MCP vs. Skills: IBM Breaks Down When to Use Each for Building AI Agents

An LLM is essentially a prediction engine trained on vast amounts of information, capable of answering many kinds of questions -- but getting the right answer depends on giving it the right context. Providing just a role and a task is prompt engineering; adding extra information, like how you want data formatted or how your team's database is configured, is context engineering. Understanding that distinction is the foundation for knowing when to reach for MCP versus Skills.

Using an agent that needs CRM data as the example, the video explains how MCP works. Instead of pasting a service's raw API docs and a token into the LLM and hoping for the best, MCP standardizes how an AI model talks to different data sources -- abstracting a service's API into a simple, LLM-ready format and handling authentication. An MCP server gets wired into your IDE or AI application, has the LLM generate a JSON request for the information it needs, and translates that into the actual POST or GET request to the service. It's a standardized layer supported by nearly every AI tool out there.

MCP solves the problem of feeding external data to an LLM, but leaves one gap: how do you give it domain knowledge it doesn't already have -- for example, a sales team wanting CRM data formatted the exact same way every time, down to a customer's name, contact info, and even their favorite type of cookie. Since LLMs are inherently non-deterministic, reliably repeating a format like that is genuinely hard.

Skills are essentially markdown files with metadata -- a title, a description of when to use it, and the actual prompt passed to the LLM -- packaged into a folder, auto-loaded into the context window only when needed. Use MCP for situations requiring real-time, tightly-permissioned data access (checking VM status, for example); use Skills when you need a lightweight, reusable custom capability (fetching and analyzing investment data, for example). Both are open source, broadly supported across today's AI tools, and usable locally starting today.