What Are Claude Code Skills? A Plain-English Guide

If you've searched "claude code skills explained" and come away more confused than before, that's fair — a lot of the writing on this treats skills like a new product category with a marketplace and a business model, when the actual thing is much smaller and more useful. A skill is a folder with instructions in it. This post covers what a skill actually is, how it differs from a subagent or an MCP server, how to create and install one, and — honestly — when you don't need one at all.

What a skill actually is

A Claude Code skill is a directory containing a markdown file (usually SKILL.md) that describes a task, plus whatever reference material the agent needs to do it well — checklists, code templates, example output, links to docs. The agent doesn't run this folder like a program. It reads the instructions the way it reads anything else in its context, then acts on them using its normal tools.

The key word is "on demand." A skill isn't loaded into every conversation by default. Each skill folder carries a short description, and the agent decides whether a given request matches it closely enough to pull the full skill into context. Ask it to fix a typo and your "review PRs for security issues" skill stays out of the way. Ask it to review a PR, and it loads.

That's different from what people assume. A skill is not a plugin running in the background, not compiled code, not a model fine-tune. It's closer to a well-organized note you leave for the agent, structured so it can find and use it when relevant.

Claude Code agent vs skill vs subagent vs MCP server

This is the question that actually generates most of the "claude code skills explained" searches, so it's worth being precise:

• The agent is the thing doing the work — reading your request, deciding what to do, calling tools.
• A skill is instructions the agent loads into its own context to do one job well. It doesn't run independently; it shapes how the current agent behaves.
• A subagent is a separate agent session the main agent can delegate a scoped task to, with its own context window, that reports back a summary instead of dumping its full transcript into the main thread. See our deeper dive on subagents specifically — subagents solve a context-pollution problem, not an instruction problem.
• An MCP server is a live connection to an external tool or data source — a database, an API, a browser — that gives the agent new capabilities it didn't have before.
• A line in CLAUDE.md is a standing instruction that's always in context, every conversation, relevant or not.

Put together: CLAUDE.md is what the agent always knows, a skill is what it knows only for a specific task, a subagent keeps a big task from cluttering the main conversation, and an MCP server is what the agent can reach out and touch. None of these replace the others — a real setup usually uses all four.

How to create and install a claude code skill

The practical shape is what most "how to create claude code skills" and "how to install claude code skills" searches are actually asking for:

1. Make a folder named for the task, with a SKILL.md file in it.
2. Write frontmatter at the top — a name and a short description of when this skill applies. The description is what the agent matches against your request, so it needs to be specific enough to trigger correctly and not on unrelated requests.
3. Write the instructions below the frontmatter — the steps, checklist, or reference material to follow once the skill applies.
4. Add supporting files if needed — templates, example inputs and outputs, scripts the instructions reference, in the same folder.
5. Place the folder where the agent looks for skills. Personal skills live in your user config; project skills live in the repo so everyone working in it gets the same skill.

No build step, no packaging, no registry account required. Write markdown, put it in the right directory, and the next matching request picks it up.

What people actually build skills for

The realistic use cases cluster around a few categories, matching what "claude code review skill" and "claude code design skill" searches suggest people are actually trying to do:

• A review skill — a checklist of what to look for in a diff before it ships: security patterns to flag, style conventions specific to the repo, mistakes the team has hit before. This is one of the highest-value categories, because review quality is inconsistent without a written checklist and a skill makes it consistently applied.
• A design or frontend skill — conventions for a specific design system, component patterns, spacing rules, accessibility requirements that are easy to forget mid-task.
• Document-handling skills — a claude code xlsx skill for spreadsheets in a specific format your team uses, or similar skills for slide decks and structured documents that need to follow a house style.

Where the "skills marketplace" framing oversells things: a marketplace implies skills are polished products you shop for and install blind. In practice the useful ones are almost always written for your own codebase, review standards, and document format — a generic "review skill" from somewhere else knows nothing about your repo's conventions. The value is in writing your own, not browsing a catalog of other people's.

When a skill is overkill

This is the section that's easy to skip and shouldn't be: most tasks don't need a skill. If you're asking the agent to do something once, a plain prompt is better than writing it down. Skills earn their cost when a task recurs — the same review checklist on every PR, the same document format every week, the same design conventions across dozens of components. A one-off task makes a SKILL.md more setup than the task itself.

The other honest caveat: a skill only helps if the agent's matching actually triggers on your request. A vague description means it sometimes doesn't load when it should — a reason to write the description as precisely as the instructions themselves, not a reason to avoid skills.

The meshcode angle

meshcode is a native desktop app for macOS and Windows built around a multi-pane agent workspace — each pane is its own agent session, and panes run in parallel, so you can watch and steer several agents at once instead of one conversation at a time. Rather than reinventing your CLI setup, meshcode drives your existing Claude Code and Codex subscriptions directly: your skills, hooks, MCP servers, and CLAUDE.md carry over and work unchanged. If you've already built a review skill or a document-handling skill, you don't rebuild it to use meshcode — same config, different workspace.

meshcode also offers its own metered mesh models, billed pay-as-you-go with no monthly fee, and ships native built-in agent tools — browser automation, Excel, a kanban board, a calendar — that the agent can reach for directly, without wiring up a separate MCP server for each one.

Bring your own skills: meshcode is free to start — download it and run the skills you've already written across several agent panes at once.

The takeaway

A skill is a folder of instructions an agent loads when a request matches, nothing more exotic than that. It's distinct from a subagent (separate context, delegated task), an MCP server (new capability), and CLAUDE.md (always-on instructions) — each solves a different problem, and most real setups use all four together. Write skills for tasks that recur; skip them for one-offs. If you want to run that setup across multiple parallel agent panes instead of one conversation at a time, that's what meshcode's workspace is built for — your existing skills included, unchanged.

FAQ

How do you actually create a Claude Code skill? Start from something you already do more than once — a review checklist, a formatting pass, a document type you keep reformatting by hand. Keep the description narrow enough that it only matches the real trigger case; too broad and it loads (or doesn't) unpredictably. Test it before trusting it: send a request you expect to match and confirm the skill actually loaded.

How do you install a skill someone else wrote? There's no installer — copy the folder into wherever your skills live, personal config for everywhere or the project's repo for the whole team. Nothing needs restarting; the next matching request picks it up. Read the SKILL.md first, though — a skill written for someone else's repo conventions may not fit yours.

What's the real difference between a skill and a subagent? A skill changes what the current agent knows; a subagent changes who's doing the work. Reach for a skill when the agent just needs the right steps for a task in the conversation you're already having. Reach for a subagent when the task would flood that conversation with exploration you don't want cluttering the main thread — it runs separately and hands back a summary instead.

Is there an actual Claude Code skills marketplace? Not a curated one worth treating as a shopping list — what exists is scattered repos of shared examples, most too generic to help, because the useful skills are tied to your own repo's conventions. The realistic path is writing your own from tasks you actually repeat, which also makes them portable: the same skill folder works whether you're driving Claude Code from a terminal or from inside meshcode's multi-pane workspace, since meshcode reads your existing CLI config unchanged.

👉 Download meshcode — Mac, Windows