General recommendations and detailed tips for chatbots and image generators.
What is a prompt
A prompt is a text request that you give to a neural network to get the desired result. In essence, it is like explaining what you want from a specialist when talking to them. The more specific and meaningful your request, the more accurate and useful the answer will be.
A prompt is based on three key elements.
- The goal is what you want to get as a result: an article, a drawing or an imitation of a photo, a recipe or a menu for a week.
- Context is any additional information: style, format. You can set it with a description or with specific examples.
- Restrictions are volume, tonality, a ban on some details.
An important nuance: the neural network does not guess thoughts, but analyzes specific words and descriptions from your prompt. If the entire query is “draw a forest,” the model will return something very basic, similar to the search results for the word “forest” in a photo stock. But if you specify that you need “a coniferous forest in the morning fog, painted with watercolors,” the result will be closer to the picture in your head. Prompt is your tool for managing the quality and detail of the answer.
What is the difference between a good and a bad prompt
The quality of the query directly affects the accuracy, completeness, and adequacy of the answer. A bad prompt gives vague, general, or irrelevant results, while a good one saves time and reduces the risk of hallucinations (that is, fictitious or simply incorrect facts).
The easiest way to understand the difference is with an example. Let’s say a neural network is required to write an article about ecology. The worst query in this situation would be:
Write an article about ecology.
It seems to be on point and you can’t find fault, but out of the three elements mentioned above, the prompt only has one: the goal. It lacks context and restrictions that would help to specify the request and get a more personalized result. At best, the chatbot will clarify the necessary details itself, at worst, it will give out a completely useless text, wasting your token limit and electricity to maintain the server.
The prompt above can be improved without resorting to any complex techniques and tricks:
Write an analytical article (500-600 words) about the state of the Amazon rainforest in 2025. The audience is ordinary readers, not environmentalists, who are concerned about the environmental situation and who may be potentially interested in volunteer programs.
It is important to indicate the main reasons for deforestation and statistics on forest loss over the past three years. Also give a couple of specific examples of successful restoration programs. Finish with practical recommendations for volunteers.
These corrections did not require delving into the mechanisms of how large language models work, or somehow specially adapting the text for the machine. It is enough to describe the task in natural language – about as to a classmate when working on a group project.
How to write a good prompt for a neural network
The word “prompt” is usually associated with chatbots, because they are the most accessible and understandable to use for people without much immersion in the topic of neural networks. But in general, prompts are needed for any neural network, they just work a little differently. For example, in image and video generators, some of the settings do not need to be written, but selected using buttons in the interface. Despite the differences, the action plan is always the same.
1. Define the goal
Formulate what the neural network should do: write an article, make a training plan, generate a picture or a set of vector icons.
If you need multiple things — like a home workout plan and a budget-friendly gym equipment selection — ask one query at a time to minimize confusion and make edits easier. In the workout example, it makes sense to first create a plan, edit it if necessary, and then choose the right dumbbells.
2. Set additional parameters
Anything that is important to consider when generating:
- Volume — number of words or characters, video length, number of slides in a presentation.
- Text or image style — strict professional tone, watercolor painting.
- Audience — experts in a specific field, students, complete beginners.
3. Add context and clarify requirements
Anything that will help the neural network better understand and implement your idea:
- Examples — fragments of texts that suit your style, famous artists or animation studios to focus on.
- Background information — dates, events, brand names or projects that may be important in the context of the topic or must be mentioned.
- Any prohibitions and restrictions – no more than 50 words per paragraph, no bullet lists, no decorations on the portrait.
4. Check and refine the prompt
If the neural network does not require a strict prompt format, the result of the previous steps can be inserted into the input window as is. For example:
Goal: make a menu for the week.
Format: table (days of the week × meals).
Audience: adult male, don’t eat meat, try to eat a varied and healthy diet, cook myself.
Context:
- Two meals a day – lunch and dinner.
- The menu should include dishes without meat, but with white fish and dairy products.
- Cooking time – no more than 30 minutes a day.
- There is a multicooker with a steamer mode.
- The food for dinner can be duplicated for lunch the next day.
- The budget is limited – the dishes should be made from ordinary, inexpensive products.
- Give preference to seasonal fruits and vegetables. I live in central Russia, it’s July now.
Restrictions:
- Do not use expensive products like quinoa, chia and truffle oil.
- No oven.
- Every day should be a different dinner, without repetitions.
It is not necessary to carefully format it, you can simply describe everything important in a couple of paragraphs of plain text. Send a request, look at the results and, if necessary, adjust the prompt. You will need to rewrite from scratch, only if the neural network did something completely wrong, but usually the result can be “finished off” within the same generation with clarifying messages: ask to correct the details that you do not like, or add something.
For example, in the example above, after generating the response, you can ask the neural network:
“Replace dinners on Thursday and Friday.”
“Add links to recipes for each dish.”
“Remove pumpkin and Brussels sprouts, add more seasonal vegetables: zucchini, eggplant, bell peppers, radishes, cauliflower.”
“Create a shopping list by days of the week with gram values of ingredients.”
A checklist for a good prompt
Before sending a request, run through this list – it will save you time, nerves and tokens.
- The goal is clearly formulated: it is clear what the output should be – text, image, script, table and so on.
- The context is specified: topic, style, audience, format, examples, sources or background information.
- There are restrictions: volume, structure, tonality, what to include or exclude.
- No vague words: instead of “make it beautiful” – a specific description of the desired result.
- A complex task is broken down into parts: a long request is divided into stages or clarified step by step.
- The model features are taken into account: text — role and language requirements, image — scene and details, video — movement, objects and atmosphere.
If everything from the list is done, most likely, the prompt will work well the first time or will require minimal edits.