A skill teaches an AI assistant how to do a job. MCP connects it to an app or a source of data, so it can reach the thing it needs. A plugin is a box that packages skills, MCP connections and other parts so you can install them in one go. Anthropic puts the first two in one line: “MCP connects Claude to data; Skills teach Claude what to do with that data.”1
- Skill: written know-how for one kind of job, read only when a task needs it.
- MCP server: a connector to an app or data, such as a drive, a calendar or a video generator.
- Plugin: a package that bundles skills, connectors and more into one install.
- Subagent: a helper AI with its own workspace and its own permissions, for a separate task.
- Custom instructions and Projects: background you set once. Instructions apply everywhere; a Project holds files for one piece of work.
- They work together. You rarely pick one; a real setup often uses several.
Picture a kitchen. A skill is a recipe card the cook reads when that dish is ordered. MCP is the door to the supplier, so the cook can fetch ingredients. A plugin is a meal kit: recipe cards, a supplier account and a gadget in one box. A subagent is a second cook working at their own bench. Custom instructions are the house rules on the wall, and a Project is a shelf of binders for one event.
Skills, MCP, plugins and the rest, side by side
The plain point: each one answers a different question: how to do it, what to reach, how to install it all, who does it, and what to always remember.
| Thing | What it is | When it is used |
|---|---|---|
| Skill | What it is: a folder of instructions, sometimes with scripts, for one kind of job. | Used: only when your request matches it. |
| MCP server | What it is: a connector that lets the AI use an outside app or data. | Used: when a task needs that app or data. |
| Plugin | What it is: a package of skills, connectors, subagents and more. | Used: its parts work as normal once installed. |
| Subagent | What it is: a separate helper AI with its own instructions and tools. | Used: when the main AI hands it a task. |
| Custom instructions | What it is: your standing preferences. | Used: in every conversation. |
| Project | What it is: a workspace with its own files and instructions. | Used: in every chat inside that Project. |
Anthropic’s own comparison table sums up what each gives: skills give “Procedural knowledge”, Projects “Background knowledge”, subagents “Task delegation” and MCP “Tool connectivity”.2 The Claude Code docs put the set in one sentence: “CLAUDE.md handles always-on context, skills handle on-demand knowledge and workflows, MCP handles external connections, subagents handle isolation, and hooks handle automation.”3 (CLAUDE.md is Claude Code’s file of always-on instructions; hooks are covered below.)
Skill = how. MCP = access. Plugin = the box they come in. Subagent = a separate worker. Instructions and Projects = background it always has.
What is a skill?
The plain point: a skill is a saved method the AI reads only when a task calls for it.
Anthropic calls skills “folders that include instructions, scripts, and resources that Claude can load when needed.”4 Until it is needed, the assistant sees only the skill’s name and a one-line description, so you can have many without crowding it.5 Since 18 December 2025 the format has been an open standard, so other AI apps can read the same skills.6 What is a skill in AI? lists which apps.
Anthropic’s test for when to make one is simple: “If you find yourself typing the same prompt repeatedly across multiple conversations, it’s time to create a Skill.”7
What is an MCP server?
The plain point: MCP gives the AI a way in to another app; it does not teach the AI what to do once inside.
MCP stands for Model Context Protocol, an open standard for connecting AI apps to outside systems. A server here is the piece of software, run by the app maker or on your own computer, that offers the AI a set of actions.7 The MCP project’s own comparison is “a USB-C port for AI applications.”8 In the Claude apps these are called connectors. What is MCP? explains it from scratch.
Anthropic’s help centre draws the line with skills like this: “MCP connections give Claude access to tools, while skills teach Claude how to use those tools effectively.”9 So if the AI cannot reach your files, you need MCP; if it reaches them but handles them badly, you need a skill.
What is a plugin?
The plain point: a plugin is packaging. It holds the other parts so you can install, share and update them together.
The Claude Code docs say: “Plugins are the packaging layer. A plugin bundles skills, hooks, subagents, and MCP servers into a single installable unit.”3 In the Claude apps, the help centre describes the same idea: each plugin “bundles skills, connectors, and sub-agents into a single package”, and plugins there need a paid plan.10 OpenAI uses the word the same way for Codex: to share skills widely, “or optionally bundle them with connectors, use plugins.”11
You get plugins from a marketplace, a catalogue of plugins; how to install AI skills shows the commands.
If one thing you install brings several skills, a connector and a few commands, it is a plugin. Check every part inside it, not just the name on the box.
Subagents, custom instructions and Projects
The plain point: these are not add-ons you download; they shape who does the work and what the AI always knows.
- Subagent: Anthropic describes subagents as “specialized AI assistants with their own context windows, custom system prompts, and specific tool permissions.” A context window is the amount of text an AI can hold in mind at once. Its rule of thumb: skills for expertise any agent can use, subagents for a separate task with its own permissions.7
- Custom instructions: the help centre says they “apply broadly to all your conversations”, while skills “only load when relevant”.9
- Projects: “Projects give Claude persistent context for a specific body of work”, while “Skills teach Claude how to do something.” Anthropic adds that copying the same instructions into several Projects is a sign you need a skill.7
One more word you may meet in Claude Code: a hook, a rule that runs automatically at a set moment. The docs’ test: “Use a hook when the action must happen the same way every time and doesn’t need Claude to think.”3
When to use which
The plain point: start from the problem, not the feature. This list is this site’s summary of the sources above.
