Claude Skills vs Plugins: What They Actually Are and Which One Your Team Needs

I used to open a new Claude chat and retype AI Operator's brand voice rules from scratch every time: how we format a LinkedIn post, our file structure, the reminder to skip em dashes. Now those rules live in a reusable Skill. That's what Skills and Plugins are built to fix, and most people building with Claude still mix the two up.

They're not the same thing. A skill is a file of instructions. A plugin is a package of capabilities. Anyone can create either one, but they solve different problems, and knowing which one you need saves you from building something that already exists or installing something generic when you needed five minutes of writing.

What a Claude Skill Actually Is

A skill is a folder Claude reads before it starts a task. Inside is a SKILL.md file with instructions, and often templates, reference files, or small scripts alongside it.

Say a company has a specific way of writing LinkedIn posts: first person for personal accounts, "we" only when it's genuinely a team effort, no em dashes, always tied to something happening in the news that week. Instead of typing that out every time, the company writes it once as a Skill. Claude reads it, follows it, and produces the same quality output no matter who prompts it.

How to Create Your Own Skill

You don't need to code. A skill is just a folder with a Markdown file inside it.

Here is the basic workflow in Claude Code.

Step 1: Create the folder

Where you put the folder decides who can use the skill:

• ~/.claude/skills/your-skill-name/ creates a personal skill you can use across projects.
• .claude/skills/your-skill-name/ creates a project skill for that specific codebase or workspace.

Use lowercase names with hyphens so the command is easy to remember.

Step 2: Add a SKILL.md file

Inside the folder, create a file called SKILL.md.

At the top, add frontmatter between --- markers. The most important field is description, because that tells Claude when the skill should load.

Below the frontmatter, write the actual instructions in plain language.

---
description: Writes LinkedIn posts in our brand voice. Use when asked to draft or review a LinkedIn post.
---

When writing a LinkedIn post:
- First person, conversational, no corporate tone
- No em dashes, use commas or periods
- Tie the post to a real news event from this week
- End with a question, not a sales pitch

Step 3: Use the skill

The folder name becomes the command. If the folder is called linkedin-post, you can run it as /linkedin-post.

Claude can also invoke the skill automatically when your request matches the description. If you do not want Claude to trigger it on its own, add this to the frontmatter:

disable-model-invocation: true

That keeps the skill available for manual use, while stopping Claude from choosing it automatically.

Step 4: Improve it over time

Start simple. Add examples, templates, scripts, or reference files only when they make the skill more reliable.

Keep the main SKILL.md lean. If the skill needs long examples or brand documentation, put those in separate files inside the same folder and reference them from the instructions.

The Fastest Way to Build a Skill: Just Ask Claude

You don't have to create the folder or write the file by hand. The quickest route is to describe what you want and let Claude build the skill for you inside Claude Code.

Type a request like this:

Create a skill called linkedin-post that writes LinkedIn posts in our voice: first person, conversational, no em dashes, always tied to a real news event from this week, and ending with a question instead of a sales pitch. Save it to my personal skills folder so I can use it in every project.

Claude creates the folder, writes the SKILL.md, fills in a description so the skill triggers at the right moment, and tells you the command to run. You review it, tweak anything you want, and it's live.

Two ways to get a sharper result:

• Ask Claude to interview you first. Add "ask me a few questions before you write it" and the instructions come from your answers instead of its guesses.
• For skills that matter, install Anthropic's skill-creator, a plugin built specifically to help you write, test, and improve skills:

/plugin marketplace add anthropics/claude-plugins-official
/plugin install skill-creator@claude-plugins-official

Then say something like "help me build a skill for summarizing sales calls" or "evaluate my linkedin-post skill." It walks you through real test cases, runs them, and shows you whether the skill actually improves the output before you rely on it.

One more trick worth knowing: you don't need a blank chat to start. If you've already solved a problem in a conversation, whether that took five messages or fifty, just say "turn this into a skill." Claude looks back at what worked in that chat and packages it into a reusable SKILL.md for you. It's the fastest path when the skill you need already exists, buried in a chat you had last week.

The Newest Way: Record a Skill

Claude has also started moving toward a more visual way to create skills: record yourself doing the task once, then let Claude turn that workflow into reusable instructions.

Instead of writing a skill from scratch, you show the process on screen. Claude watches the sequence, drafts the skill, and gives you something to review before saving.

This is most useful for repeatable workflows that are easier to demonstrate than explain, like preparing a weekly newsletter summary, formatting a recurring report, or turning raw notes into a standard client update.

Before you record, update the Claude app, grant any screen-recording or accessibility permissions it asks for, and close anything private. Everything visible on screen becomes part of the recording context, so keep passwords, client data, and sensitive tabs out of view.

