Meta-prompting: Let ChatGPT Write Your Prompts

What if ChatGPT could write better prompts than you? That’s the core idea behind meta-prompting: instead of drafting the instruction yourself, you ask the model to design it. This technique shines in two very different scenarios: when you’re staring at a new task and don’t know where to start, and when an existing prompt works “okay” but you can’t figure out why it’s hitting a ceiling.

Delegate the Design, Not the Intent

The classic approach has you write the prompt, test it, and iterate until it holds water. The meta approach flips the order: you describe your goal in plain English, ChatGPT drafts the prompt, and you simply put it to work. What you delegate is the formatting and completeness of the instruction—the sections you forget, the constraints you wouldn’t have thought to set. What you keep is the definition of the need, which no one else can formulate for you. The skeleton below is enough to get started.

You are an expert in prompt engineering for ChatGPT.
Your job is to create optimized prompts to achieve
the best possible results.

I want a prompt that allows me to:
[DESCRIBE YOUR OBJECTIVE]

The prompt you create must:
- Be immediately usable (copy-paste into ChatGPT)
- Include a relevant role
- Provide clear context
- Specify precise constraints
- Define the expected output format
- Remain simple and direct, avoiding formulas like "think step by step"

Write the complete prompt.

Filled with a real need, this yields a request like this:

You are an expert in prompt engineering for ChatGPT.

I want a prompt that allows me to transform my raw
meeting notes into a professional report sendable to my team.

The notes are usually messy, with abbreviations
and incomplete sentences. The report must be clear,
structured, and include decisions and action items.

Write the complete and optimized prompt.

The returned prompt will almost always contain sections you wouldn’t have thought of: handling abbreviations, managing missing information, rules for separating decisions from discussions. It is precisely those three lines that make the difference between a distributable report and a draft that needs reworking.

Revise a Prompt That’s Hitting a Ceiling

Meta-prompting also serves as after-sales support. Submit your current prompt, explain how the result disappoints you, and ask for a commented rewrite.

Here is a prompt I am currently using:

"""
Write a sales email to sell our accounting software
to SMEs. The email must be short and convincing.
"""

This prompt gives me correct but too generic results.

As a prompt engineering expert, rewrite this prompt
to get more targeted and impactful results.
Explain every improvement you make.

The requirement for an explanation changes the nature of the exercise: you don’t just walk away with a better prompt, but with the reason for every modification—why this role, why this length constraint, why this prohibition. After a few rounds of this, you stop needing the meta-prompt for this whole family of tasks: you have internalized the reasoning.

Get Interrogated Before Being Served

The model produces a more accurate prompt the more it knows about your context, and it knows nothing until you tell it. So reverse the order: ask it to question you first.

I want to create a prompt to help me prepare job
interviews (as a recruiter).

Before creating the prompt, ask me the 8 most
important questions to understand my need and produce
the best possible prompt. Ask the questions one by one,
waiting for my answer each time.

The questions will resemble these: what type of role do you recruit for most often, what skills do you prioritize evaluating, what is the typical duration of your interviews, do you use an evaluation grid, and so on. Once you provide your answers, the resulting prompt sticks to your recruiting practice rather than a textbook interview—it is the variant that yields the best results, by far.

Compare Multiple Versions Before Choosing

Nothing forces you to settle for a single candidate. Asking for three variants built on different principles, followed by a reasoned arbitration, teaches you in a single response what level of sophistication your task actually deserves—sometimes the simple prompt suffices, and you spare yourself the few-shot.

Generate 3 versions of a prompt for writing engaging
LinkedIn posts on the theme of management.

Version 1: Simple and direct prompt
Version 2: Prompt with role and detailed constraints
Version 3: Prompt with few-shot (include examples)

For each version:
1. Write the complete prompt
2. Indicate its strengths and weaknesses
3. Specify which use case it is best suited for

Then recommend the best version and explain why.

Design Entire Systems

This is where meta-prompting becomes truly profitable: on chains of prompts. Written one by one, a few days apart, the links in a process often end up misaligned: step 4 expects a format that step 3 doesn’t produce. Designed in one go, they fit together.

I want to create a complete prompt workflow to manage
the blog post publication process.

The workflow includes the steps:
1. Research and idea brainstorming
2. Creation of a detailed outline
3. Drafting the first version
4. Revision and SEO optimization
5. Creation of the title and meta description
6. Generation of promotion posts (LinkedIn, Twitter, newsletter)

For each step, create an optimized prompt that:
- Takes the result of the previous step as input
- Produces a clear deliverable
- Includes quality constraints

The whole thing must be coherent and form a complete pipeline.

The same logic applies to a thematic collection, when you equip an entire function rather than a process: the salesperson who receives the five prompts below has a complete toolkit, coherent from one moment of the sales cycle to the next.

You are an expert in prompt engineering specializing in
the B2B sales domain.

Create a collection of 5 optimized prompts for a B2B salesperson
selling SaaS solutions to companies with 50 to 500 employees.

The 5 prompts must cover:
1. Research and qualification of a prospect before a call
2. Preparation of a personalized pitch
3. Handling common objections
4. Writing a sales proposal
5. Follow-up email after radio silence

Each prompt must be standalone, ready to use, and include
variables to fill in between [BRACKETS].

Install a Permanent Coach

For delicate tasks, where you feel the first draft won’t suffice, open a two-person work session instead: the model accompanies you iteration by iteration rather than delivering a block you have to evaluate alone.

You are an expert coach in prompt engineering.
Your mission is to help me formulate the best possible prompt
for my task.

Process:
1. I describe my task in simple language
2. You ask me clarification questions (3 max)
3. You propose a first draft of the prompt
4. I give you my feedback
5. You refine the prompt
6. We iterate until I am satisfied

Rules:
- Explain every formulation choice
- Flag potential pitfalls
- Propose alternatives when relevant

My task: [DESCRIBE YOUR TASK]

Your Turn

Try this now. Take the task you delegate most often to ChatGPT and launch the following meta-prompt, which combines clarification and production.

You are an expert in prompt engineering.

I need a prompt for [YOUR MOST FREQUENT TASK].

Before creating the prompt:
1. Summarize in one sentence what you understand of my need
2. Identify 3 pieces of information you are missing
3. Propose the prompt including [BRACKETED] variables
   for the missing information
4. Explain how to optimize the prompt based on the answers
   to the missing information

Three reflexes condition the success of the exercise. What you get is a candidate prompt, not a validated one: run it on a real case before adopting it, because an elegant prompt can very well produce an off-topic result. Don’t stop at the first proposal either, which is rarely the best; ask for variants or a reasoned improvement. And never forget that the model knows nothing about your context: the meta-prompts that work are those where its prompting expertise meets your domain knowledge.

Official sources: Prompt generation · Reasoning best practices (OpenAI API documentation) · Prompting (ChatGPT documentation)

Key Takeaways

• Meta-prompting entrusts ChatGPT with writing the prompt; the intent and validation remain yours
• The interrogative meta-prompt, where ChatGPT asks questions first, produces better-adapted prompts
• Ask for multiple versions and a reasoned comparison to choose the right approach
• Design prompt chains in one go when their steps must fit together
• Always test the generated prompt on a real case before adopting it