- “I keep typing the same instructions.” Make or install a skill.
- “It cannot see my files, calendar or account.” Add an MCP server (a connector).
- “It can see them but does the job badly.” Add a skill that explains your way of doing it.
- “I want a ready-made setup, or to share mine with a team.” Use a plugin.
- “This side task clutters the main conversation, or needs tighter permissions.” Use a subagent.
- “It should always know this about me.” Use custom instructions.
- “It needs these documents for one piece of work.” Use a Project.
Most real setups combine them: a connector to reach the data, a skill to handle it well, and a plugin to install both at once.
One example of each from AI video
The plain point: the same three ideas show up in AI video tools.
- A skill: Remotion’s agent skills. Remotion is a tool for making videos by writing code in React, a popular way to build web pages. Its makers say: “Remotion maintains a list of Agent Skills that define best practices for working in Remotion projects.”12 They teach the AI how to write a Remotion video well; they do not connect it to any outside account. See Remotion agent skills.
- An MCP server: Runway MCP. Runway makes AI video models. Its MCP server lets an assistant generate images and videos through your Runway account: “No separate API key is required – generations are tied to your existing Runway plan.”13 That is access, not know-how, and every video it makes uses your plan. (ElevenLabs, the voice company, also runs an MCP server, but its docs describe it as managing voice agents in your workspace, not general voiceover.)14
- A plugin: HyperFrames. HyperFrames, by HeyGen, turns web pages into video. It ships as a Claude Code plugin that holds many skills; its README says “HyperFrames ships 21 skills agents load on demand.”15 One install brings them all. See HyperFrames, and the best AI video skills for the rest.
What each one can do on your computer
The plain point: the more a thing can do, the more carefully you check it before installing.
- A skill is mostly text, but it can include code. Anthropic warns that “a malicious Skill can direct Claude to invoke tools or execute code in ways that don’t match the Skill’s stated purpose.”5
- An MCP server passes your requests, and anything you hand it, to the service it connects to.Reasoning, 13 Oct 2026 — this follows from what a connector is for (reaching an outside service), as described in the MCP section above; it is not a quote from a vendor. A connector on your own computer can run with your permissions; the help centre warns that local MCP servers in plugins “run on your computer with the same permissions as any other program you run.”10
- A plugin can carry all of the above. Claude Code’s docs: “A Claude Code plugin you install can execute arbitrary code on your machine with your user privileges.”16
The checklist before installing a skill and the permissions checklist for agents and plugins cover what to look for.
Questions people ask
Do skills replace MCP?
No. They do different jobs and are meant to be used together: Anthropic’s advice is “Use both together: MCP for connectivity, Skills for procedural knowledge.”7
Is a plugin the same as a skill?
No. A plugin can contain skills, but it can also contain connectors, subagents and hooks. A single skill is just one folder (sources above).
Which should a beginner start with?
In this site’s view, a skill from a source you trust: it is the easiest to read in full before you install it. Add connectors once you know which app you need the AI to reach.
Where to go next
- “How do I install one?” Follow how to install AI skills.
- “Tell me more about skills.” Read what is a skill in AI? and skills, repos and MCP.
- “Tell me more about MCP.” See what is MCP? and what is tool use?
- “Where do agents fit?” Start with AI agents for beginners.
- “Show me the whole section.” Go to AI skills.
- Related: What Is a Plugin in AI? Plugin Meaning in Plain English.
- Related: What Is a Subagent? Subagent Meaning in Plain English.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What does an MCP server do, according to the page?
- Teaches the AI your house style
- Connects the AI to an outside app or data
- Stores your standing preferences
Show the answer
Answer: Connects the AI to an outside app or data. The page says MCP gives the AI a way in to another app or data; a skill is what teaches it how to do the job.
2. What is a plugin?
- A package that bundles skills, connectors and more into one install
- A prompt you type once for one task
- A helper AI that works on a separate task
Show the answer
Answer: A package that bundles skills, connectors and more into one install. The page calls a plugin the packaging layer: one install that can hold skills, connectors, subagents and hooks.
3. The AI can reach your files but handles them badly. What does the page suggest?
- Add a second MCP server
- Delete your custom instructions
- Add a skill that explains your way of doing it
Show the answer
Answer: Add a skill that explains your way of doing it. In the page’s when-to-use list, access problems need MCP, but doing the job badly needs a skill.
