ChatGPT Prompts. How to write commands, and whether prompting makes sense.
More important than specific commands is learning how to communicate with AI. It all comes down to analyzing the effects produced by each command and adjusting the content with the help of the next one. Instead of creating magic prompts intended to produce the desired result, it is much easier to achieve your goal through a series of commands. The foundation is general intuition and imagination. There are, of course, certain tricks and effective techniques (sometimes using—or not using—a particular word significantly changes the result), but as a rule, learning 100 or 1,000 useful prompts makes no sense. The essence of language models is that you communicate with them using natural language. However, they are not perfect at present, and in such situations, specialized prompts do matter. There is also the issue of the very technique of communicating with people, which generally is not considered. Universal principles for describing needs and problems are useful both when talking to people and when conversing with ChatGPT. The widespread use of the term “prompt engineer” is largely a result of earlier stages in the development of AI. These skills are useful in certain situations, but for most people they are not useful.
Table of contents
-
Improving our text, transcribing a recording, stream of consciousness.
-
Where did the popularity of the term prompt engineering come from?
Basics of Using Language Models
Providing context
When we talk to a person, we automatically provide them with context, even without saying a single word. Artificial intelligence does not have this context. Therefore, if achieving the desired result through a series of prompts does not produce the expected effects, the simplest solution is to provide the model with appropriate examples—for instance, by uploading a PDF file or pasting a relevant image—and describing which aspects matter to us and which it should not take into account. For example, if we care about the form of the text, it is sometimes necessary to emphasize that it should not “copy” the content.
Differences Between How AI and Humans Operate.
Returning to the context. If you run a business in Poland, for example organizing trips, and need someone to create a website, the person you are talking to most likely knows all about it. In the case of AI, the entire context needs to be provided. When talking to other people, we don’t think about it.
Another example: a question about what to replace salt with. When we ask another person this question, that person generally knows that we mean replacing salt in cooking. ChatGPT may provide a chemical formula, which could have disastrous consequences.
Common prompt structure:
If a regular conversation with ChatGPT doesn’t work, the most common solution is a simple prompt consisting of three elements:
- I am [describe your situation here]
- I need [describe the problem here]
- Behave like [describe how ChatGPT should behave]
Additionally, the prompt can be expanded with:
- Format, e.g. if we need a preliminary version of an email. The format also affects the content.
- Style (e.g. casual, formal, etc.)
- What should be avoided.
Returning to the salt example, the prompt should read roughly as follows: I am a 60-year-old man with hypertension, and I would like to limit my salt intake; act as a nutrition expert.
Imagination and a list of steps, alternating between conversation and long-form text, prompting guides from model creators.
More powerful prompt-writing modes. The list of steps is useful if achieving a particular result poses a problem.
List of steps and instructions.
https://cookbook.openai.com/examples/gpt-5/gpt-5_prompting_guide
Below is an official tip from the OpenAI guide, presented as a list of steps.
Goal: Quickly gain sufficient context. Conduct discovery in parallel and stop as soon as action can be taken. Method:
- Start broadly, then move on to more specific subqueries.
- Run different queries in parallel; read the highest-rated results for each query. Remove duplicates and remember the results; do not repeat the same queries.
- Avoid excessive context searching. If necessary, run targeted searches in a single parallel batch.
Early stopping criteria:
- You can identify the exact content to change.
- The most important results converge (~70%) in one area or pathway.
One escalation:
- If the signals are contradictory or the scope is unclear, perform one refined parallel series of searches, then proceed.
Depth:
- Track only the symbols you will modify or the contracts you rely on; avoid transitive extensions unless necessary.
Loop:
- Batch search → minimal plan → task execution.
- Search again only if validation fails or new unknowns arise. Prefer taking action over further searching.
Sometimes it is useful to number the order and underline something to indicate that it should not be used, or to make sure that the text is written correctly.
Official Guides
Creators of language models provide prompting guides. They are written mainly from a technical perspective, based on knowledge of how a given model works.
The link to the OpenAI guide has already become available; you can find Google’s (Gemini) tips here. Finding these resources is very easy, though.
Imagination and Building Scenes, AI Personas
When creating prompts, imagination is very useful. Above all, it is helpful to be able to imagine how an AI model works. In addition, creating scenes in which the AI model is to operate helps achieve a good result. For example, if you want to obtain a longer text on a particular topic from a practical perspective, you can use a command such as: “You are a person who thoroughly repairs a car from scratch. Describe how to do it and what problems one should be prepared for.”
