A prompt is an input that directs an agent’s behavior by specifying what to do, how to do it, or what constraints to observe. In a language-based system, the prompt is the primary interface between a principal and an agent — it is how the principal communicates intent.

The word comes from theater. A prompter sits offstage and feeds lines to actors who have forgotten them. The prompter does not perform; the prompter tells the performer what to say next. The theatrical sense carries into computing: a shell prompt ($, >) tells the user the system is waiting for input. The command prompt is not itself a command — it is the invitation to give one.

In artificial intelligence, prompting became a central technique with the development of large language models. Rather than training a model for each specific task, researchers discovered that providing the right input text could steer a general model toward specific behavior. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou demonstrated in “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models” (2022) that including reasoning steps in a prompt caused models to produce reasoning steps in their output. The prompt shapes the response not by changing the model but by changing what the model conditions on.

A prompt has three components. Instructions specify what the agent should do — the task, the format, the goal. Context provides information the agent needs — background, examples, relevant state. Constraints specify what the agent must not do — boundaries, safety requirements, scope limits. A well-constructed prompt balances all three: enough instruction to direct, enough context to inform, enough constraint to bound.

In agentic systems, prompts compose. A system prompt sets baseline behavior for all interactions. Rules inject standing constraints. Skills inject domain-specific knowledge when activated. A user prompt specifies the immediate task. The agent’s effective prompt is the composition of all these layers — system, rules, skills, and user input merged into a single context that governs the agent’s next action. This layered composition is why agentic systems can be customized without modifying the agent itself: each layer adds to the prompt without replacing what came before.