Meta prompting is using a prompt to write, critique or improve another prompt, usually by asking a capable model to draft a system prompt from a short task description. Anthropic, OpenAI and Google all ship tools built on this idea, but a generated prompt is a first draft that still needs testing on real inputs.
This guide covers the two meanings of the term, the official prompt generators and optimizers, four meta prompts you can copy, and how to test what they produce. Vendor details were checked against official docs on 28 October 2026.
What is meta prompting?
In everyday use, it is prompting about prompts. OpenAI's cookbook recipe on the topic (now marked as archived) puts it plainly: "Meta-prompting is a technique where you use an LLM to generate or improve prompts." It adds that this is "typically" done with a more capable model improving prompts for a less capable one.
OpenAI's prompt generation guide gives the core definition: "A meta-prompt instructs the model to create a good prompt based on your task description or improve an existing one."
The term also has a research meaning, and you will see both in search results:
- Structure over examples. Meta Prompting for AI Systems by Yifan Zhang, Yang Yuan and Andrew Chi-Chih Yao (first posted November 2023) introduces "a framework that emphasizes the formal structure of a task rather than content-specific worked examples."
- One model as a conductor. Meta-Prompting: Enhancing Language Models with Task-Agnostic Scaffolding by Mirac Suzgun and Adam Tauman Kalai (January 2024) has a single model break a task into subtasks, hand each to "expert" instances of the same model, then check and combine the results. With GPT-4 and a Python interpreter, the authors report it beat standard prompting by 17.1% averaged across their tasks.
Most developers searching the term want the first meaning: get a model to write a better prompt than you would on a blank page.
How is it different from few-shot prompting?
Few-shot prompting puts worked examples of the task in the prompt, so the model imitates them. Meta prompting works one level up. The input is a description of a task, or a weak prompt, and the output is a new prompt.
The two combine well. A good meta prompt often asks the model to include examples in the prompt it writes, and Anthropic's metaprompt is itself "a long multi-shot prompt filled with half a dozen examples of good prompts," per its Cookbook notebook.
Which built-in prompt generators and improvers exist?
Each of the three big model vendors has documented at least one.
| Vendor | Tool | What it does | Status as documented |
|---|---|---|---|
| Anthropic | Metaprompt notebook (Claude Cookbook) | Takes a task and optional variable names, returns a prompt template | Linked from the current prompt engineering overview |
| Anthropic | Console prompt generator and prompt improver | Generator drafts a template from a task; improver rewrites an existing template | Announced May 2024 and October 2024 |
| OpenAI | Playground Generate button | Generates prompts, functions and schemas from a task description | Documented in the prompt generation guide |
| OpenAI | Prompt optimizer | Rewrites a prompt using best practices, optionally guided by a dataset with annotations, critiques and grader results | Dataset-backed version is being deprecated with the Evals platform |
| Prompt optimizer (zero-shot, few-shot, data-driven) | Improves system instructions, from a single prompt up to batch runs against labeled data | Documented for Gemini Enterprise Agent Platform |
Anthropic. The prompt engineering overview sends readers without a first draft to the metaprompt recipe. The notebook describes itself as a tool "designed to solve the "blank page problem" and give you a starting point for iteration," and lists two caveats: it is "designed for single-turn question/response prompts, not multiturn," and the result "is not guaranteed to be optimal by any means." Its templates mark variables as {$NAME}. Anthropic's 2024 articles on the Console prompt generator and prompt improver describe the improver adding a reasoning section, standardizing examples into XML and rewriting for clarity.
OpenAI. The Playground's Generate button uses "meta-prompts that incorporate best practices to generate or improve prompts," and "meta-schemas" for JSON and function definitions. OpenAI publishes the meta prompts themselves in the prompt generation guide, which is a useful reference if you write your own. The prompt optimizer is a chat interface in the dashboard. Note the notice on that page: OpenAI "is deprecating the dataset-backed prompt optimizer as part of the Evals platform," with Evals becoming read-only for existing users on October 31, 2026 and the platform scheduled to shut down on November 30, 2026.
