Research
What Is a Prompt? How AI Prompts Work in 2026
A prompt is the instruction you give an AI model to tell it what to do. Here is what a prompt actually is in 2026, the parts that make one work, and the prompting techniques worth knowing.
A prompt is the input you give an AI model to tell it what to do. It can be a question, command, statement, code, or even an image or audio clip, and it is the only thing conveying your intent to the model. Clearer, more specific prompts produce more useful output.
By 2026, talking to an AI model has become an everyday act. OpenAI's CEO said in October 2025 that more than 800 million people use ChatGPT every week, and tools like Google's Gemini and Anthropic's Claude reach hundreds of millions more. Every one of those interactions begins the same way: with a prompt. Yet most people use prompts without ever stopping to ask what one actually is or why the wording matters so much. This guide answers that plainly and vendor-neutrally.
What is a prompt?
A prompt is the input submitted to a generative AI model to guide it in producing a response. In practical terms, it is whatever you type or attach to tell the model your goal. As TechTarget defines it, an AI prompt is "the input submitted to a large language model (LLM) via a generative artificial intelligence (GenAI) platform" — the foundation the model uses to generate its answer. A prompt can be a single word, a full paragraph, a code snippet, a question, or a command. It is not a setting or a button; it is plain language, and it is the entire bridge between what a human wants and what the model does. Because the model has no other window into your intent, the prompt carries all of the meaning. Vague in, vague out.
What are the parts of a good prompt?
A throwaway prompt can be one line, but the prompts that produce reliable results tend to share the same building blocks. Naming each part deliberately is the single biggest lever most people have over output quality.
| Component | What it does | Example fragment |
|---|---|---|
| Instruction | States the task and the goal | "Summarize the email below" |
| Context | Gives background, audience, or tone | "for a busy executive" |
| Input data | The material the model works on | The pasted email text |
| Output format | Specifies structure, length, or style | "in two bullet points" |
You rarely need all four for trivial questions, but for anything that matters, stacking instruction plus context plus input plus a format specification turns a generic answer into a precise one. Amazon Web Services frames the discipline of doing this well as prompt engineering: writing, refining, and optimizing inputs to encourage a model to produce specific, high-quality outputs. A prompt is what you write; prompt engineering is the practice of getting that writing right on purpose.
System prompts vs. user prompts: what is the difference?
Not every prompt comes from you in the moment. Modern chat systems separate two kinds. A system prompt sets the model's persistent behavior — its role, rules, and tone — and stays constant across the conversation. A user prompt is the specific request you send each turn. A developer might set a system prompt of "You are a concise legal assistant; never give legal advice," while the user prompt is simply "Explain what an NDA is." Keeping them separate reduces conflicting instructions and makes AI applications easier to maintain. In most consumer apps the system prompt is hidden, so when you type into a chatbot you are almost always writing only the user prompt — which is why the same product can behave very differently inside two companies that wrote different system prompts.
Can a prompt include images or audio?
For most of the history of chatbots, a prompt meant text. That is no longer the boundary. Frontier models in 2026 are multimodal: a single prompt can mix text with images, PDFs, audio, or video. You can attach a photograph and ask what is in it, paste a chart and ask the model to read the trend, or hand it a multi-page contract and ask for the key risks. Models differ in their inputs — some accept native video, others specialize in document and PDF reasoning — but the principle is constant: the text portion of the prompt still steers the interaction, telling the model what to do with the attached media. Multimodal prompting effectively makes the prompt the model's perception layer, not just a text box, and it widens what counts as a prompt in the first place.
What are the main types of prompting techniques?
Once you can write a clear prompt, the next question is how to structure it for harder tasks. Three techniques cover most of what everyday users and builders need, and they trade effort for control and accuracy.
| Technique | Examples provided | Best for |
|---|---|---|
| Zero-shot | None | Simple, well-defined tasks |
| Few-shot | A few worked examples (often 2-5) | Consistent, structured output |
| Chain-of-thought | A request to reason step by step | Math, logic, multi-step problems |
Zero-shot prompting asks the model to do something with no examples, leaning on what it learned in training; it is fast but can be generic. Few-shot prompting includes a small set of worked examples — often two to five — so the model recognizes the pattern you want; a systematic survey of prompting techniques catalogs it as one of the most widely used methods for steering output toward a consistent format. Chain-of-thought prompting asks the model to work through its reasoning step by step before answering, which markedly improves accuracy on arithmetic, logic, and other multi-step problems; the simplest version is appending "let's think step by step" to a request. The practical rule is to reach for the lightest technique that reliably gets the result you need, and only add examples or reasoning steps when a plain prompt falls short.