A safe workflow looks like this:

1. Open the skill creation flow in Claude.
2. Choose the Record a Skill option if it appears in your account.
3. Start recording.
4. Do the task normally, narrating the key decisions as you go.
5. Stop the recording when the task is complete.
6. Review the draft skill before saving it.
7. Test it on a second example before using it in real work.

Treat the first version as a draft, not a finished system. Recorded skills are fast, but they still need cleanup. Check that the steps are clear, remove anything too specific to one example, and make sure the skill explains what to do when the input changes.

For AI Operator, this is useful when the process is visual or hard to describe cleanly in a prompt. If the workflow is simple, ask Claude to write the skill. If the workflow is nuanced, show it once and then refine the result.

What a Claude Plugin Actually Is

A plugin is a bundle of Claude Code upgrades packaged together in one install.

If a skill teaches Claude how to do one repeatable task, a plugin gives Claude a larger working setup. It can include instructions, shortcuts, tool connections, automations, and supporting files, all organized in a structure Claude knows how to load.

Think of it like installing an app extension. You do not copy each prompt, script, template, and integration by hand. You install the plugin once, and Claude gets access to the pieces inside it when they are needed.

A plugin can include:

• Skills: reusable instructions for specific tasks.
• Commands: shortcuts you run with /, like /report or /review.
• Connectors or MCP servers: the wiring that lets Claude work with outside tools, APIs, databases, or internal systems.
• Agents: specialized assistants built for one type of job.
• Hooks: automations that run at certain moments in the workflow.
• Assets: templates, examples, brand files, reference docs, images, or other supporting files.
• Structure files: configuration that tells Claude what the plugin contains and how to load it.

Not every plugin has all of these. A small plugin might only include one skill and a template. A larger team plugin might include skills, commands, MCP servers, hooks, and shared assets together.

The point is packaging. Instead of sending teammates a folder of random prompts, scripts, and setup notes, you give them one organized install.

That makes plugins useful for shared systems: onboarding workflows, code review standards, reporting packs, customer research workflows, or integrations with tools your team already uses.

Most plugins people use are built by someone else, because the hard part is often the setup: structure, authentication, updates, and tool wiring.

You can still build your own plugin for your team. Just keep it simple: start with one repeatable workflow, add only the pieces it needs, and organize everything in the expected plugin structure.

How to Install a Plugin

The easiest way is visual, no commands needed.

• In the Claude app, open the Customize menu in the left sidebar. Chat and Cowork now share one window (switch between them from a toggle in the message box), and Customize is the same place for both: plugins, skills, and connectors together.
• Open the Plugins tab and click Browse plugins. This opens Anthropic's directory, where there are lots to choose from.
• Click Install on the one you want. Done.
• Prefer a browser? The same catalog lives at claude.com/plugins.

A few from the official directory worth a look:

• github connects Claude to your repositories, issues, and pull requests.
• atlassian wires up Jira and Confluence for issues, docs, and sprints.
• security-guidance reviews your code for security issues as you work.

Prefer the terminal? In Claude Code it's two commands:

Step 1: Add a marketplace. A marketplace is just a GitHub repo that lists plugins. Anthropic runs the official one:

/plugin marketplace add anthropics/claude-plugins-official

Any public GitHub repo set up as a marketplace works the same way with /plugin marketplace add owner/repo.

Step 2: Install the plugin you want:

/plugin install plugin-name@claude-plugins-official

Or skip the commands entirely: type /plugin, open Discover, and browse what's available. Installed plugin skills show up namespaced as /plugin-name:skill-name, which means they can never accidentally overwrite a skill you wrote yourself.

Where to Find Good Skills and Plugins

• The official marketplace at github.com/anthropics/claude-plugins-official is Anthropic's curated directory and the safest starting point.
• agentskills.io documents the open Agent Skills standard that skills follow, with guides on writing and evaluating them. Skills built to this standard work across Claude Code and other compatible tools.
• Community marketplaces on GitHub. Many companies publish their skills publicly (Sentry's engineering skills repo is a good example of a well-structured one). Add any of them with /plugin marketplace add owner/repo. One caution: plugins run code on your machine, so stick to repos from teams you trust.

When to Write a Skill vs When to Install a Plugin

Write a Skill when the task involves knowledge only your business has: your onboarding checklist, your reporting format, or the exact way a sales call should be summarized before it goes into your CRM. Nobody has packaged that, because nobody else works the way your business does.

Install a Plugin when Claude needs to reach something outside its native capabilities and someone already built the bridge: web search, database access, Slack, Google Drive, or another external system. Building that yourself means writing an MCP server from scratch, which almost nobody needs to do when a well-maintained option already exists.

Most teams end up using both. Plugins handle the infrastructure. Skills handle the "here's how we specifically do this" layer on top of it.

