What Is a Prompt?
Author: Domeny.ForSale editorial team Published
In brief: A prompt is the input you give an artificial intelligence model to get a response or assign a task. It can be a question, an instruction or a combination of instructions and material to analyse. A good prompt clearly defines the expected result, but does not guarantee that the AI’s response will be true or correct.
What does a prompt mean in practice?
In a conversation with generative AI, meaning artificial intelligence that creates things such as text, images or code, a prompt could be a sentence like: “Explain how an internet domain works.” A more detailed prompt may include a description of the situation, data, constraints and an example of the expected response. It does not have to be written in a special language or use technical syntax.
Depending on the tool’s capabilities, you can attach a photo, document or recording to your instructions. When you upload a photograph of a plant and ask about visible symptoms, the model receives both the question and the material needed to answer it. The text “What do you see?” on its own, without an attachment, does not give it the same information.
Outside the context of AI, the word “prompt” can also mean a command prompt in a terminal: a signal that the program is waiting for you to enter a command. In discussions about chatbots, however, it usually means an instruction or other input for the model.
What makes up a useful instruction?
The simplest prompt can be a single sentence. Expanding it makes sense when the extra information helps the model identify the task rather than forcing it to guess. Five elements are most often useful:
- Goal: what should be produced, such as a comparison, summary, message or piece of code.
- Context: who the response is for and the situation in which it will be used.
- Input material: the text, data or description on which the model should base its response.
- Constraints: what not to add, what to preserve and what scope to adopt.
- Output format: for example, three paragraphs, a comparison table or a list of steps.
Not every task requires all of these elements. The question “How many centimetres are in a metre?” does not need a description of the audience or a specified style. When preparing a proposal for a client, however, missing information about the service, price or deadline can lead to an unhelpful response.
You can also specify a role, for example: “Assess the text from an editor’s perspective.” This instruction helps define the approach to the analysis, but does not give the model any real qualifications. The instruction “You are a lawyer” does not turn the response into professional legal advice.
Example: from a general request to a specific task
Suppose you run a shop and want to inform a customer about a delayed shipment. A general prompt might read:
Write an email to a customer about a delay.
The model does not know how long the delay is, what can be promised or how to sign the message. It may therefore produce a vague text or fill in the gaps with unconfirmed details. A more useful instruction looks like this:
Write a short message to a shop customer about a delay in dispatching order no. 123. The planned dispatch day is now Thursday instead of Tuesday. Apologise for the inconvenience and mention that we will send dispatch confirmation separately. Do not give a delivery date or promise a discount. Use a calm, polite tone and no more than 100 words. Sign off as: The Shop Team.
The second version is not better because it is longer. It is better because it contains the information needed to complete a specific task and limits unwanted additions.
If you have an example of a message that matches your company’s style, you can include it. Make clear, however, that it should serve as a model for the tone, not as a source of data about the new order. This allows you to clearly separate instructions from supporting material.
How can you improve a response without rewriting everything?
Working with a prompt often resembles editing a text over several rounds. Within the same conversation, you can refer to the previous response and specify a particular change: “Shorten the introduction”, “Remove specialist terminology” or “Keep the dates, but make the tone less formal”.
Instead of writing “This is wrong”, identify the problem and the expected result. For example: “The message suggests guaranteed delivery on Friday, but we only know the dispatch date. Remove that promise.” This gives the model a clear criterion for the revision.
For a larger task, you can first ask for a plan, approve the scope and only then ask it to develop individual sections. If key data is missing, add: “Before preparing your response, ask questions about the missing information.” Check, however, that the model actually waited for answers rather than making its own assumptions.
What can even a good prompt not guarantee?
A precise instruction does not eliminate the risk of hallucinations: generated information that sounds credible but is untrue. Asking for sources is not enough either: the publications, quotations and web addresses provided need to be checked. The instruction “use up-to-date data” does not, by itself, give the tool access to the internet.
Do not paste passwords, access keys or customer data into a prompt without authorisation and without checking the service’s policies. How conversations are stored and used depends on the tool, its settings and the terms of the agreement. For practice, replace information that identifies a person with fictional data.
To get started, choose one small task and write down three things: what should be produced, what information it should be based on and what the model should not add. Once you receive the result, check it against the same criteria before using it in practice.
See also: What Is a CDN?
Read also