Why prompts matter
A prompt is a small thing with outsized leverage. The model does not read your mind; it reads your prompt, and the difference between a vague request and a well-structured one is often the difference between a useless answer and an excellent one — at no extra cost and no extra model. As AI moves deeper into daily work in 2026, writing a clear prompt has quietly become a core literacy, on par with knowing how to phrase a good search query a generation ago. You do not need to be an engineer to benefit. State the task, give the context, supply the input, and say what you want back, and most of the value of these systems is already within reach.
Frequently asked
What is a prompt in simple terms?
A prompt is the input you give an AI model to tell it what you want it to do. It can be a question, a command, a statement, a code sample, or even an image or audio clip, and it can run from a few words to several paragraphs. When you type "summarize this email in two sentences" into ChatGPT, that whole instruction is the prompt. The model reads it and generates a response based on what the prompt asks for. The clearer and more specific the prompt, the more useful the output, because the prompt is the only thing telling the model your intent. In short, a prompt is how a human conveys a goal to a generative AI system in plain language.
What is the difference between a prompt and prompt engineering?
A prompt is the actual input you send to a model. Prompt engineering is the practice of designing, testing, and refining those inputs to get reliably better results. Writing one quick question is prompting; systematically tuning the wording, structure, examples, and constraints until the model performs a task consistently is prompt engineering. The distinction matters because the same request can produce very different output depending on how it is phrased. Prompt engineering treats the prompt as something to optimize, not just type, using techniques like adding context, giving examples, specifying the output format, and breaking complex requests into reasoning steps. For everyday use you only need good prompts; building dependable AI applications usually requires real prompt engineering.
What are the parts of a good prompt?
A strong prompt usually combines four ingredients. First, a clear instruction stating the task and the goal, such as "write," "summarize," or "classify." Second, context, the background the model needs to interpret the request correctly, including the audience, the tone, or relevant facts. Third, the input data the model should work on, like the text to rewrite or the numbers to analyze. Fourth, an output specification telling the model the format, length, or structure you want back. You do not need every part for simple tasks, but for anything important, naming the goal, supplying context, providing the input, and stating the desired format dramatically improves accuracy and cuts down on vague or off-target answers.
What is the difference between a system prompt and a user prompt?
A system prompt sets the model's overall behavior and persists across a conversation. A user prompt is the specific request you type in each turn. Developers use the system prompt to define the AI's role, rules, and tone, like "You are a careful financial assistant; never give investment advice." The user prompt is everything that should change per request, such as "Explain what a 401(k) is." Keeping the two separate reduces conflicting instructions and makes applications easier to maintain. In chat APIs these arrive as distinct message roles. Most consumer chat tools hide the system prompt, so when you type into ChatGPT you are usually writing only the user prompt.
Can a prompt include images or audio, not just text?
Yes. While most prompting is still text, frontier models in 2026 are multimodal, meaning a single prompt can combine text with images, PDFs, audio, or even video. You can hand a model a photo and ask what is in it, paste a chart and ask it to read the trend, or attach a contract and ask for a summary. Models differ in what they accept: some take native video, others specialize in document and PDF reasoning. The text part of the prompt still does the steering, telling the model what to do with the attached media. Multimodal prompting effectively turns the prompt into the model's perception layer, not just a text box.
What are zero-shot, few-shot, and chain-of-thought prompts?
These are three common prompting techniques that trade effort for control. A zero-shot prompt asks the model to do a task with no examples, relying on what it already learned during training; it is fast but can be generic. A few-shot prompt includes a small number of worked examples, typically two to five, so the model recognizes the pattern you want and produces more consistent output. A chain-of-thought prompt asks the model to reason step by step before answering, which sharply improves accuracy on math, logic, and other multi-step problems. A simple version is adding "let's think step by step" to a request. Pick the lightest technique that reliably gets the result you need.