Four Things Most Explainers Get Wrong

Most explainers make Skills sound either too simple or too technical. The useful truth sits in the middle. These four details are what make Skills practical for real teams.

1. Skills barely cost anything until they're used. Only the skill's name and short description sit in context by default. The full instructions load only when you or Claude invoke the skill. That means you can keep dozens of skills available without quietly eating up the context window.

2. You can control who triggers a skill. Set disable-model-invocation: true in the SKILL.md frontmatter and only you can run it, never Claude on its own. That matters for anything with real consequences, like sending a message or deploying something. You can also use user-invocable: false when a skill should work in the background without showing up as a manual command.

3. Skills can bring in live data before Claude starts. A skill can run a shell command first and place the real output into the prompt. That means it can hand Claude today's numbers, files, or project state instead of forcing Claude to guess or hunt for context after the task starts.

4. Skills stack across different levels. Skills can live at the enterprise, personal, and project level, plus inside plugins. Claude Code has clear priority rules when names overlap: enterprise beats personal, and personal beats project. Plugin skills are namespaced, like plugin-name:skill-name, so installing someone else's plugin won't overwrite your own skill.

Power Moves Worth Knowing

Once a team has written a Skill or two, these options turn it from a saved prompt into a real operating layer for repeatable work.

1. Pass details on the fly. A skill can take input right after the command. Write $ARGUMENTS or $1, $2 for specific values in your instructions, then run something like /fix-issue 123. Claude drops 123 straight into the task.

2. Auto-trigger only for the right files. Add a paths line to the frontmatter and the skill loads automatically only when you're working on matching files. It stays available when needed and stays out of the way the rest of the time.

3. Pre-approve the safe tools. List trusted tools under allowed-tools so Claude can use them without stopping to ask permission every time. This is useful when the skill has a clear, repeatable workflow and the allowed actions are low-risk.

4. Run heavy work separately. Add context: fork and the skill runs in its own isolated space before reporting back. This keeps the main conversation clean, especially for research, reviews, or longer checks.

5. Adjust the model and effort. A skill can override the model or reasoning effort with model and effort. Use more thinking power for careful reviews, and keep quick formatting or cleanup skills lightweight.

6. Keep the main file lean. Put the essentials in SKILL.md, aim for under 500 lines, and move long reference material into separate files in the same folder. Claude loads extra files only when it needs them, so the skill stays fast and focused.

How We Use This at AI Operator

At AI Operator, we build our own Skills in-house rather than relying only on what's available off the shelf, because the value of a Skill is how specific it is to how a team actually operates. One enforces our exact brand voice down to which words are banned. Another checks drafts for AI writing patterns line by line before they go near a client or a LinkedIn post. Each one exists because repeated instructions should become reusable systems.

That's the same approach we walk teams through in the Operationalize phase of the ADOPT Method™, where the goal is to move from scattered AI usage to tools the whole team can run the same way, every time.

Frequently Asked Questions

What is a Claude Skill?

A skill is a set of instructions, usually a Markdown file plus supporting templates or scripts, that teaches Claude to handle a specific, repeatable task the way you want it done. You write it in plain language, no coding required.

What is a Claude Plugin?

A plugin is a packaged bundle of capabilities installed in one step. It can include one or more skills, MCP servers, hooks, and subagents. Anyone can build one, but most people install plugins others have already published.

How do I create a skill?

Create a folder at ~/.claude/skills/your-skill-name/, add a SKILL.md file with a short description and your instructions, and it's live. The folder name becomes the /command.

Can I just ask Claude to create the skill for me?

Yes, and it's usually the fastest way. In Claude Code, describe the skill you want and where to save it, and Claude writes the SKILL.md for you. For skills that matter, install the skill-creator plugin to test and refine them before you rely on them.

How do I install a plugin?

Add a marketplace with /plugin marketplace add anthropics/claude-plugins-official, then run /plugin install plugin-name@claude-plugins-official. Or browse everything through the /plugin menu.

What is Record a Skill?

A Claude Cowork feature, Pro/Max/Team plans, Claude for Mac only. Record yourself doing a task once, narrating as you go, and Claude proposes a skill to review and save. Only screenshots are kept afterward, not the video or audio. Treat the first pass as a draft and refine it before you rely on it.

Can I turn an existing chat into a skill?

Yes. If you've already worked through a task in a conversation, tell Claude to turn it into a skill. It packages the approach you already used into a reusable SKILL.md, no need to start from a blank file.

The Takeaway

Skills are what you write because nobody else knows your business the way your team does. Plugins are what you install because someone already solved the general problem. Confusing the two either wastes hours building something that already exists, or leaves you stuck re-explaining yourself to Claude every single day.

If you want help figuring out what your team should build versus install, that's exactly what we walk through in AI Operator live sessions and inside the ADOPT Method™.

Explore the ADOPT Method™