Skills (Agent Skills), CLAUDE.md, and MCP all share the goal of extending Claude for your specific situation, but each solves a different problem. This article lays out what each one is responsible for, when to reach for which, and how the three naturally combine. If you're new to Skills themselves, or to CLAUDE.md or MCP, this site's introductory guides on each cover the basics this comparison builds on.
🟢 Model references match the current lineup · model notice · Fable subscription
Fable 5 and 5.1 subscription (updated September 7, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits (
What Skills are
Skills (Agent Skills) are modules of instructions, metadata, and optional resources (scripts, templates) that Claude loads automatically when it encounters a task they apply to. Official documentation compares them to an onboarding guide you'd hand a new team member — just as a new hire gets a document explaining how things are done, a Skill hands Claude a packaged explanation of how to handle a particular kind of task.
Skills differ from a prompt (a one-off instruction typed directly into a conversation). A prompt has to be retyped every time, while a Skill is pulled in automatically whenever it's relevant, so you don't have to repeat the same guidance. The documented benefits are threefold: specializing Claude for domain-specific tasks, reducing repetition, and composing multiple Skills together for complex, multistep work.
There are two kinds of Skills. Pre-built Agent Skills, provided by Anthropic, cover common document tasks like PowerPoint, Excel, Word, and PDF. Custom Skills are ones you create yourself — in Claude Code, uploaded via the API, or added through claude.ai settings — and they work consistently across Claude's products.
Skills stay lightweight because of progressive disclosure (loading information in stages as it's needed). At the top of a Skill file sits a metadata block with a name and description; only this part is always loaded into the system prompt. The Skill's actual body instructions and any bundled resources load only when the task at hand actually calls for them. This structure is why, according to the documentation, you can install many Skills without consuming much context.
One thing to watch for: that description field is the only basis Claude has for deciding whether to use a given Skill. The documentation is explicit that the description must state both what the Skill does and when to use it. A vague description means the Skill may not get triggered when it's actually needed.
What CLAUDE.md is
CLAUDE.md is a file that holds context and rules meant to apply across an entire project, at all times. Unlike a Skill, which is pulled in selectively only when a matching task comes up, CLAUDE.md is read in as soon as a conversation starts, with no trigger required.
It typically holds things like a project overview, directory structure, coding conventions, and commit or deployment principles — the kind of "always do it this way in this project" guidance. Rather than the task-specific procedures a Skill provides, CLAUDE.md tends to carry the background knowledge and standing principles that should stay consistent across every task in the project. This site's CLAUDE.md guide covers how to write one in more detail.
What MCP is
MCP (Model Context Protocol — an open protocol for connecting external tools and data) is a public communication standard that lets Claude connect to external resources like internal systems, databases, or GitHub. Where Skills and CLAUDE.md deliver instructions about what to do and how, MCP acts as the connection itself, letting Claude actually use outside tools or read and write outside data.
Put another way: if a Skill supplies "knowledge and procedure" and CLAUDE.md supplies "standing context," MCP supplies "new capability." Filing an issue in a tracker, or pulling a document from an internal search system — things Claude couldn't do on its own — are the kind of tasks MCP makes possible. This site's MCP guide covers it in more depth for readers new to the concept.
The core differences
The biggest distinctions between the three come down to when they load and what they provide. The table below summarizes it.
| Skills | CLAUDE.md | MCP | |
|---|---|---|---|
| One-line definition | A packaged set of instructions loaded automatically when relevant | A file of standing project context and rules | A protocol for connecting to external tools and data |
| When it loads | Name/description always; full body only when relevant | Always, from the start of the conversation | Only when the relevant tool or data is needed |
| What it mainly holds | Task-specific instructions, scripts, templates | Project overview, conventions, standing principles | The ability to access and act on external systems |
| Best suited for | Recurring, specific task types (e.g. document work) | Background knowledge shared across all tasks | Access to internal systems, APIs, and other external resources |
| How it's provided | Pre-built or custom Skill | A guide file at the project root | Connected via an MCP server |
When to use which
The deciding question is whether the content is always needed, needed only for a specific task, or requires reaching an external resource. If it's a project rule that should apply to every conversation, use CLAUDE.md. If it's a specialized procedure or template needed only for a specific kind of task, use a Skill. If you need direct access to data or tools in an external system, use MCP.
An example of using all three together
The three tools aren't mutually exclusive — they're designed to combine. Consider a workflow that ties into an internal issue tracker to write up bug reports. CLAUDE.md might hold a standing principle like "bug reports always list reproduction steps, expected result, and actual result in that order." The detailed procedure for formatting that report into a specific document type (say, a Word document with a table) is handled by a Skill, pulled in automatically when relevant. And the actual work of looking up or filing the issue in the tracker is handled by a tool connected through MCP. In this kind of combination, CLAUDE.md carries the standing principles, the Skill carries the procedure, and MCP carries the ability to act.
Frequently asked questions
Q. Do Skills and CLAUDE.md conflict with each other?
Not by default — their roles are different enough that using them together is the natural pattern. CLAUDE.md provides background context that applies throughout the conversation, while a Skill is pulled in additionally only when a matching task comes up, to handle that task's specialized procedure.
Q. Does installing many Skills waste context?
According to the documentation, not significantly. Only a Skill's metadata — its name and description — is always loaded; the full body of instructions loads only when the relevant task actually comes up, thanks to the progressive disclosure structure.
Q. Does connecting MCP make Skills unnecessary?
No — they solve different problems. MCP only provides the connection that lets Claude reach external tools or data; the specialized instructions for how to actually use that connection can still live in a Skill.