How Claude’s New ‘Skills’ Transform AI Agents and Outpace MCP

What is a Skill

Skills give a large model a specific ability, similar to the MCP concept originally proposed by Anthropic.

For example, to check tomorrow’s weather the model must perform an online search; a Skill can encapsulate this as a Python method that queries a weather site.

Claude’s settings include several built‑in Skills.

A Skill is a set of instructions stored in a folder containing a SKILL.md file and any supporting files such as Python scripts. When an AI agent tries to solve a problem it references these files.

The Skill.md file must start with a YAML front‑matter that includes name and description metadata. At startup the agent pre‑loads each installed Skill’s name and description into its system prompt.

Additional markdown files can be referenced from Skill.md, allowing markdown to play a major role.

Skill vs. MCP

Both aim to provide tools, but they differ mainly in how they manage the context window.

Connecting three MCP servers can consume about 16 % of the context window (~32,000 tokens) because each server loads all its tool descriptions.

Sub‑agents isolate their context from the main agent, returning only final results without sharing intermediate context.

Claude’s Skills rely on the SKILL.md file, which contains a system prompt describing the Skill and any associated tools.

Key differences:

Skills are specialized for repeatable workflows, explicitly telling the AI how and in what order to use tools.

When multiple Skills are available, only their lightweight metadata (≈100‑150 tokens each) is loaded initially, unlike MCP which loads full tool descriptions.

Upon a user request, the system selects relevant Skills by similarity, ignores unrelated ones, and loads the full content of the chosen Skill only when needed.

Anthropic calls this “progressive disclosure,” meaning Claude loads information only when required, making Skills flexible and scalable.

Potential of Skills

The approach could become an industry standard, following the earlier MCP concept, but its adoption will depend on broader acceptance.