Both ChatGPT and Bing are two extremely powerful language models of artificial intelligence, whose potential can be counted in billions of parameters and thousands of books from which models have gained their knowledge. However, despite this, sometimes the content generated by artificial intelligence is of at least dubious quality, if not simply wrong, too general or poorly formulated.

The key to getting what you are asking for is to skillfully formulate the queries themselves – prompts. In the face of artificial intelligence, which is to revolutionize not only the labour market but also the world as such, even a new field has been created: prompt engineering (also called prompt engineering), whose main area of interest is the development and optimization of queries sent to artificial intelligence models.

If there’s anyone who knows “how to talk” to generative AI, it’s prompt engineers. The process of becoming one of them is quite long, but to get the perfect what you expect from generative AI, you don’t need to spend a few years in college and have several years of work experience related to machine learning – just apply a few tips in practice.

Give your speech context

Although generative AI hallucinates and is still far from human potential, it shares one thing in common: it hates general questions. For example, imagine that you work in a flower shop. A customer comes to the florist and asks, “I want a bouquet.” What bouquet? Big, small? For what occasion? What flowers should it consist of? A bouquet may well be to put on the table or to give to the president as congratulations for 20 years of work. Now consider that the same customer comes and says, “I want to order a bouquet for my wife for their fifth wedding anniversary. My wife likes white flowers, but she doesn’t like roses. I want the bouquet to be big, but easy to hold in one hand.” Easier to do, right?

It’s the same with artificial intelligence: the more details you provide about your query, the more satisfying content will be generated. Asking ChatGPT “I want to buy a dog, what breed do you recommend”, the model will give the standard formula “There is no one breed that would be best for all people…” and lists the most popular breeds that you can read about on the first page of Google search results in Labrador Retriever, Golden Retriever or Border Collie. However, when I clarified my housing conditions and expectations for the dog, ChatGPT changed its tone to “Based on your preferences, lifestyle, and housing conditions, I can recommend several breeds of dogs that may fit your situation.” and listed less obvious but more suitable breeds for me, such as the Boston Terrier or Cairn Terrier.

Be precise

The precision of expression is directly related to the context. The more details the AI knows, the better it will be able to generate a response and match it to your expectations. As an example illustrating the precision of speech, I will again use the choice of dog breed.

I want to buy a dog. I live in an apartment in a block of flats with my wife and two children. We want a dog that is energetic but small. We are ready to devote time to care. What breed do you recommend?

– original prompt used in the previous query

However, this prompt can be clarified, because:

• The apartment can be located in the heart of a large city or on the outskirts of a small one, which affects the ability to provide the dog with movement
• There is a big difference between, for example, two children aged 2 and 6 and two children aged 16 and 17

I want to buy a dog. I live in an apartment in a block of flats on the outskirts of town, close to green areas, with my wife and two children aged 16 and 17. We want a dog that is energetic but small. We are ready to devote time to care. What breed do you recommend?

– new, clarified prompt

After clarifying the prompt, ChatGPT proposed four new breeds, which he had not mentioned before, and they fit into the previously stated requirements better than the breeds given with an unspecified prompt.

Use simple, understandable language

The language is extremely colourful, having many regionalisms, accretions from other languages, dialects or neologisms and linguistically inconsistent forms from everyday language. When talking to ChatGPT, Bing or any other artificial intelligence, simplify the language, and beware of complex art forms and vocabulary that could potentially be incomprehensible to artificial intelligence.

Play roles with AI

This tip is more about using generative AI as a creative text generator, but it can be successfully applied to an interesting conversation. Models such as ChatGPT are loaded to the brim with information from different sources and formulated in different ways. Starting from scientific publications written by experts in a given field, through plain language explanations of why an egg must be heat-treated before eating, downloaded from a forum about cooking.

Taking into account the multitude of topics, narratives and content, writing a prompt can direct artificial intelligence to the desired tone and way of speaking. For example, you will get completely different answers after typing in the prompt “Explain how the aurora borealis is formed” and after typing “Explain to a five-year-old child how the aurora borealis is formed”. You can also play roles the other way around, for example by changing the prompt “Explain why it is worth buying a smartwatch” to “Speaking from the perspective of a person actively practising sports, explain why it is worth buying a smartwatch”.

Outline the AI technical requirements of the text

Not once or twice it will happen that artificial intelligence will write too long or too short text. While in the case of Bing in Edge, the option to adjust the length of the text in the Compose tool does the trick, in ChatGPT and other popular generative models of artificial intelligence, the length of the generated text is a lottery – a lottery that can be won in advance.

When writing the prompt text, remember to take into account the properties of the desired content. You can do this by adding to the content of prompts such reservations as “The written text should be no longer than 10 sentences”, “The text should be exactly 200 words”, “Explain in 5 sentences …”, “Write a description of the photo on Instagram, which is between 300 and 400 characters. Include hashtags…” You can experiment not only with length but also with form, mentioning, for example, “List the reasons why […]. Write in a string, don’t use a bulleted list, and don’t number your arguments.”

Gently change the question, try again and use the conversation

Even if, after following all the tips, artificial intelligence generated an unsatisfactory text, it is worth not giving up and continue working with one prompt. Maybe you formulated your problem in a way that is incomprehensible to AI. Exchanging a few words can change miracles. Often, however, you do not even need to change the prompt – just ask to regenerate the answer.

It is also worth remembering that ChatGPT, but also other popular models (including GPT-4 or Bing chatbot) have a kind of “short-term memory”. As long as you don’t close the window or switch to a new conversation, you can refer to previous responses in subsequent prompts, such as “Rewrite the previous response” or “Summarize the previous answer in three sentences”.