Boundary guide
Claude Skills vs MCP, Projects, Plugins, and Instructions
A boundary guide for deciding whether a workflow belongs in a Skill, MCP server, project instructions, plugin, or ordinary prompt.
Primary source takeaway: Anthropic documents tools and MCP separately from Skills, so it is useful to keep their jobs separate.
Decision table
| Need | Best fit | Why |
|---|---|---|
| Repeatable procedure with examples, references, or scripts | Skill | The agent needs packaged know-how. |
| Live external system access | MCP or tool | The agent needs a connector or function call. |
| Repository-wide facts and preferences | Project instructions | The agent needs persistent local context. |
| Distribution of commands, hooks, and skills | Plugin | The package needs multiple capabilities. |
| One-off task context | Prompt | The information is not reusable yet. |
Skills vs MCP
Use a skill when the valuable thing is procedure: how to review, write, transform, test, or evaluate. Use MCP when the valuable thing is live capability: querying a system, taking an action, retrieving data, or exposing a tool surface from another process.
They can work together. A skill can tell the agent how to use a particular MCP server safely. But the skill is the guide; the MCP server is the connector.
Skills vs project instructions
Project instructions are always-on background guidance for a repository or workspace. Skills are conditional packages for specialized tasks. If the agent should remember a rule on every turn, it probably belongs in project instructions. If the agent should activate a procedure only for a matching job, it probably belongs in a skill.
A common mistake is placing every team preference into a skill. That makes the skill noisy and weakens its trigger. Put stable project facts where the agent can always see them; put specialized methods in skills.
Skills vs plugins
A plugin can distribute skills, commands, hooks, or integrations. A skill is one kind of capability that may be packaged by a plugin. When a user says "install this plugin," they may be installing a bundle that includes several skills.
For directory purposes, list the skill by the job it performs, not merely by the plugin that ships it. That makes review easier and prevents plugin branding from hiding the actual behavior.
FAQ
Is this site affiliated with Anthropic?
No. Claude Has Skills is an independent editorial guide. It links to Anthropic and Agent Skills primary sources so readers can verify product details directly.
Does a skill replace normal prompting?
No. A skill packages repeatable instructions and resources so the agent can load them when a task matches. The user still asks for a task, and the agent still decides what to read and run.
Can a skill call tools?
A skill can instruct the agent to use available tools or scripts, but the tool capability itself is provided by the environment, API, or connector.
Should company policy live in a skill?
Only if it supports a specialized repeatable workflow. Always-on policy belongs in a more global instruction surface.
Primary Sources
Claude Code Skills docs
Anthropic Docs
Canonical source for Claude Code skill locations, types, and runtime behavior.
Tool use overview
Anthropic Docs
Primary source for function-like tools, distinct from Skills.
Model Context Protocol connector docs
Anthropic Docs
Primary source for MCP as an external connector protocol.
Agent Skills specification
Agent Skills
Portable format reference for SKILL.md and bundled files.
Best practices for skill creators
Agent Skills
Source for scoping, trigger descriptions, and progressive disclosure guidance.