What Is a Prompt? How to Write AI Instructions That Work

👁 120 reads

A prompt is the instruction or input material you give an AI model to get a response or complete a task. It can contain a question, instruction, example, and data - and in tools that support multiple media types, it can also include images or audio.

A prompt tells the model what to do with the material

When you type “explain photosynthesis,” you’re providing a prompt. When you add “for an elementary school student, in one paragraph,” you’re specifying the audience and format. If you paste text and ask to shorten it without changing the facts, the source material also becomes part of the input. The OpenAI help documentation on prompts describes this definition covering text and other input types.

The most important thing is separating what needs to be done from the material that needs to be processed. A pasted email shouldn’t accidentally replace your instruction.

A good prompt defines the task and success criteria

You don’t need to start every conversation with an elaborate expert role. For simple tasks, a concrete request is enough. For harder ones, add context, data source, format, and constraints. OpenAI’s recommendations emphasize clarity, appropriate task scope, and refining instructions based on the results you get.

You can try this in free ChatGPT or Google Gemini. The same prompt doesn’t guarantee identical results. Models change, as do settings, tools, and conversation context. That’s why for comparison, you should save the exact input text and response.

An example with a note gives you a checkable answer

I prepared my own editorial example, not a model test transcript. The material reads: “Anna will send the proposal on Wednesday. Mark will prepare photos. The client will approve the project on Friday.” There are three actions here, but only two have specified deadlines. This distinction lets you immediately notice whether the response filled in missing information.

Extract actions, people, and deadlines from the note.
Return a table with columns: action, person, deadline.
Use only the note. Mark missing data as "not specified."
Do not convert days of the week to calendar dates.
NOTE
Anna will send the proposal on Wednesday. Mark will prepare photos.
The client will approve the project on Friday.
END

The correct answer has three rows. For Mark, the deadline should remain unspecified. If the model enters Thursday, even a very neat table is wrong. I also counted data completeness: deadlines are specified for 2 out of 3 actions, or about 66.7%. This isn’t AI accuracy - it’s a property of this particular note.

Image prompts require specifying what should remain

In a graphics task, “improve the photo” leaves open questions about framing, colors, and scope of change. It’s better to name the element to edit and the features you want to preserve. Example: “Change the background to a solid light wall. Keep the framing, clothing, facial proportions, and light direction.” This is an instruction to test in a specific tool, not a guarantee of faithful editing.

Compare the result with the original, especially in areas the model wasn’t supposed to change. It’s useful to check text, hands, object edges, and product details. In the guide on how to edit photos with AI, we expand on working with images. Don’t assume that one generator’s rules work the same way in another.

Where even detailed prompts fail

Instructions don’t supply missing knowledge. If you ask about current pricing without a source and a tool to check it, the model may generate a credible-looking number. The instruction “answer only truthfully” isn’t proof. Provide a document, point to a current page, or ask it to mark things that can’t be determined.

Too many requirements can also create contradictions. “Keep all details” and “fit everything in one short sentence” can’t always be fulfilled simultaneously. Decide what takes priority. When editing, I rank preserving facts higher than perfect length, and for automatic imports I require correct format first and explicit problem reporting.

In an application that performs actions, the instruction alone doesn’t replace access control. Document content may contain an instruction you didn’t intend to give. Separating material from task makes work easier, but doesn’t provide a technical security guarantee. For important operations, you need tool restrictions and result verification before execution.

Fix one error at a time and save successful versions

Start with a short instruction, and after the response, name the specific error: an added deadline, omitted condition, or wrong format. Change the instruction to address that problem and repeat the test on the same material. Then check it on another example so you don’t fit the solution to just one note. Save the successful prompt along with acceptance criteria, not as a magic formula that works without verification.

From definition to practice

This article explains the concept itself. Once you know what a prompt is, the next step is learning to write them. We’ve broken that down element by element in the guide on how to write prompts.

// read moreHow to Write Suno Prompts: Six Elements and Four MistakesGuidesPrompt Engineering: Techniques, Examples & Test Results (2026)GuidesSmall Language Models Learn from Large Ones via Retrieved InstructionsBenchmarks

Kinetyra