Skills and plugins in Claude Code: what they are, how to install them and when to build your own

Short answer: a skill is a file of instructions that Claude Code loads only when a task calls for it, and a plugin is a package that installs several skills, subagents and connections at once. To install a plugin, type /plugin in a Claude Code session and pick it from the catalog.

People who start using Claude Code for brand and marketing work tend to go through the same phase: pasting the same instructions into every conversation. The post checklist, the way to put together a brief, the review rules. Skills and plugins exist to take that work off your hands, and knowing how the two differ saves time and tokens.

This is the second piece in our series on AI agents applied to branding. In the first, we explained what subagents and agent teams are. This one covers skills, which store procedures, and plugins, which bundle skills, agents and other parts so you can install them in one go. The technical details come from the official Claude Code documentation, checked in October 2026.

If AI agents are still a new subject for you, our guide to agentic AI explains the concept, and the piece on how to use AI in branding walks through the phases people go through when they start.

What a skill is

A skill is a folder with a file called SKILL.md, which holds instructions for a specific task. Claude uses the skill on its own when a request matches it, or you call it by name by typing /skill-name. In practice, it’s a procedure written once and reused whenever the task comes up.

A skill has one advantage over pasting instructions into the conversation: it costs little while it sits unused. Claude Code keeps only each skill’s description in context, so it knows the skill exists. The body, with the full instructions, comes in only when the skill is used. That’s why you can store a long manual without weighing down every conversation.

The old custom commands, which lived in .claude/commands/, have been folded into skills. A .claude/commands/briefing.md file and a .claude/skills/briefing/SKILL.md skill create the same /briefing. The skill adds a folder for supporting files and lets Claude call it without you asking.

What a skill looks like inside

The file opens with a short header followed by instructions in plain text. Here is an example for someone who publishes on a blog:

---
name: post-checklist
description: Checks a blog post before publication. Use when a draft is ready for final review.
---

Check the draft in this order and list what's missing, without rewriting the text:
1. Title of up to 60 characters, with the keyword at the start.
2. Description of up to 132 characters.
3. At least three internal links with descriptive anchor text.
4. Paragraphs of up to 60 words.
5. Alt text on the cover image.

As with subagents, the field that carries the most weight is description. It tells Claude when the skill applies. The documentation recommends keeping SKILL.md under 500 lines and warns that the description gets cut off in the listing beyond 1,536 characters.

The folder can hold more than SKILL.md. Examples, reference material and scripts go in files next to it, and SKILL.md says when each one should be opened. A script gets executed without its text entering the context. Personal skills live in ~/.claude/skills/ and project skills in .claude/skills/.

Skill, subagent or CLAUDE.md: which one to use

All three store instructions, but each one enters the conversation in its own way, and that difference decides which to use.

A rule of thumb: if the instruction applies to everything done in the project, it goes in CLAUDE.md. If it applies to a task that comes up now and then, it becomes a skill. If the task produces a lot of intermediate material, it goes to a subagent, which can even load skills of its own.

What a plugin is

A plugin is a folder that brings together skills, subagents, hooks and MCP servers so they’re installed and loaded as a set. Hooks are commands that run at fixed moments, such as after every edit. MCP servers connect Claude to external tools. A plugin is identified by a manifest file, .claude-plugin/plugin.json, with its name, version and description.

All of this works without a plugin. A skill saved in ~/.claude/skills/ already applies to all your projects. A plugin starts to make sense when you want to distribute a set: install it with one command, receive updates and make sure the team uses the same version. A plugin’s skills show up with its name in front, as in /my-plugin:review.

How to install a plugin

Plugins come from a marketplace, which is a catalog: a repository with a file listing the plugins and where to fetch each one. Claude Code adds Anthropic’s official marketplace, called claude-plugins-official, the first time you open a session in the terminal. To see what it offers, type /plugin and open the Discover tab.

To install by name, the format is plugin@marketplace:

/plugin install commit-commands@claude-plugins-official

To use another company’s catalog or your team’s, add the marketplace once and then install its plugins:

/plugin marketplace add owner/repository
/plugin install plugin-name@marketplace-name

During installation you pick the scope. User scope turns the plugin on for you in every project on the computer. Project scope turns it on for everyone working in that repository, through the shared settings file. Local scope turns it on only for you and only in that project.

What to check before installing

A plugin runs with your permissions. If it ships scripts or servers, they run on your machine as if they were you. The documentation recommends reviewing a plugin before installing it, and that also applies to third-party marketplaces, which means every one that doesn’t belong to Anthropic, a colleague’s included.

Active plugins also take up room. The name and description of every skill and agent stay in context on every message, even in conversations where nothing from the plugin is used. So it pays to turn off what you aren’t using. In the installed tab, /plugin shows the plugins with no recent use, and the official marketplace displays an estimate of the context cost before you install.

How we organize our plugins

At Headcore, the brand operation runs on Growth System™, six plugins we distribute through a private marketplace. brandbrain builds each brand’s Brand Brain, with positioning, voice and rules. craft produces pieces, surf handles pages and brand-studio handles video. growthmkt measures what went live, and mesa-criativa judges each piece before it’s published. Together they add up to 17 subagents.

The decision that carried the most weight was turning rules into structure. In craft, a concept that hasn’t been approved can’t generate a piece, because the code won’t allow it.

In growthmkt, the output has no field for money, and a test checks this on every version: the plugin measures, while the budget decision stays with a person. We built it this way because the model can let an instruction written in the prompt slip by, and it has no way around a lock in the code.

We also learned something about updates. An installed plugin lives in a separate copy of the repository, and editing the file in the repository doesn’t change what the session loads. The change only arrives after it’s pushed to GitHub and updated through /plugin. When an adjustment seems to have no effect, the first thing to check is whether the session is running the new version.

Where to start

Start with skills. Pick a procedure you explain to the AI every week, like a post checklist or a briefing outline, and turn it into a SKILL.md. Test it by calling /name, then see whether Claude uses it on its own when a request matches the description.

Once you have a few skills and one or two subagents the whole team should use, package everything into a plugin and create a marketplace in a company repository. Before installing other people’s plugins, read what they execute, and turn off whatever you stop using.

The full series

This piece is part of a five-part series on AI agents applied to brand work:

1. Subagents and agent teams in Claude Code
2. Skills and plugins in Claude Code (this piece)
3. CLAUDE.md and Brand Brain: brand memory for AI
4. How to review AI-generated content
5. How we use AI in Headcore’s operations

Frequently asked questions

Sources: Claude Code documentation on skills and plugins, checked on 21/10/2026.