The art of writing good prompts, or commands for artificial intelligence (AI) systems , is one of the market competencies that is becoming an absolute must-have on the market. Thanks to it, you can discover more possibilities hidden in generative AI, solve additional problems, or build entire action scenarios. Many users abandon ChatGPT after the "first play" with commands, obtaining answers of varying quality. Without knowing the full context, it is easy to fall into the trap of testing different tools and that is where the adventure ends. It is good to know all the possibilities, limitations and principles of operation of ChatGPT in order to use it better.
Contents
Let's start with the major limitations of the ChatPGT model. First of all, it is a knowledge base that is current until January 2022 (version 3.5). So there is no access to events, publications and content available after that date. There is a risk that you will get false information in the scope of questions constructed in this way. The model will also have a problem understanding the context regarding events after that period.
Additionally, ChatGPT has no consciousness or the ability to experience emotions. It only processes information and provides answers on an ongoing basis. You can check this by asking questions like:
o czym myślałeś wczoraj? co myślisz o ... ?
On the other hand, ChatGPT answers are always limited to the given context and prompt you use. The more general the query, the more general (and imprecise!) the answer. Try asking about space, for example, and the answer will probably be short and general
jak duży jest kosmos?
Each newly started chat has its own limited memory . At some point in the conversation, ChatGPT may get lost and ask you to start a new chat window. It is a good practice to start a new chat window in each new context. Remember that you can go back to the previous conversation and edit it to change the direction of the conversation that you do not want to start over.
Want to learn AI effectively? Sign up for the first CampusAI school in Poland!
The free version of ChatGPT (the one marked 3.5) will be more than enough for private users or small projects that need basic text generation and workflow functionality. The paid version opens up more possibilities, showing significant improvements in understanding the subtleties of human language, including humor and sarcasm. Its responses are more polished, coherent, and creative. ChatGPT 4.0 can process graphics , allowing for a wider range of applications, such as analyzing charts, explaining memes, and summarizing documents containing both text and images. ChatGPT 4.0 is capable of processing 25,000 words of context , compared to 3,000 words in version 3.5.
Unlike the free version, ChatGPT Plus can access online resources, providing up-to-date information on current events.
Functionality | ChatGPT 3.5 | ChatGPT 4.0 |
---|---|---|
Input data | Text only | Multi-modal (text and image) |
Processing | High but limited in solving complex problems | Much higher, can solve complex scientific problems |
Understanding the details | Basic understanding | Better understanding, especially of humor, sarcasm, but also mathematics |
Creativity | Basic | More creativity and consistency, especially in long texts |
Maintaining context | Up to 3,000 words | Up to 25,000 words |
Accuracy | Sometimes prone to mistakes | Reduced susceptibility to creating “false stories” |
Cost | $0 | $20 or $25 for Team Edition |
A good prompt is, above all , a consistent, precise prompt that contains full context , supplemented with the expected format and style. Over time, you will notice that mastering prompts is not easy, and the results you get can vary significantly.
Let's start with an exercise. You ask the first prompt:
napisz artykuł na temat zdrowej diety
You get a result. The quality of the content answers the query, but it is not very detailed, it does not delve into any specific aspect of the diet.
So let's try to give this some context and check the quality of the information.
napisz artykuł na bloga na temat zdrowej diety skupiając się na poniższych zagadnieniach? - jakich produktów powinno się koniecznie unikać w diecie? jakie są jednoznacznie złe? - jakie badania warto wykonać, aby sprawdzić czy moja dieta jest prawidłowa? - czym są makroskładniki w diecie i jak wpływają na zdrowie?
We already have a pretty specific article that takes into account the context. We can go further and add a specific group and more precision, e.g.
napisz artykuł na bloga na temat zdrowej diety dla mężczyzny w wieku 50 lat, który chciałby schudnąć skupiając się na poniższych zagadnieniach? - jakich produktów ta osoba powinna koniecznie unikać w diecie? - jak powinna obliczyć swoje zapotrzebowanie kaloryczne? - jakie badania powinna wykonać, aby sprawdzić czy jego dieta jest prawidłowa? - czym są makroskładniki w diecie i jak wpływają na zdrowie?
You can add a remaining wait to the entire prompt, for example:
As you can see, this way you can create context, clarification, format, or style almost endlessly.
A good prompt is, above all, a consistent, precise prompt that contains full context , supplemented with the expected format. What errors do you need to watch out for?
If you ask for a post about travel, you will get a very vague article full of generalities . If you ask about preparing a travel plan for two adults for 3 days in Milan, supplemented with recommended places, necessary budget, define yourself as a fan of specialty cafes and local restaurants, then you can count on a better guide.
Avoid questions that contain any thoughts or suggestions , e.g. what do you think about all-season tires? The model does not think, and the question indicates subjective feelings.
If you build complex sentences that contain many commands the model can get lost. In this case either you will get the answer to some of the questions or ChatGPT will lose the context. If you notice such results you can split the prompts into separate ones and get better answers.
When building prompts, in addition to the rules I described above, it is worth considering some additional ones.
For convenience, any language can be used, but English often provides better answers. You can experiment here, and also tell ChatGPT what language you expect the answer to be in!
If you want more specific language and interesting answers, you can prompt for a role change. In our diet example, you could start with “Act as if you were a dietician helping me create a customized meal plan.” By adopting this persona, ChatGPT will provide diet advice, personalized meal plans, and motivation tailored to the context.
An interesting addition to the content is the instruction to prepare positive and negative elements of a problem. In this case, you will get better answers than the prompt about "comprehensive analysis".
You can tell the model to make sure it understands the context by prompting additional questions before providing a full answer, e.g. ask me a few questions first to clarify the content of the article
Most ChatGPT answers use direct but general language. If you indicate that you expect language tailored to your target audience, you will get a better result.
Multi-prompt is a technique of waiting until a few commands have been typed before responding. The model will wait for all the information before it acts.
After receiving the model's response, you can indicate which part of it you want to expand, correct, edit. You can also evaluate as negative or positive part of the statement.
Prompt crafting is a skill that develops over time. The more you work and train with the model, the better you will know it. In the whole process, it is worth remembering about the so-called fact-checking, because the model tends to deviate from the truth.
The more specific and clear the prompt, the better the result. Always remember the context, target group, form and style. Analyze the model's answers and evaluate what effect an additional command has. Use the role-play, scenarios and questioning functions to refine your prompts!
Want to learn AI and prompting? Sign up for the first CampusAI school in Poland!