A prompt is the natural-language input given to a generative AI model, such as a Large language model or an image generator, to elicit a desired output. Prompts range from a single short instruction to elaborate, multi-part specifications that include examples, constraints, and formatting requirements.
Types of prompts
Prompts take several distinct forms depending on the interaction. A System prompt sets persistent instructions and persona for an entire conversation, typically hidden from the end user, while the ordinary conversational input is often called a user prompt. In image and video generation, a Negative prompt specifies what should be excluded from the output, separately from the main descriptive prompt. Prompts that include worked examples of the desired input-output behavior are described as Few-shot learning prompts, in contrast to zero-shot prompts that provide only an instruction.
Prompt engineering
The practice of deliberately crafting prompts to reliably produce better outputs is known as Prompt engineering. Techniques such as Chain-of-thought prompting, which asks a model to reason step by step before answering, emerged from empirical observation that the phrasing and structure of a prompt can substantially change a model's output quality, even when the underlying model is unchanged. This sensitivity is closely related to In-context learning, the ability of language models to adapt their behavior based on information present in the prompt itself, without any update to their underlying weights.
Risks
Because prompts are the primary interface through which users and, increasingly, external data control a model's behavior, they are also a significant attack surface. A Jailbreak (AI) is a prompt crafted to bypass a model's safety training, while Prompt injection refers to malicious instructions hidden in content that a model processes, such as a web page or document, which can hijack an AI system's behavior without the user's knowledge. Prompt injection has become a particularly serious concern as AI agent systems are given the ability to take real-world actions, since a successful injection can cause an agent to misuse its tools rather than simply produce bad text.
Significance
The centrality of prompts to how generative AI systems are used has given rise to informal communities and marketplaces devoted to sharing, refining, and trading effective prompts, alongside academic interest in understanding why particular phrasings succeed or fail. Some researchers argue that as models become more capable and better at inferring intent, the value of manual prompt crafting will diminish relative to broader practices sometimes described as Context engineering, which concerns curating the fuller set of information, tools, and history a model has access to, not just the immediate instruction.