Claude Code skills and plugins: what a one-person AI agency needs to know

A web designer finishes a site for a hair salon and asks Claude Code to check it before the client sees it. It checks some things well and forgets others: the booking link on the mobile menu, the alt text on the gallery, the salon's phone number in the footer. Next week, for a different client, she pastes her checklist into the chat again. And the week after.

That repeated paste is what Claude Code skills are for. Anthropic added skills in October this year, a few days after plugins, and between them they change how you teach your agent the way you work. This article explains both in plain words, what they're good for when you use AI agents for client work, and what to check before installing one written by someone else.

What are Claude Code skills?

A skill is a folder of instructions for one kind of task. At its simplest, it's a folder containing a file called SKILL.md: a short name and description at the top, then the instructions, written the way you'd explain the task to a new assistant. A skill can also include scripts and other files the instructions refer to.

The clever part is how Claude uses them. Anthropic's engineering write-up describes it as progressive disclosure: Claude reads only the short description of each skill at the start, and loads the full instructions when a task looks relevant. So you can have many skills without filling every conversation with all of them.

Anthropic introduced Agent Skills on 16 October 2025, for Claude's apps, its API and Claude Code. In Claude Code, you can add skills to a folder in your own settings, share them with a project through version control, or install them through plugins. On 18 December Anthropic also published Agent Skills as an open standard, so the same format can work beyond Claude.

What are Claude Code plugins?

A plugin is a package for sharing Claude Code setup. Anthropic announced plugins on 9 October 2025, in public beta, as a way to bundle any combination of:

• slash commands, shortcuts you type for things you do often;
• subagents, helpers set up for a particular kind of task;
• MCP servers, connections to other programs and information through the Model Context Protocol;
• hooks, small steps that run at set points in Claude Code's work.

You install and manage plugins with the /plugin command, and you can switch them on and off. Anthropic suggests turning off the ones you don't need, to keep Claude's working context small. Plugins come from marketplaces, which are lists of plugins anyone can host, often in a code repository. Anthropic's own example marketplaces include plugins for code review and security guidance.

Skills and plugins overlap, which confuses people. The simplest way to hold it:

	Skill	Plugin
What it is	Instructions for one kind of task, in a folder	A package that bundles setup for sharing
What it contains	A SKILL.md file, plus optional scripts and files	Commands, subagents, MCP connections, hooks, and skills
When Claude uses it	When a task matches the skill's description	Whatever it contains becomes available once installed and switched on
Who it's for	Anyone who repeats a task	Anyone sharing a setup across projects or people

What can a web designer use them for?

Start with the instructions you already repeat. For a small web business, good first skills look like this:

• Your pre-handover check. The list you paste every time: mobile menu, forms, alt text, contact details, page titles, the favicon. Written once as a skill, Claude uses it whenever you ask for a review.
• A client's brand notes. Their tone of voice, words they never use, colours, how they write their own name. One skill per client keeps each voice straight.
• Your page structure. How you lay out a service page or a menu page for a local business, so every build starts from your pattern instead of a generic one.
• Your quote format. The headings your quotes always have, and the "not included" list you never want to forget.

Here's what a first brand-notes skill might look like for a family dentist, in plain text:

Name: dental-practice-voice. Description: Use when writing or editing any page, email or post for the dental practice. Instructions: Warm and calm, never clinical jargon without an explanation. Say "the practice", not "the clinic". Always give the phone number in full. Never promise pain-free treatment or results.

That's the whole thing. Claude reads the description, recognises a task for the practice, and follows the rest. When the practice changes how it wants to sound, you edit one file.

Plugins make sense once you have a few of these and want them to travel together, or when you work across several projects and want the same setup in each.

If you're weighing whether to run an agency with Claude Code at the centre, this is the practical side of it. Skills are how your agent learns your way of working, instead of starting from scratch in every chat. They sit alongside MCP connections, which give the agent your information rather than your instructions; what is MCP, and what it changes for agency work covers that side, and the short definition of what MCP means is a quick reminder.

What should you check before installing someone else's?

A skill can include scripts, and a plugin can add connections and hooks. That's the point of them, and it's also the risk.

Anthropic is direct about this. Its engineering post warns that a malicious skill could introduce weaknesses, direct Claude to send data somewhere it shouldn't, or take actions you didn't intend. It recommends installing skills only from trusted sources, and reading a less-trusted skill thoroughly before using it, including any instructions that tell Claude to connect to outside services.

Before installing, check:

1. Who made it. A known company or person, or an account you've never heard of?
2. What's inside. Open the SKILL.md and any scripts. If you can't tell what a script does, don't run it on client work.
3. What it connects to. A plugin that adds an MCP connection or a hook can reach beyond the chat. Know where.
4. Whether you need it. Fewer, well-understood skills beat a folder of downloads.

And keep one rule regardless of what's installed: anything a client or the public will see waits for you. A skill can make your agent's drafts better; it doesn't decide what gets sent. Approving what your agent sends stays with you.

Where should you start?

Pick the one checklist you paste most often. For most web designers it's the pre-handover review.

1. Write it out as plain instructions, the way you'd explain it to someone new. Be specific: "check the phone number in the header and footer matches the Google listing", not "check contact details".
2. Add a one-line description at the top that says when to use it: "Use when reviewing a client website before the client sees it."
3. Save it as a skill in Claude Code, following the skills documentation.
4. Ask for a review on your next site, and see what it catches and what it misses. Edit the instructions, not the chat.

After a month, write a second one. For developers building client sites with an AI agent, and for designers too, a handful of skills that describe how you actually work is worth more than any large collection you could download.

Skills and plugins are a few months old as this year closes, and they'll keep changing. The habit underneath them won't: write down how you work, once, and let your agent read it every time.