What are Claude Skills?

Definition

Skills were introduced by Anthropic so that domain experts can capture how a job should be done once, then hand that procedure to anyone using Claude. A skill is not a model fine-tune and not a tool call: it is a Markdown contract Claude reads at runtime. The `SKILL.md` file declares a name, a description (the trigger), and a body of instructions. Optional companion files — examples, scripts, schemas, prompts — sit alongside it. At inference time Claude scans the descriptions of every available skill, chooses the best match for the current task, and follows the contract. That makes skills portable, reviewable, and version-controllable in ways that prompt engineering alone is not.

A short history

Anthropic shipped skills as a first-class capability in October 2025 alongside Claude Code, then extended the format to the Claude apps and the Claude Agent SDK. The motivation was practical: prompts grew long, examples became unwieldy, and teams kept reinventing the same recipes. Skills standardised the unit of reusable expertise. Within months a public ecosystem appeared — open-source skill libraries, vendor-published skills, and aggregators like the ElasticFlow catalogue. The format remains intentionally simple: a single Markdown file with frontmatter is the minimum viable skill, and that simplicity is what allows the ecosystem to grow.

Skills, plugins, and MCP servers

Skills are one of three composable primitives Anthropic exposes for extending Claude. The other two — plugins and MCP servers — are often confused with skills because all three appear in marketplaces and all three change what Claude can do. They sit at different layers of the stack.

Plugins

A plugin is a bundle. It groups one or more skills (and sometimes commands, hooks, and configuration) under a single installable unit so a user can run one command and get a coherent set of capabilities. If a skill is one recipe, a plugin is the cookbook. Plugins are how publishers ship a domain — a plugin called "sales" might bundle skills for forecasting, pipeline review, call summaries, and outreach drafting. Installing the plugin installs every skill inside it; uninstalling removes them together.

MCP servers

An MCP server is a separate process Claude can talk to over the Model Context Protocol. It lives outside the model. It exposes tools (functions Claude can call), resources (data Claude can read), and prompts (templates Claude can use). Skills tell Claude what to do; MCP servers give Claude the side-effects to do it with — querying a database, hitting a SaaS API, reading a file, sending an email. Many real workflows pair the two: a skill describes the procedure, and inside that procedure Claude calls MCP tools to act on the world. The catalogue lists skills, plugins, and MCP servers separately for that reason.

Rule of thumb: a skill is a procedure, a plugin is a bundle of procedures, an MCP server is a connector to the outside world. You can ship a skill without an MCP server, and you can ship an MCP server without a skill, but most production workflows use both.

Why skills matter

Skills move expertise from prompt-of-the-day into reviewable artefacts. A team can pin a skill to a git commit, audit it in a pull request, and roll it back if a new version regresses. Authors get attribution; consumers get reproducibility; security teams get a place to apply policy. For SEO and discovery, skills also create a long tail of specific, indexable capabilities — "forecast accuracy review" is a skill someone can find, install, and run, in a way "good general prompt for forecasting" is not. The catalogue exists to surface that long tail.