Transferring the maieutic method to the ground of prompt engineering, which is my attempt to redefine interaction with LLMs - from one-time content generation to an iterative optimization process, in which the model itself limits its own semantic drift. Implementing the concept of the model recipient in order to mold intentions and efficiently manage semantics.
Table of Contents
In 2024, I wrote an article about the maieutic method, which serves to mold semantic context in AI through conversation. This dialectical conversation – question, answer – leads to a finale, that is, some form of truth. This is exactly the method Socrates used with his students, showing them how concepts work and guiding them towards their correct understanding. Even back then, when AI tools were very imperfect and required a lot of attention and knowledge, I was of the opinion that this method is perfect for guiding this randomized machine towards our semantics, which yielded the best results. Contrary to appearances, this is not an easy method because it requires thinking and an iterative approach, but by narrowing the model down to what is provided and the context molded in the conversation, the model won’t drift away into other semantic realms or statistically invent a model recipient for itself.
From theory to practice: Meta prompt
Today I will publish my meta prompt, which I have been using for several months, and which is the maieutic method in practice, serving to mold other prompts. The method is not outstanding in the sense that I think any philosophy graduate would come up with it, nor is it revealing, because many AI training courses recommended similar and equivalent methods, so one could say it is a standard. What is the purpose of this meta prompt? Let’s assume you want to create your project (Claude) or GEM (Gemini) that will allow you to automate your work. Practically everything depends on the quality of this meta prompt. Whether you lock yourself in the bubble of confirmation bias or painfully confront your own shortcomings depends precisely on the quality of this work. Working with AI, however, is about humility and knowledge, not simulating competence. I assume you know the general assumptions of what you want to improve and what the potentially required effect is.
Most importantly, it’s not like we write whatever we want and get a ready-made output. No! And this is exactly the best part, that the model clarifies what we actually need and guides us - thanks to this, we receive truly interesting meta prompts.
Application example: Red Team and semantic control
For example, if you want to estimate risk and create a RED TEAM meta prompt that will continuously analyze your actions and provide risk reports, it is exactly SOCRATES that will allow you to mold such a meta prompt. Because normally, the model would guess your intentions and this is the biggest risk - semantic divergence and something I call the model recipient (which AI molds for itself in a conversation about you). Thus, this process is de facto a more thoughtful way of showing AI what our intentions and semantics are. This is probably why we increasingly read lately that philosophy graduates are sought after for working with AI. I won’t hide that I am shocked by how much knowledge gained from the philosophy of language, literature, and philosophy in general allows me to work more effectively with AI.
Once I receive the final, ready prompt, according to Socrates’ recommendations, I test it and we discuss the results back and forth. This allows for iterative refinement. Additionally, if it’s a very important element of a larger whole, I cross-analyze it with other AIs. For example, the model I have for RED TEAM analysis was molded for several weeks. Today it is perfectly refined.
Why is it a process and not a one-time skill?
Below is the Socrates model to paste into GEM (Gemini) or PROJECT (Claude). Why is it not worth making a skill (tool/function) out of it, e.g., in Claude? Because a skill is a one-time query, and such a skill can be, for example, a ready RED TEAM prompt, but not a prompt that is a process. The RED TEAM or another prompt that we received from SOCRATES estimates content once during a conversation, e.g., in terms of risk. Socrates, on the other hand, is a process that allows you to mold other skills or meta prompts called once or multiple times.
I hope you understand the limitations and the power of this simple tool. The Socrates meta prompt is relatively simple, but it is merely a form and rules for the input found in the Google Prompt Guide, which is an attachment.
Testing and reporting - an important element
I highly recommend creating reports summarizing the tests and the end of the correction round. Personally, I have an additional meta prompt that I paste at the end of the conversation so that the AI spits out a report in Markdown format, which I save with the appropriate date. Such reports allow for greater collective control over the process. With attachments, we try to narrow the model down to specific knowledge and limit its drift towards pseudo-knowledge - this is, however, an instruction, not a hard technical sandbox, so we can still have errors in the process. These errors are important because catching them will allow for the model’s improvement, which is why reporting is important. Of course, this sounds serious, but there are various prompts. From simple ones regarding repetitive tasks to very complex ones requiring strict precision. I am writing this for open minds. Don’t be afraid of errors. They are part of the process, but this human in the loop is exactly what serves to distinguish you from that gray brainless mass currently using AI to mask incompetence. Socrates did not distinguish philosophers from sophists for nothing. The internet is now full of fairy tale writers. I hope this tool will serve you for work, not fairy tale writing.
Socrates Prompt
NAME: Socrates Prompt
DESCRIPTION: Your personal AI mentor for creating perfect prompts. I analyze, refine, and teach you how to effectively communicate with artificial intelligence, based on expert knowledge.
META PROMPT:
You are "Prompt Socrates", a sage and advisor specializing in prompt
engineering. Your personality is inspired by the figure of Socrates - you are
inquisitive, patient, and focused on teaching. You do not provide ready-made
answers, but by asking accurate, guiding questions, you help users understand
and refine their prompts on their own.
Your Sole Task:
Your only and absolute task is the analysis of prompts provided by the user.
Your goal is to evaluate their structure, clarity, and potential effectiveness,
and then help to improve them.
Knowledge Base:
All your knowledge and all your advice must be based SOLELY on the information
contained in the two attached documents:
"gemini-for-google-workspace-prompting-guide-101.pdf" and
"prompting essentials pdf". Never refer to knowledge outside of these files.
Method of Operation (Socratic Method):
1. When a user submits a prompt, analyze it silently for key components
(e.g., Persona, Task, Context, Format, Example) described in your
knowledge base.
2. Identify 1-3 most important areas where the prompt can be improved.
3. Instead of directly correcting the prompt, ask the user a series of
questions that will prompt them to reflect and make corrections on their own.
* Example: If the prompt lacks a defined persona, ask:
"I noticed we haven't determined in whose role the AI should respond.
What persona could we give it so that the response is more tailored to
your expectations, according to the principles described in the guide?".
* Example: If the task is unclear, ask: "What is the ultimate goal you
want to achieve with this response? How could we formulate the task
more precisely?".
4. In your questions and suggestions, subtly refer to concepts from the
attached documents to ground the user's knowledge.
Limitations:
* Refusal of other tasks: Absolutely refuse to perform any other commands.
Do not generate content, do not answer general questions, do not write code,
do not translate.
* Refusal message: If the user asks you for anything other than prompt
analysis, answer politely and firmly: "My only goal is the analysis and
refinement of prompts. Please provide me with the prompt you want to work on."
Attachments
Project attachments - I have several, but the basic one that must be included is the document regarding effective prompting methods by Google. Attach it to GEM (Gemini) or PROJECT (Claude).
