDashboard

Meta-prompting

  • Context
  • Research term
4 steps

Meta-prompting is using a prompt to write, analyze, compare, or improve other prompts or instructions, so the AI helps design the instructions for a task as well as doing it.

Writing instructions means putting into words what you already judge by feel. A model can read a stack of results in seconds, describe what the better ones have in common, and write that description as instructions, which makes it a fast first drafter of prompts.

Where you’ll meet it

After each client call, you paste your notes into a saved prompt that drafts a follow-up email. Some drafts are ready to send. Others bury the next step or sound stiff, and you've kept a handful of each kind.

What to do

Give the AI both sets and say what separates them, then ask for three versions of the prompt. Run each one on notes from calls it hasn't seen, and keep the shortest version whose emails you'd send as written.

Easy to mix up

Prompt engineering covers every way of designing prompts, by hand or with AI help. Meta-prompting is one technique within it: using a prompt to write or improve other prompts.

Check yourself

Your saved prompt for summarizing meetings keeps burying the action items. You paste the prompt into a chat and ask the AI to improve it, and it returns a longer, more detailed version. Will the new prompt fix the buried action items?

Go deeperResearch uses the name for more than one idea

In a 2024 paper, Mirac Suzgun and Adam Tauman Kalai used the name for a setup where one model splits a task among fresh copies of itself and combines their answers. The sense here is closer to automatic prompt optimization, as in Automatic Prompt Engineer (APE) by Yongchao Zhou and colleagues. There a model proposes candidate instructions, each is scored on test cases, and the best scorer is kept.

With a dozen test cases, the version that scored best will usually score a little lower on new ones, since part of its lead was luck. What the process produces, such as a system prompt (an app's standing instructions to the model) or a skill file (instructions an agent loads for one kind of task), is an ordinary instruction once written. The meta-prompt is only the request that asked for it.

Practice Modern AI

Or take the learning quiz on a mix of every set