OpenAI recently released a new prompting “cookbook” for GPT-5 and confirmed what many of us have seen in practice: prompt engineering isn’t a temporary hype cycle-it’s a core skill...
OpenAI recently released a new prompting cookbook for GPT-5, and in essence confirmed what many of us have long suspected in practice: prompt engineering is not just a temporary hype, but an absolutely key skill for anyone serious about AI-first products.
You can have the best and newest model, but if you can't brief it properly, you're barely halfway there. GPT-5 is undoubtedly a new league, bringing smarter reasoning, more precise work with tools (tool calling) and overall more robust agentic workflows. Even so, the real key to unlocking its potential remains what instructions you actually send it. And this is where the fascinating topic of "meta prompting" opens up.
Meta prompting: the hidden weapon most people don't use
Put simply, meta prompting is "prompting about prompting". In practice it means you don't think only about how to phrase one specific command for the AI as well as possible. Instead you think about how to create a system for creating, testing and managing prompts. It's a shift from one-off commands to building scalable and sustainable "prompting frameworks".
Often it means letting the AI itself revise, evaluate and propose improvements to your own prompts. I do this day in, day out: I iterate prompts with the help of a smarter model, test them in real scenarios, collect feedback from end users and run miniature "A/B tests" for every new version of a prompt. It's like having an AI coach who helps you shorten the road from a "good prompt" to an "excellent prompt".
While most people are still working out how to write one good prompt, meta prompting lets you create a system that generates dozens of quality prompts for you in a fraction of the time. It's the difference between making yourself one hammer and building a hammer factory.
The official cookbook from OpenAI finally formalises what I've been observing in practice for months. Instead of trying to find the perfect wording by hand, you use a stronger model (in the OpenAI examples it's o1-preview) to generate an optimised prompt for a specific task, which you then deploy on the target model (say gpt-5). That not only speeds up development, but also gets you systematically better results.
A concrete example from the OpenAI cookbook
Let's look at a practical example. You start with a simple prompt:
"Summarize this news article: {article}"Then you use a meta-prompt for improvement:
Improve the following prompt to generate a more detailed summary. Adhere to prompt engineering best practices. Make sure the structure is clear and contains the type of news, tags and sentiment analysis. {simple_prompt} Only return the prompt.The result? A comprehensive prompt that demands structured output with categorisation, tags and sentiment analysis. The difference in output quality is dramatic, instead of a simple summary you get a richly structured analysis.
In the Czech environment meta prompting isn't very well known yet, which is a pity. I see a huge opportunity in it for those who start using it before the competition. It's not just about efficiency, it's about being able to achieve systematically better results than those who still rely on tuning prompts by hand.
What the new OpenAI cookbook brings: concrete techniques and my experience
The new official cookbook from OpenAI isn't just theoretical material, it's full of practical techniques you can use right away. Let's look at several key concepts and how they match my experience from practice.
Practical examples of using AI (and why meta prompting plays a role)
Czech Retail Efficiency
- In the Czech Republic retailers use AI to optimise stock and planning → pilots show up to a 30% reduction in operating costs and ~20% fewer out-of-stock items.
- Hyundai sped up planning from days to 5 minutes and saved 520,000 EUR.
- Bednar FMT moved from Excel to AI forecasts, which led to revenue growth.
Logistics / Order Management Transformation
- AI transformed the order system → faster synchronisation, fewer errors and higher efficiency in deliveries.
AI Product Managers in 2025
- According to Product School, PMs who don't integrate AI are quickly falling behind.
- A study shows ~40% higher productivity for PMs with generative AI in their workflows.
Generative AI in Agile
- A Fortune 500 technology company deployed generative AI to evaluate epics in an agile process, resulting in faster backlog assessment and better prioritisation.
- Result: faster quality assessment, but also limits in context and nuance → shows the need for "meta prompting" for better results.
Controlling "agentic eagerness": how to tame AI's excessive initiative
In the cookbook OpenAI describes how to steer the "eagerness" of GPT-5. By default the model is very thorough and wants to explore every detail. Sometimes that isn't desirable. The cookbook offers concrete solutions:
For less proactive behaviour:
<context_gathering> - Search depth: very low - Bias strongly towards providing a correct answer as quickly as possible, even if it might not be fully correct. - Usually, this means an absolute maximum of 2 tool calls. - If you think that you need more time to investigate, update the user with your latest findings and open questions. </context_gathering>For more autonomous behaviour:
<persistence> - You are an agent - please keep going until the user's query is completely resolved. - Only terminate your turn when you are sure that the problem is solved. - Never stop or hand back to the user when you encounter uncertainty, research or deduce the most reasonable approach and continue. </persistence>From my own experience of building AI-first products I can confirm that these principles work:
Structure and clearly defined rules are the foundation. Exactly what OpenAI formalises in the cookbook. Agentic workflow means the AI decides "for you". If it doesn't have precisely set guardrails, it easily slides into either doing too little or, on the contrary, starting to take over control. The cookbook solves this with new parameters like reasoning_effort and verbosity, plus clearly defined rules in the prompt.
Tool preambles: how the AI communicates its progress
Another great technique from the cookbook is "tool preambles", a way for the AI to explain what it's doing and why:
<tool_preambles> - Always begin by rephrasing the user's goal in a friendly, clear manner. - Then, immediately outline a structured plan detailing each logical step. - As you execute your tasks, narrate each step succinctly and sequentially. - Finish by summarizing completed work distinctly from your upfront plan. </tool_preambles>This is something I had to painfully learn in practice. When AI does more complex tasks, users need to know what's going on. Without these "progress reports" the AI comes across as a black box.
One prompt for everyone is a myth. What works for one team or use-case fails in another context. The most effective approach is to see a prompt as a flexible template, a base to which every product, team or process adds its own specific "sub-rules" and rules. Meta prompting is the ideal tool for managing these templates and adapting them.
Debugging prompts is a new discipline. What I enjoy most is that I can use GPT-5 to analyse and revise my own prompts. You just ask the model what it would change so the prompt produces a better result. It often proposes edits that wouldn't have occurred to me even after ten manual iterations. It is exactly the principle OpenAI formalises in meta prompting.
Responses API: a game changer for agentic applications
The cookbook strongly recommends switching to the Responses API, where reasoning persists between individual tool calls. OpenAI gives concrete numbers: on the Tau-Bench Retail benchmark the score improved from 73.9% to 78.2% just by moving to the Responses API. That isn't a marginal improvement, that's a dramatic difference.
A prompt engineer today is effectively a product manager. In AI-first workflows a prompt is often more valuable than part of the backend code. Prompts are versioned, documented, tested and iterated just like any other digital product. It's not just text somewhere in the code, it's a central element of the product that defines its behaviour.
What to take from this?
The new cookbook from OpenAI isn't just a set of tips and tricks. It's confirmation that we are moving from simple "assigning tasks" to systematic "managing AI systems". Prompts are no longer static labels in a text box. They are living documents that have to be regularly revised, scaled and tested.
The biggest accelerator in this process is precisely meta prompting: letting the AI analyse its own prompts, look for weak spots and propose miniature changes. If you want to move your AI-first workflow forward quickly, meta prompting will save you an incredible amount of time and effort.
A prompt today isn't just an add-on, it's a building block of a digital product, automation and scalable innovation. Whoever tunes prompts as carefully as code has a lead. Whoever expects the model to "compensate for it all" will always be a step behind.
How about you? Do you have your own best practices for debugging prompts, or are you still expecting it to "turn out somehow"?
Sources:
- [1] OpenAI. (2024, October 23). Enhance your prompts with meta prompting. OpenAI Cookbook: https://cookbook.openai.com/examples/enhance_your_prompts_with_meta_prompting
- [2] OpenAI. (2025, August 7). GPT-5 prompting guide. OpenAI Cookbook: https://cookbook.openai.com/examples/gpt-5/gpt-5_prompting_guide
The article was first published on LinkedIn. Original article on LinkedIn
