Prompt

What it means

A prompt is whatever a person or a program gives an AI model to work from. It might be a question, an instruction, a few examples, or the start of a sentence for the model to finish. Google's machine learning glossary lists all of these as kinds of prompts.¹ Say a manager types: "Summarize this contract in five bullet points and flag any renewal dates." That sentence is the prompt, and the summary that comes back is the output.

The prompt is not always only what you typed. Software makers often write instructions of their own that sit behind yours. When an AI tool reads an email or a document for you, that text can be mixed into what the model works from. The U.S. National Institute of Standards and Technology (NIST) describes exactly this mix of trusted and untrusted text.² That detail matters more than it seems, as the story below shows.

Where the word comes from

Theater had prompts long before computers did. Dictionary.com defines a prompter as "a person who is offstage and follows a play in progress from the book, repeating missed cues and supplying actors with forgotten lines," and dates the word to 1400–50.³ Oxford Languages traces "prompt" through Old French to the Latin promptus, meaning "'brought to light', also 'prepared, ready'."⁴

The AI meaning grew in a few steps. In 2020, OpenAI researchers reported that their GPT-3 model could take on new tasks with "tasks and few-shot demonstrations specified purely via text interaction with the model," with no retraining.⁵ ("Few-shot" means giving the model a handful of examples.) In 2021, Pengfei Liu and colleagues surveyed this new approach and described how "prompt-based learning is based on language models that model the probability of text directly."⁶ By 2023, Oxford Languages had made "prompt" a finalist for its Word of the Year, noting that "As AI became more prevalent, so too did the need for human skills to moderate and direct these advances."⁴ So a word that once meant a line whispered from offstage became the main way people direct AI.

How others define it

• Oxford Languages: "An instruction given to an artificial intelligence program, algorithm, etc., which determines or influences the content it generates."⁴
• Google: "Any text entered as input to a large language model to condition the model to behave in a certain way."¹
• OpenAI, defining the craft: "Prompt engineering is the process of writing effective instructions for a model."⁷
• NIST, defining an attack on prompts: "An attack which exploits the concatenation of untrusted input with a prompt constructed by a higher-trust party such as the application designer."²

Where they agree and differ

All four agree that a prompt is input that steers what an AI system produces. They differ on two points.

The first is what counts as a prompt. Oxford calls it an instruction. Google counts any text at all, including examples and half-finished sentences.¹ The second is who writes it. The dictionaries picture one person typing. NIST describes layers: a prompt written by a trusted party, such as the company that built the tool, joined to text from somewhere less trusted.²

They also differ on how exact the craft is. Google calls prompt engineering "The art of creating prompts that elicit the desired responses from a large language model."¹ OpenAI calls it "a mix of art and science."⁷

The strongest objection to treating prompts as a leadership issue is that, for most people, a prompt is just a question typed into a chatbot. On that view, getting better answers is a personal skill, and that is how vendor guides teach it.⁷ For one person at one screen, that is fair. It stops being enough once AI tools read text written by strangers.

In the news

In early October 2026, the security company Barracuda described an email campaign aimed at two readers at once: the person and the AI assistant that summarizes their inbox. "Another layer carries hidden malicious instructions known as prompt injection," Barracuda wrote. "These are aimed at the AI assistant." It added that "None of these instructions are visible to the user."⁸ Prompt injection means slipping your own instructions into text an AI system will read. Infosecurity Magazine reported the research and noted that Barracuda did not say how widespread the campaign was.⁹

How sure we are: we are confident the campaign was reported as described, because we read Barracuda's own post. We do not know how common attacks like this are.

What it means for your organization

Every prompt is an instruction, so the real question is whose instruction it is. A model cannot reliably tell its owner's words from a stranger's when both arrive as text.² That is why this publication holds that an agent takes instructions only from its [principal](/definitions/principal), with [a person in the middle](/definitions/a-person-in-the-middle) of every chain of command.

Barracuda's advice follows the same logic. It recommends "Human approval requirements for financial transactions, vendor changes, and other sensitive decisions," and that "external content of any kind is always treated as data only and kept separate from instructions."⁸

Question to ask your team: When our AI tools read outside email or documents, which words do they treat as instructions, and which actions still need a person's approval?