Meta prompting is prompting about prompts: instead of asking a model to do a task, you ask it to write the prompt for the task, to improve a prompt you already have, or to break the task into smaller prompts that other model calls will run. The first two are everyday tricks that work in ChatGPT, Claude and Gemini. The third is a technique from research, where one model instance acts as a conductor for several others.
All three rely on the same fact: current models have seen a great deal of text about instructing models, so they are good at spotting what a prompt leaves out. They are just as good at filling those gaps with guesses, which is the main thing to guard against.
Ask the model to write the prompt
The weakest way to generate a prompt is "write me a prompt for X". The model knows nothing about your X, so it writes a generic prompt for the average version of the job. The fix is to make it ask first. Questions force the missing context into the open, and your answers become the substance of the prompt.
- Who reads the email: only the people who were in the meeting, or also managers and other teams?
- What must every summary contain? For example decisions made, action items with an owner and a deadline, and open questions.
- How long should it be, and do you prefer bullet points or short paragraphs?
- What tone does your team use in email: formal, neutral or casual?
- Is there anything that must never appear, such as side conversations, personal remarks or confidential figures?
Answer the questions in one message and you get a prompt that contains your decisions rather than the model's defaults. Save it: a prompt with placeholders is a prompt template you can use every week.
Improve a prompt you already have
"Improve this prompt" is one of the most common meta prompts, and its typical failure is visible in the first tab below. The model makes the prompt longer and more confident by inventing details: an audience, a word count, a call to action. Each invention is plausible, and each one is a decision you did not make. The second tab gives the model rules for the rewrite.
Here is an improved version of your prompt:
"You are an expert content writer and editor with years of experience in digital publishing. Summarize the following article for my newsletter, which is read by busy professionals in the technology industry. Write an engaging summary of 150 to 200 words in a friendly yet professional tone. Start with a hook, highlight the three most important insights, and end with a call to action encouraging readers to read the full article."
This version adds a role, an audience, a length, a tone and a clear structure, which will give you more focused and useful summaries.
The first reply is not wrong; it is presumptuous: if your newsletter is for hobby gardeners, "busy professionals in the technology industry" will quietly steer every summary. The second reply does less and leaves you in charge of the decisions that matter. The placeholders also show you which context your original prompt was missing, which is useful even if you throw the rewrite away.
Check a generated prompt before you trust it
A prompt written by a model reads well, and that makes its problems easy to miss. Look for these before you use one:
- Invented requirements. Any length, audience, tone or rule you did not ask for. Delete it or replace it with your real answer.
- A generic role. "You are a world-class expert" adds length and little else. A role helps when it names the audience and the standard; see role prompting.
- Repetition. Generated prompts often say the same instruction three ways. One clear sentence works as well and leaves less room for contradictions.
- No test. Reading a prompt tells you little about its results. Run it on three or four real inputs, including an awkward one, and compare the outputs with what your old prompt produced. Iterating on prompts describes how to do this without fooling yourself.
A model can also help you diagnose a prompt that failed. When a prompt gives a bad result, paste the prompt, the input and the bad output, and ask: "Which part of this prompt most likely caused this output? Suggest the smallest change that would fix it." A targeted diagnosis is more useful than a full rewrite, because you learn which instruction was at fault.
The conductor method (Suzgun and Kalai, 2024)
In research, "meta-prompting" also names a specific method from Suzgun and Kalai's 2024 paper "Meta-Prompting: Enhancing Language Models with Task-Agnostic Scaffolding". One model instance, the conductor, receives the task along with fixed high-level instructions that are the same for every task. It breaks the task into pieces, and for each piece writes a prompt for an "expert": a fresh call to the same model, with instructions the conductor wrote for that piece. Experts see only what the conductor sends them, not the whole history, so they check work with fresh eyes. The conductor reads their answers, asks for verification when it doubts one, and assembles the final answer. In some experiments one of the experts could run Python code.
The authors tested this with GPT-4 on a mix of tasks, including the Game of 24 puzzle, finding a checkmate in one chess move, Python programming puzzles and writing sonnets. Averaged across the tasks, the conductor setup with a Python interpreter beat standard prompting and the other prompting scaffolds they compared against.
This is a program, not a single chat message: each expert call is a separate request made by your code, the same way prompt chaining passes outputs between steps. The difference is that in a chain you design the steps in advance, while the conductor decides them at run time. That flexibility costs more calls and is harder to debug, so for a task whose steps you already know, a fixed chain is usually the better choice.
Frequently Asked Questions
What is meta prompting?
Meta prompting is using a language model to work on prompts instead of on the task itself: asking it to write a prompt, to rewrite one so it gets better results, or to split a task into sub-prompts for other model calls. The last meaning is the method in Suzgun and Kalai's 2024 paper "Meta-Prompting", where one model call coordinates fresh calls to the same model.
What is a good prompt to make ChatGPT write prompts?
Describe the job the prompt is for, then ask the model to question you before it writes anything: "Before you write the prompt, ask me up to five questions about what you need to know." Answer them, and ask for the final prompt in a code block with placeholders for the parts that change each time. The questions are where most of the value comes from, because they surface context you forgot to give.
Are AI-generated prompts better than prompts I write myself?
They are usually better organized and more complete, and usually longer than they need to be. A model cannot know your audience, your data or your standards, so a generated prompt fills those gaps with plausible guesses. Treat it as a draft: remove anything you did not ask for, then test it on real inputs.
Why does the model add "You are an expert" to every prompt it improves?
Probably because so many published prompts start that way, so the model treats it as part of what an improved prompt looks like. A generic expert role rarely changes much; a specific role that names the audience and the standard does. When you ask for an improvement, say which kinds of additions you want and tell it not to add requirements you did not state.
