Claude Code skills vs prompts, commands and plugins

Claude Code has several ways to give it instructions, and the names overlap. This page puts them side by side. The definitions come from Anthropic’s Claude Code docs, linked in each row and again at the foot of the page. For the idea of a skill on its own, read What are agent skills?. For the habits that make a set of them easier to live with, see Claude Code skills best practices.

The six, side by side

Skill or prompt

A prompt is a description you type into the chat. A skill is a file Claude Code keeps. Anthropic’s skills page says to create one when you keep pasting the same instructions, checklist or multi-step procedure into chat. That is the whole test: if you have pasted it twice, it is a skill waiting to be written.

A skill can hold more than text. It is a folder, so it can carry scripts, templates and examples beside SKILL.md. What are agent skills? shows the layout.

Skill or slash command

You do not have to choose. The skills page says custom commands have been merged into skills: a command file and a skill folder of the same name both give you /name and work the same way. Skills add three things: a directory for supporting files, frontmatter that controls whether you or Claude invokes them, and the ability for Claude to load them automatically when relevant.

That frontmatter is where the old command behaviour lives. The docs say disable-model-invocation: true stops Claude loading a skill by itself, for workflows you want to trigger manually with /name. And user-invocable: false is for a skill only Claude should invoke.

Skill or subagent

A skill is instructions. A subagent is a worker: its own context window, system prompt, tools and permissions. They can meet. The skills page says setting context: fork in a skill runs it in a forked subagent context. A plugin can also ship subagent definitions that Claude can delegate to.

Skill or CLAUDE.md

Both are text Claude reads, and they differ in when. The memory page says CLAUDE.md is loaded at the start of every conversation. The skills page says a skill’s body loads only when it is used, so long reference material costs almost nothing until you need it. It suggests a skill when a section of CLAUDE.md has grown into a procedure rather than a fact. Facts that should always be in view go in CLAUDE.md. Procedures go in skills.

Skill or hook

This is the choice between asking and making. A skill is read by the model, which then decides what to do. A hook is a command Claude Code runs itself, so the action always happens. The memory page makes the same point from the other side: CLAUDE.md is context, not enforced configuration, and to block an action regardless of what Claude decides, use a PreToolUse hook. If a rule must hold every time, it belongs in a hook, not in a skill or in CLAUDE.md.

Skill or plugin

A plugin is not a rival to a skill. It is a way to carry one. The plugins page says skills, subagents, hooks and MCP servers all work on their own, and that a skill you save in ~/.claude/skills/ is available in every project on your machine. Use a plugin when you want several of them packaged as one unit, to share with teammates, install in many projects or publish versioned releases. One cost to know: an enabled plugin is part of every session, and for each of its skills that Claude can invoke on its own, the name and description sit in Claude’s context on every turn, while the full text loads only when used. Where skills live lists the folders.

A worked example: one hero, as a prompt and as a skill

Say you want an animated website hero. MotionSites sells prompts for this, to paste into an AI builder, and the result depends on the tool, the model and the run. A prompt you might type is:

Build a retro 1950s diner hero for my ordering app. A waitress on roller skates behind a big headline. Add a link bar, two buttons and a ticker.

You type that again next time, and each run can come out differently. Now the same job as a skill. Roller Diner Hero is a free skill whose SKILL.md header holds a name and a long description saying what it builds and when to use it. That description is what Claude Code matches your request against, and it names retro, diner, restaurant and hospitality heroes. The body then gives Claude the steps:

1. Collect the words from you: a name, a short headline, two sentences, buttons, links and ticker lines. The skill says every number and claim must come from you.
2. Write hero.json by copying the example and replacing every value.
3. Build with node <skill>/scripts/build.mjs hero.json out, which lists every problem at once and writes nothing until they are fixed.
4. Check with node <skill>/scripts/check.mjs out/index.html.

The footage, colours, fonts and layout stay as they are; you change the words and links. The design lives in the skill’s template, which is why the second run looks like the first, and the length limits are enforced by the build, not left to the model. The skill also writes a paste-in embed and a 1080p MP4 with sound, which a pasted prompt does not carry with it. Its limits are real too: it needs Node.js 18 or newer, Chrome, Chromium or Edge, and ffmpeg for the video.

Nothing about this needs a plugin or a hook. If you wanted the build to run every time you finished editing hero.json, that is the hook’s job. If you wanted to hand the skill to a team, a plugin or a committed .claude/skills/ folder does it.

Free

web

Roller Diner HeroClaude CodeCodexCursor+3 more agents

web

Cloud Sky HeroClaude CodeCodexCursor+3 more agents

web

Cursor Tour HeroClaude CodeCodexCursor+3 more agents

To judge descriptions like the one above, see Write a description that makes a skill trigger. To compare more hero skills, see the best hero animations for Claude Code.

Which one to reach for