When Claude Skills was released Simon Willison said they might be a bigger deal than MCP. Willison is the co-creator of the Django framework and astute observer of the AI landscape. I'm inclined to agree with him on skills.
Skills are a unit of context for agents. Generally they are instructions or documentation for how to perform a specific task. For example, Anthropic maintains a “Frontend Design Skill” that dramatically improves the quality of Claude’s design. Remotion, a library for creating videos, recently published Remotion Skills that enables agents effectively to create videos with React.
At the most basic level, skills are markdown files that follow a specific format. Each skill lives in a folder named for the skill and has a SKILL.md file. For Claude this looks like .claude/skills/skill-name/SKILL.md. SKILL.md requires a few specific frontmatter elements but is otherwise completely open. You can add additional markdown files or code within the folder as well. SKILL.md is the entry point for the agent but it can leverage everything in the folder.
The core of skills are pieces of context that are primed for repeat usage. You define the context once and then agents can read that context when appropriate. This concept is not particularly new. Personally I've had Raycast snippets since at least 2024 to serve a similar purpose - context to repeatedly give agents.
Skills are important and valuable because they have a standard location and basic formatting. This enables Skills to be shared, within a team or publicly, and enables them to be used by all agents. Vercel has released a skills CLI that enables anyone to publish skills and anyone to install these community skills.
I cheekily say I’ve been using Skills in Cursor since 2024, but this portability and sharability is truly what makes them different now. Previously, I had all of these “units of context” scattered across Raycast Snippets, a personal prompt library, and markdown directly in repos. Having a defined, preset location removes a lot of friction from adoption.
MCP Servers are one way to make data from external systems available to an agent. Skills can be used in conjunction with MCP Servers. MCP Servers can be used with or without Skills to guide the agent on how to use them. I think of Skills as the higher level instructions. MCP Servers are a gateway to external data. But the boundaries are fuzzy.
Skills unlock a form of continual learning for agents. Whenever an agent is using a Skill to complete a task you can have the agent reflect on what went well and what went poorly and then update the corresponding Skill. If the agent oneshot the task the skill is probably fine. But if the agent required additional prompting and guidance from you, those are literally teaching moments. The fundamentals of having skills in predefined locations makes it easier to close the loop. The whole flow is a few prompts: “Use X Skill to do Y task” “Now that you are done, reflect on your task completion and update X Skill with any relevant new learnings” Of course, as with all writing, you might end up with a bit of slop if you don't prune the agent’s writing. But basically this is enough to unlock continuous improvement out of your agent for any task!
At Firstloop the way we are using Skills is evolving. We’ve consolidated our skills into a single repo. This repo is broken up into multiple plugins like “core” and “monorepo template”. Core contains skills for: Product Management and Skill Management and the Monorepo Template skill contains skills related to building with our in-house template. The Skill Management Skill contains two types of instructions. One for creating Skills locally within a project, and one for creating/updating the Skills in our Skills Repo. It’s getting meta! The Meta Skill Management contains instructions for how to manage the skills repo, how to update skills and how to increase their version numbers. This means that while we are working we can have Claude fire off a sub agent to make an improvement to a skill. This results in a GitHub issue being created in the Skills repo (which we then assign to Claude to implement).
We’re still in the early innings of using this approach but it’s already starting to pay big dividends. Our own personal learnings can be documented and automatically shared across the team. Any productivity improvement for one of us is instantly shared across all.
Get insights on AI development and behind-the-scenes updates from our team.