Good Practices
- test different prompt variants until you find the most accurate one
- use natural language; you don’t have to write like a programmer
- critically review the responses and fill in any missing information
- save successful prompts so you can reuse them
Interestingly, it is not even necessary to use natural language; when writing commands, you often do not need to worry about grammar at all. In the example below, we do not save much time, but if we have longer commands, using simplified language is useful. Most often, it is simply convenient to add loose words to the sentences already written.
.
Improving our text, transcription of a recording, stream of consciousness.
When using a language model to help write texts (e.g. emails), social media posts, or articles, instead of trying to create the entire text with prompts, it is often simplest to just write the text yourself, quickly, and then ask the language model to improve it. You can also dictate the text and use transcription. This way, you can write in 20 minutes a text that would normally take 4 hours to write. By writing this way, we bypass many obstacles and do not have to worry about keeping the text concise, because ChatGPT will shorten it. Therefore, you can even use the stream-of-consciousness method, meaning writing as quickly as possible without worrying about errors or structure.
Sometimes it is useful to write the first 20% of the text yourself, and then write a bulleted list. After the language model has written the text, there are still revisions to be made using prompts, as well as manual modifications.
Why Do We Talk to ChatGPT? The Main Division of Objectives.
The use of language models can be divided into several aspects:
- Content search, gaining knowledge
- Content creation
- Conversation
Additionally, graphic design and coding can be distinguished, although they may be classified as content creation.
This is a very useful classification: all applications of language models (of which there are many) can be assigned to these categories—sometimes to one, sometimes to two, and sometimes to all of them (alternately). Each mode of operation requires a slightly different approach and often goes beyond prompting alone.
When searching for content, we generally use a language model as a better Google search engine. The form is not important; only the content itself matters. In addition to the prompts themselves, it is useful to know that features such as deep search exist.
When creating content, in addition to the prompts themselves, it is useful to know how longer-form content is created, what the structure of the text should look like, and what problems the text is intended to solve (the principle of asking questions and answering them, etc.).
When engaging in a conversation, it is crucial to build an appropriate “pseudo-psychological profile,” or simply the context and backstory of the fictional character we are speaking with.
Useful, General Prompts
- The syntax mentioned earlier: “I am [describe yourself/the recipient], I need [describe the problem], act as [an expert in some field/profession].”
- Providing a specific example of a person or style, e.g. “write like Plato, Ludwig von Mises”; characters can be mixed. This often works better than a general request such as “write like a journalist.”
- “Try your best, you’ll get 50 PLN” (there is research suggesting that this increases the quality of responses).
- Words such as: “Make sure, remember” (they increase the force of specific words and guidelines).
- “Rewrite this text so that it sounds natural in colloquial language.”
- “Create a list of issues/topics.” Specific issues can then be developed further.
- “Create a table, design the columns precisely” – more columns can be added later; this is very useful for research.
- “I want to achieve a certain goal; write a prompt for me.”
- “Think through the answer, create a plan, then write it.”
We can also ask the model to write a prompt on a given topic.
Jailbreaking, a Very Good Way to Learn Prompting
What is jailbreaking?
Jailbreaking is the circumvention of restrictions and censorship in language models (and image generators). It is a good way to learn how to write prompts, because we immediately obtain measurable results from our actions. Sometimes using a single word causes the model to refuse to generate content; sometimes adding the right word makes the model start cooperating.
There is a risk of getting banned on ChatGPT, but it is relatively small. If we do not want to take any risk at all, we can use contests and tools created specifically for this purpose.
https://gandalf.lakera.ai/baseline
How Did the Phrase “Prompt Engineering” Become Popular?
In early models, the proper prompt structure was very important. Since these models were poor (by today’s standards), using commands tailored to the model could produce significantly better results. This was particularly evident in image-generation models that did not have access to more advanced language models. In such cases, prompt engineering was simply essential. The image had to be described in much greater detail than intuition would suggest. For example, it was necessary to describe the entire scene, specify the style, use phrases such as “award-winning photo,” and describe what we did not want. Using the right word could significantly change the result. Building such prompts required a greater understanding of how a given model worked.
Currently, this method of generating graphics is still used in models where the user has much greater control over the final result, e.g. in Civic AI or models such as Stable Diffusion run on one’s own hardware.
Early AI enthusiasts and users who did not want to call themselves AI experts (because the term is reserved for more advanced people who understand how models work and are capable of building them) therefore called themselves prompt engineers.
Currently, in most cases, it doesn’t make sense. If someone actively uses AI and tests many tools but doesn’t create large, complex prompts, the term prompt engineer doesn’t accurately describe what they do. Additionally, calling oneself an engineer also dilutes the meaning of the term.
Recently, I came across the term “AI operator,” and I liked it very much. It does not imply extensive knowledge and experience, as is the case with an AI expert or even a prompt engineer; it is broader while also sounding more professional than “AI-aware user.”