- 1Anthropic, Skills explained: How Skills compares to prompts, Projects, MCP, and subagents, 5 Mar 2026, read at source 13 Oct 2026“MCP connects Claude to data; Skills teach Claude what to do with that data.” ↑
- 2Anthropic, Skills explained, 5 Mar 2026, read at source 13 Oct 2026, comparison table, “What it provides”: Skills “Procedural knowledge”; Prompts “Moment-to-moment instructions”; Projects “Background knowledge”; Subagents “Task delegation”; MCP “Tool connectivity”.↑
- 3Claude Code Docs, Extend Claude Code, read at source 13 Oct 2026“Each extension solves a different problem: CLAUDE.md handles always-on context, skills handle on-demand knowledge and workflows, MCP handles external connections, subagents handle isolation, and hooks handle automation.” ↑a“Plugins are the packaging layer. A plugin bundles skills, hooks, subagents, and MCP servers into a single installable unit.” ↑b“Use a hook when the action must happen the same way every time and doesn’t need Claude to think.” ↑c
- 4Anthropic, Introducing Agent Skills, 16 Oct 2025, read at source 13 Oct 2026“Skills are folders that include instructions, scripts, and resources that Claude can load when needed.” ↑
- 5Anthropic, Agent Skills (Claude developer docs), read at source 13 Oct 2026“until a Skill is triggered, only its name and description occupy context.” ↑a“a malicious Skill can direct Claude to invoke tools or execute code in ways that don’t match the Skill’s stated purpose.” ↑b
- 6Anthropic, Skills for organizations, partners, the ecosystem, 18 Dec 2025, read at source 13 Oct 2026“We’re also publishing Agent Skills as an open standard. Like MCP, we believe skills should be portable across tools and platforms” ↑
- 7Anthropic, Skills explained, 5 Mar 2026, read at source 13 Oct 2026“If you find yourself typing the same prompt repeatedly across multiple conversations, it’s time to create a Skill.” ↑a“MCP servers expose data and capabilities; MCP clients (like Claude) connect to these servers.” ↑b“Subagents are specialized AI assistants with their own context windows, custom system prompts, and specific tool permissions.” and “Use Skills to teach expertise that any agent can apply; use subagents when you need independent task execution with specific tool permissions and context isolation.” ↑c“Projects give Claude persistent context for a specific body of work”; “Skills teach Claude how to do something.”; “If you find yourself copying the same instructions across multiple Projects, that’s a signal to create a skill instead.” ↑d“Use both together: MCP for connectivity, Skills for procedural knowledge.” ↑e
- 8Model Context Protocol, What is the Model Context Protocol (MCP)?, read at source 1 Oct 2026 (as cited on this site’s skills explainer): “Think of MCP like a USB-C port for AI applications.”↑
- 9Claude Help Center, What are skills?, read at source 13 Oct 2026“MCP connects Claude to external services and data sources. Skills provide procedural knowledge—instructions for how to complete specific tasks or workflows. You can use both together: MCP connections give Claude access to tools, while skills teach Claude how to use those tools effectively.” ↑a“Custom instructions apply broadly to all your conversations. Skills are task-specific and only load when relevant, making them better for specialized workflows.” ↑b
- 10Claude Help Center, Use plugins in Claude, read at source 13 Oct 2026“Plugins are available to all paid plans (Pro, Max, Team, Enterprise).” and “Each plugin bundles skills, connectors, and sub-agents into a single package” ↑a“Plugins may include local MCP servers that run on your computer with the same permissions as any other program you run.” ↑b
- 11OpenAI, Build skills, read at source 13 Oct 2026“When you want to distribute reusable skills beyond a single repo, or optionally bundle them with connectors, use plugins.” ↑
- 12Remotion, Agent Skills, read at source 13 Oct 2026“Remotion maintains a list of Agent Skills that define best practices for working in Remotion projects.” and, for the remotion-markup skill, “Best practices for writing Remotion React markup.” ↑
- 13Runway, Introducing Runway MCP, 27 May 2026, read at source 13 Oct 2026“add Runway as a custom connector using the Runway MCP server URL and sign in with your Runway account. No separate API key is required – generations are tied to your existing Runway plan.” ↑
- 14ElevenLabs Documentation, Hosted MCP server, read at source 13 Oct 2026“Once connected, an assistant such as Claude can create, configure, and manage the agents in your workspace through natural language” ↑
- 15HeyGen, hyperframes README, read at source 13 Oct 2026HyperFrames is “an open-source framework for turning HTML, CSS, media, and seekable animations into deterministic MP4 videos.”; “HyperFrames ships 21 skills agents load on demand.” and, for Claude Code, “install the versioned plugin”: claude plugin marketplace add heygen-com/hyperframes, then claude plugin install hyperframes@hyperframes. ↑
- 16Claude Code Docs, Plugin security and trust, read at source 13 Oct 2026“A Claude Code plugin you install can execute arbitrary code on your machine with your user privileges.” ↑
Read at source 13 Oct 2026: Anthropic, Skills explained (5 Mar 2026) and Introducing Agent Skills (16 Oct 2025); Claude Help Center, What are skills? and Use plugins in Claude; Claude Code Docs, Extend Claude Code and Plugin security and trust; OpenAI, Build skills; Remotion, Runway, ElevenLabs and HeyGen pages linked above.
The kitchen comparison, the “when to use which” list and the beginner advice are this site’s own, not claims from these sources. The USB-C comparison is the MCP project’s own.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK