Make your AI work better by using meta prompting to write your instructions, you make sure the AI completely understands what you want, giving you the exact answers you need every time.
The why
Meta-prompting is the act of instructing an LLM to modify or write a new prompt for another LLM. Like in human communication, the other person's interpretation of our instructions is heavily affected by how specific we are with said instructions. Or, in other words, how much room we leave for interpretation. The same phenomenon can be observed when interacting with an LLM as well.
You can think of it as a scale: on the left side resides freedom, and on the other, strictness. If we leave the instructions too vague, the model might produce unexpected behaviour, especially if we are using a smaller model or the thinking level of the model is not set well. On the other hand, if we set too many guidelines and rules, we might limit the model too much, especially when a task requires multiple steps to solve; and in such cases, we might not be fully aware of the nature of these steps.
The how
A good prompt has to have the right amount of guidelines, a persona for the model, and perhaps an output structure defined. The good thing is that we can use another model to help us achieve this task.
It's been shown that today's models are getting very good at writing prompts. One thing has to be stated: each model is different in many respects, especially if it was produced by a different lab; hence, models can have different ideas about the same topic. We can use this to our advantage by asking different models to review our prompt and suggest changes.
Also, when instructing the model to review our prompt, it can be useful to tell the model what we are going to use this prompt for.
Example (using gemini 3.1 pro):
Input:
Review the folowing LLM research prompt:
I want to find most popular event hosting websites in Budapest
Output:
Your draft is a solid starting point because it clearly sets the geographic location. However, the phrase "event hosting websites" can be interpreted by an AI in a few distinctly different ways, which might lead to a scattered response.
To get the most accurate and actionable results from an LLM, here is a breakdown of how to sharpen your prompt.
### What to Improve
* **Clarify "Hosting":** Are you looking for **ticketing platforms** to sell tickets and promote an event you are organizing, or are you looking for **venue directories** to rent a physical space (like a conference hall or a party boat)?
* **Define the Audience:** Budapest has a massive expat community alongside its local market. Mentioning whether you want to reach Hungarians or internationals/tourists will completely change the recommendations (e.g., local sites like Eventim.hu versus international hubs like Eventbrite or Meetup).
* **Specify the Event Type:** The best platform for an underground music event is very different from the best platform for a corporate tech summit.
* **Request a Format:** Tell the LLM exactly how you want the data organized, such as a categorized list, a comparison table, or a breakdown of pricing models.
---
### Upgraded Prompt Examples
Here are three ways to rewrite your prompt depending on your actual goal. You can copy and paste the one that fits best:
**Option 1: If you want to sell tickets and promote an event**
> "List the most popular event ticketing and promotion platforms in Budapest. Please categorize them by 'Best for Local Hungarian Audiences' and 'Best for Expats/Internationals'. Include a brief summary of the types of events usually hosted on each."
**Option 2: If you are looking to rent a physical space**
> "What are the best websites or directories for finding and booking physical event venues in Budapest? Focus on platforms that allow me to filter by capacity, price, and event type. Please present the results in a bulleted list."
**Option 3: If you are an attendee looking for things to do**
> "I want to find the best websites for discovering events, concerts, and festivals to attend in Budapest. Please list the top 5 platforms, noting whether they are in English or Hungarian, and what niche they serve."
Even if we don't have subscriptions for multiple model providers, most of the providers have good enough free tiers: Qwen, Gemini, Deepseek, ChatGPT, GML.
Just make sure you enable thinking before you proceed, you can read more about meta prompting here.
Happy prompting!
Top comments (0)