Prompt

What is a prompt?

A prompt is the input given to a [large language model](/glossary/large-language-model/) to influence what it produces. It may contain a question, task, instructions, examples, reference material, desired format, or constraints. In an AI-agent system, prompts can also include tool descriptions, retrieved information, current state, and observations from earlier steps.

A prompt is therefore more than the short message typed by a user. Applications often assemble several pieces into the final model input.

What a prompt can contain

A useful prompt may include:

• the outcome the model should help achieve;
• relevant facts or documents;
• rules about tone, format, or allowed behavior;
• examples of acceptable inputs and outputs;
• tool definitions available to the model;
• the current task state and recent tool results.

For example, a support prompt might ask the model to summarize a case, include the customer’s message and order status, require a three-bullet output, and prohibit guessing missing facts.

Prompt versus system prompt and context

A prompt is the overall input used to guide a model response. A system prompt is a higher-priority part of that input supplied by the application to establish behavior and boundaries. A user prompt expresses the user’s current request.

[Context](/glossary/context/) is broader than the prompt wording. It means the information available to the model for the current inference. The prompt is one way that context is packaged and presented.

[Agent instructions](/glossary/agent-instructions/) are also a conceptual layer. They define how the agent should operate and may be delivered through a system prompt, configuration, or application logic. Critical restrictions should not depend on prompt text alone.

What makes a prompt useful?

Good prompts reduce ambiguity without burying the important objective. They clearly separate instructions from reference data, state what a successful output should contain, and identify relevant constraints. Examples help when a format or judgment standard is difficult to describe.

Longer is not automatically better. Unnecessary text consumes tokens, can distract the model, and may create conflicting instructions. Prompt quality also cannot compensate for missing data, weak tools, or unclear goals.

Why prompts matter in agents

An agent may construct a new prompt on every loop. It selects relevant memory, state, observations, and tools so the model can decide the next step. Poor context assembly can make an otherwise capable model forget constraints or act on stale information.

Builders should treat prompts as testable system components: version them, evaluate representative cases, make instruction priority clear, and keep untrusted content separate from trusted rules. See [Build Your First AI Agent](/build-your-first-ai-agent/) for a practical implementation path.