Notes7 min read
Agents, skills, MCPs and hooks, explained with Figma
An agent does the work and a subagent takes one piece of it. A skill is a saved recipe it follows, a slash command is how you run one, an MCP server connects it to another app, a hook is a rule that always runs, and a plugin bundles them all to share.
From the course · lesson 0.8
Open any thread about Claude Code and you'll see the same seven words: agent, subagent, skill, slash command, MCP, hook, plugin. Each one names a simple idea, and the difference between an agent, a skill and an MCP is smaller than it looks. Most of them have a close match in Figma, so start there.
Agent vs skill vs MCP: seven words, seven Figma analogies
Agent: a collaborator in your file
An agent is an AI that takes a goal and works toward it in steps. It reads, decides, acts, and checks the result. Claude Code is the agent. Picture a teammate with edit access to your Figma file who can open any page, make changes, and tell you what they did. You set the brief and approve the work.
Subagent: a teammate with their own scratch file
A subagent is a helper the main agent hands one job to. It works in its own context window (its working memory for the conversation), with its own instructions and its own set of allowed tools. When it finishes, it sends back a summary instead of every step it took.
In Figma terms, you give a teammate one frame and a brief. They explore in their own scratch file and come back with the result, not their forty discarded attempts. That keeps your main file, and your main conversation, clean. Claude Code ships with a few built in, such as Explore for searching a project. You can write your own as small files in .claude/agents/.
Skill: a main component
A skill is a folder with a SKILL.md file in it: a name, a description of when to use it, and the steps to follow. Claude reads only the descriptions at first, then loads the full instructions when a task matches.
This is your main component. You build it once, and every instance follows it. Improve the main component and every use gets better. Here's how to write your first skill. A Kudos skill might be "write a reply thanking a customer for their testimonial, in our voice."
Slash command: Quick actions
A slash command is anything you run by typing / and a name, like /clear to start a fresh conversation. It works like Figma's Quick actions menu: one search box for everything you can do.
The two ideas now overlap. Anthropic's docs say custom commands have been merged into skills, so every skill you write can also be run as /its-name. A skill can be picked up by Claude on its own; a slash command is you asking for it by name.
MCP: a connection to another app
MCP stands for Model Context Protocol. It's an open standard for connecting AI tools to other apps and data. An MCP server is the connector for one app. Connect one and Claude can read from and act in that app: Notion, a database, Stripe, or Figma itself. Before you connect anything, read how to connect Claude to your tools safely.
The Figma match is an integration that pulls real content into your file, so you stop pasting lorem ipsum. Figma has its own MCP server. It gives agents your variables, components and layout data, and it can also create and edit frames on the canvas.
Hook: a constraint
A hook is a command that Claude Code runs automatically at a set moment: before a file is edited, after an edit, when Claude needs your input, when a session starts. The docs call this deterministic control. The action happens every time, without relying on the model to remember.
Think of constraints in Figma. Pin a layer to the right edge and it stays pinned every time you resize the frame. Nobody has to remember. An instruction in a skill is a request; a hook is a guarantee. Two common examples are a desktop notification when Claude is waiting on you, and a hook that blocks any edit to a file of secret keys.
Plugin: a published team library
A plugin is a package of skills, subagents, hooks and MCP servers that installs as one unit. Plugins are listed in marketplaces, which are catalogs you can browse with /plugin.
It's a published team library. Instead of rebuilding someone's components by hand, you enable their library and get everything at once. Install only plugins you trust: anything a plugin does on your computer, it does with your permissions.
How they fit together
The course groups these pieces into the Agent Stack: seven layers, each answering one plain question.
| Layer | The question | In Claude Code | In Figma terms |
|---|---|---|---|
| Context | What does Claude know? | CLAUDE.md, subagents | The cover page and brief |
| Connect | What can it reach? | MCP servers | Integrations that pull in real data |
| Codify | How does it do things my way? | Skills | Main components |
| Control | What must always or never happen? | Permissions and hooks | Constraints and edit access |
| Cadence | When does it run without me? | /loop, scheduled tasks, routines | (no real twin) |
| Crowd | Can it do 50 at once? | Subagents, workflows | A team working in parallel |
| Share | How does my team get it? | Plugins | A published library |
CLAUDE.md is a plain text file at the top of your project that Claude reads at the start of every session. Put your project's facts and rules there. Skills hold the procedures.
Here's how it looks in a made-up project called Kudos, a testimonial app. CLAUDE.md says what Kudos is and who it's for. An MCP server connects Claude to the database where testimonials are stored. A skill turns a raw customer message into a short pull quote, using only the customer's own words. A hook blocks any edit to the file that holds API keys. A subagent reviews the finished page with fresh eyes. A plugin is how you'd hand the whole setup to a co-founder.
Which to learn first
Anthropic's docs make a sensible point: you don't need all of this on day one. Add each piece when a specific problem shows up.
- Context first. When Claude gets a convention wrong twice, add it to
CLAUDE.md. - Then one skill. When you've pasted the same steps into chat for the third time, save them as a skill.
- Then one connection. When you keep copying data from a browser tab Claude can't see, connect that app with an MCP server.
- Hooks when something must always (or never) happen. If a rule can't depend on Claude remembering it, make it a hook.
- Subagents and plugins later. Use subagents when side tasks clutter your conversation. Make a plugin when a second project needs the same setup.
Try it: a five-minute tour
Open any project in Claude Code and run these one at a time:
- Type
/and scroll. That's every command and skill available to you right now. /skillslists your skills and where each one comes from./mcpshows connected MCP servers and whether they're working./hooksshows any hooks that are set up. Expect it to be empty at first./pluginopens the plugin browser. Look through it, but don't install anything yet.- Ask in plain words:
What subagents can you use in this project?
To see these pieces working together day to day, read directing a crew of agents. Foundations in Chief of Agents covers these pieces in this order, then has you write your first skill and connect your first app.
Sources
- Extend Claude Code (features overview), Claude Code docs
- Subagents, Claude Code docs
- Extend Claude with skills, Claude Code docs
- Commands reference, Claude Code docs
- Connect Claude Code to tools via MCP, Claude Code docs
- Automate actions with hooks, Claude Code docs
- Plugins overview, Claude Code docs
- How Claude remembers your project (CLAUDE.md), Claude Code docs
- Skills explained, Claude blog
- Steering Claude Code: when to use CLAUDE.md, skills, hooks, and subagents, Claude blog
- Figma MCP server, Figma developer docs
Learn this inside Claude Code.
Join the waitlist for Chief of Agents. You'll get the free beginner's handbook and skills right away, and hear first when the course opens.
Keep reading
6 min read
Claude Code for designers: what it is, and why it isn't just for engineers
What a coding agent does, why directing one is a design skill, and a safe first hour for designers new to Claude Code.
7 min read
Connecting Claude to your tools safely: MCP for non-developers
Is it safe to connect Claude to Gmail or Notion? What an MCP server can see and do, why to start read-only, and which hooks help.
6 min read
From designer to builder: directing a crew of agents at work
What working with AI agents looks like day to day after ten years in design: the work of direction, and what stayed the same.