Prompt, skill, command, hook, agent, loop. Six terms that constantly get mixed up in day-to-day AI work. Yet they describe something concrete: how much work you hand off to the AI.

Level 1 does everything by hand. Level 6 runs almost on its own. Each step up takes over something you used to type or trigger yourself. These aren't competing tools — you combine them. But once you know the difference, you stop operating a 6-level instrument like a pocket calculator.

---

1) Prompts — the base unit

Plain text input to the model. One instruction or question, processed once. No memory, no automation: input in, output out.

Example: „Summarize these three customer emails in one sentence each." You type, the AI answers, done.

When to use it: For anything one-off. A quick question, a draft, a rewrite. If you only have a task once, the prompt is the right and only tool. The moment you type the same instruction for the third time, you're on the wrong level — keep reading.

2) Skills — stored knowledge

Predefined folders with best practices and instructions for a specific type of task. The AI loads them automatically when the context fits. Reusable knowledge you don't have to paste into every prompt.

Example: A team sets up a blogpost skill: house format, tone, structure, SEO rules, banned phrases. After that, „write a blog post about topic X" is enough — the format gets applied automatically, without anyone explaining it. Nobody has to paste the style guide again.

When to use it: When there's a right way to do a type of task that you don't want to describe from scratch every time. Skills are your saved standard — defined once, always present.

3) Commands — stored routines

Custom shortcuts, usually as a slash command (/command). You define a routine in a file once, then call it by a short keyword. Like a saved prompt with a fixed structure.

Example: A /standup that builds a standup update from the latest git log automatically. Or a /release-notes that writes clean notes from the commits since the last release. Instead of typing the instruction every time, the slash command does it.

When to use it: For recurring routines that you trigger. The difference from a skill: a skill is knowledge that kicks in behind the scenes; a command is an action you start deliberately.

4) Hooks — automatic triggers

Triggers that fire on certain events, without you typing anything. For example before or after a tool call. They run deterministically in the background — no „please", no typing.

Example: A hook that runs the formatter and linter automatically after every file edit. You save, the hook checks and formats right after — without your involvement. Or a hook that blocks every commit if an API key shows up in the code.

When to use it: When something should always happen, reliably and without exception. Hooks are where you offload discipline from your head into the pipeline. Whatever a hook handles, you never forget again.

5) Agents — delegated goals

Independent instances that work through a multi-step task on their own, use tools, and make decisions along the way. You delegate a goal, the agent works it off. Far more autonomy than a single prompt.

Example: A research agent gets a topic, searches for sources itself, reads them, reconciles contradictions, and returns a finished briefing. Or a pipeline setup where several agents collaborate — one researches, one writes, one edits — and a finished piece of content comes out the end.

When to use it: When the task has several steps, you know the what but don't want to dictate each individual step. You give the goal and the boundaries, not the click-by-click instructions.

6) Loops — the iteration principle

Repeated passes: the AI works a task, checks the result, fixes it — until a goal is met. Act, observe, correct, repeat. This is the mechanic that makes agents useful in the first place.

Example: A task with a clear success criterion — say „all tests must pass". The AI writes code, runs the tests, sees a failure, corrects it, runs again. That loop keeps turning until the criterion is met. No human clicks between rounds.

When to use it: When you have a measurable goal and the path there takes trial and error. The loop replaces the „check it and try again" you'd otherwise trigger by hand. One caveat: it needs a clear stop condition — otherwise it spins forever.

The map

Level	What it is	Who triggers it	What you hand off …
1 — Prompt	Single input	You, manually	nothing — you do it all
2 — Skill	Stored knowledge	AI, context-driven	repeating your standards
3 — Command	Stored routine	You, by keyword	typing fixed routines
4 — Hook	Automatic trigger	Event, deterministic	remembering to do it
5 — Agent	Delegated goal	You, then autonomous	the individual steps
6 — Loop	Iteration to a goal	AI, self-correcting	checking and fixing

In short: a prompt is the base unit. Skills and commands are stored, reusable building blocks — one as knowledge, the other as action. Hooks automate triggers. Agents delegate whole goals. Loops are the principle that makes agents reliable.