Prompts, Skills, Commands, Agents: A Field Guide to the Jargon

Agentic AI has accumulated a vocabulary faster than most teams have had time to agree on definitions, and the same four words — prompt, command, skill, agent — get used loosely enough in casual conversation that people often mean subtly different things by them without realizing it. What follows is a short, plain-language field guide: a definition and a concrete example of each, so a team can use the same words to mean the same things.

Prompt

A prompt is a single request to the model, answered once, on the spot, with no persistence beyond that turn. "Summarize this paragraph" is a prompt. There's no packaging, no reuse mechanism, and no expectation the same wording will be issued again — although in practice, it very often is, which is exactly the repetition that motivates everything else on this list.

Command

A command is a direct, explicit instruction a person issues into an active session, matched deterministically by a parser to a fixed action — not inferred by the model. Typing /apply job-description.md is a command: the runtime recognizes the literal string and runs the same underlying directive every time, regardless of what the model "thinks" about the request. The defining trait of a command is that the model doesn't get a vote on whether it runs. This is worth distinguishing carefully from an agent action, which looks similar on the surface but is a model-inferred decision, chosen dynamically from context, with no explicit trigger string typed by a person.

Skill

A skill is a reusable, packaged set of procedural instructions and workflow guidance that an agent loads on demand to perform a recurring task consistently. A skill file sitting unused in a library costs nothing — no tokens, no attention — until a task actually calls for it, at which point its contents merge into the active context and the agent follows that guidance for the rest of the task. The key trait to remember is that skills stay dormant until invoked; that dormancy is what makes it practical to maintain a large library of specialized procedures without every one of them taxing every conversation.

Agent

An agent is an autonomous entity that manages its own perception, reasoning, planning, and tool execution to accomplish a goal, without step-by-step human direction on every action. The simplest way to hold this is as a repeating cycle — perceive, reason, plan, act — that runs every turn. An agent is goal-directed, tool-using, stateful within a session, and bounded: everything it does happens inside its allotted context window.

Keeping the four straight

A useful test, when a term's meaning is unclear in conversation, is to ask: is there an explicit trigger a person typed (command), is it dormant reference material invoked on demand (skill), is it a single unrepeated request (prompt), or is it a self-directed loop deciding its own next step (agent)? Most confusion in practice comes from conflating a command with an agent action — both can look, from the outside, like "the AI just did something" — but only one of them is deterministic. Getting this distinction right early saves a great deal of miscommunication later, particularly when a team is deciding where to place a human checkpoint: that decision depends entirely on whether the step in question is a fixed, predictable command or a model-inferred agent action.