Google. The prompt optimizer docs list three modes. The zero-shot optimizer "is a real-time low-latency optimizer that improves a single prompt or system instruction template." The few-shot optimizer refines system instructions "by analyzing examples where a model's response did not meet expectations." The data-driven optimizer "is a batch task-level iterative optimizer" that scores responses to labeled prompts against evaluation metrics.
How do you write a meta prompt?
Give the model what a skilled prompt writer would ask you for. A useful meta prompt names:
- The task and the audience of the final output, in one or two sentences.
- The target model or app, since prompts that suit one model can need changes for another.
- Inputs that change each time, so the generated prompt marks them as variables instead of hard-coding them.
- The output contract: format, length, and what a correct answer looks like.
- What to return: only the finished prompt, or the prompt plus a short list of assumptions.
OpenAI's published meta prompt is a good model of the last two points. It tells the model to "Explicitly the most appropriate output format, in detail" (sic) and to output only the completed system prompt with no extra commentary.
Which meta prompts can you copy?
These four cover most needs. Each uses {{name:default}} blanks, so you can fill them in by hand or in any tool that supports that syntax.
Draft a prompt from a task
You are an expert prompt engineer. Write a complete system prompt for this task: {{task:summarize customer support tickets for a weekly report}}.
The prompt will run on {{model:a current Claude model}} and its output is read by {{audience:the support team lead}}.
Include: a one-line role, the goal, step-by-step instructions, the exact output format, two short examples, and what to do when the input is unclear.
Mark every part that changes per run as a variable in double curly braces with a lowercase name.
Return only the finished prompt, then a list of the assumptions you made.
Critique an existing prompt
Review the prompt below as a strict prompt engineer. Do not rewrite it yet.
List the five biggest problems, ordered by impact on output quality. For each, quote the line, explain the failure it causes, and suggest a fix in one sentence.
Then list three test inputs most likely to break this prompt.
Prompt to review:
{{prompt:paste your prompt here}}
Rewrite a prompt for another model
Rewrite this prompt so it works well on {{target_model:an OpenAI GPT model}}. It was written for {{source_model:Claude}}.
Keep the goal, constraints, examples and variables exactly as they are. Change only structure and wording that the target model handles differently.
After the rewritten prompt, list each change and why you made it.
Prompt:
{{prompt:paste your prompt here}}
Turn a one-off prompt into a reusable template
Convert the prompt below into a reusable template.
Find every detail that would change between uses, such as names, topics, tone, length or file paths, and replace each with a variable written as a lowercase name in double curly braces, with a sensible default after a colon.
Do not change anything else. Return the template, then a table of variables with their defaults.
Prompt:
{{prompt:paste your prompt here}}
How do you test a prompt that an AI wrote?
Treat the generated prompt as a candidate, not a result. Anthropic's overview says prompt engineering assumes you already have "a clear definition of the success criteria for your use case" and "some ways to empirically test against those criteria." OpenAI is just as direct about its optimizer: "it's possible for the optimized prompt to perform worse than your original on specific inputs."
A simple loop works:
- Collect 10 to 30 real inputs, including the awkward ones.
- Run the old prompt and the new prompt on the same inputs.
- Grade both against the same written criteria, by hand or with a grader model.
- Keep the new prompt only if it wins on the cases you care about, and keep the failing inputs as regression tests.
LLM evals covers graders and test sets in more depth, and the prompt engineering workflow shows where this step fits in a full draft, test and ship cycle.
Where should generated prompts live?
Somewhere you can find, compare and reuse them, with the winning version clearly marked. Generators use different variable styles (Anthropic's notebook writes {$NAME}, OpenAI's meta prompt suggests placeholders in brackets), so normalize them when you save.
Kipt keeps prompts as Markdown in a private GitHub repo you own, with {{name}} or {{name:default}} blanks, so the tested version of a generated prompt is the one you copy next time. If a prompt becomes a repeatable workflow for an agent, the free SKILL.md generator turns it into a skill file in your browser.
What should you read next?
The prompt engineering guide covers the techniques a good meta prompt asks for: structure, context placement and output contracts. Few-shot prompting explains how to choose examples, and the prompt templates collection has ready-made starting points you can run through the critique prompt above.