Skill (Claude Skills)

A Claude Skill is a folder containing a SKILL.md file (and optional supporting files) that defines a named capability Claude can invoke. The SKILL.md has YAML frontmatter with metadata — name, description, when to use it — and a markdown body with instructions, examples, and links to helpers. Drop it in ~/.claude/skills/ and Claude can use it. Skills load by progressive disclosure, in three stages. At startup Claude reads only each skill's name and description from the YAML frontmatter, which costs on the order of tens of tokens per skill. When a request matches that description, Claude loads the full SKILL.md body into context. Only then, while following the instructions, does it open any bundled reference files or scripts the skill points to. This is why a skill's description field matters more than anything else in the file: it is the only part Claude sees when deciding whether the skill applies, so it has to state both what the skill does and when to use it. It is also why you can install many skills without paying a context penalty for the ones you are not using. Think of them as discrete, shareable units of capability that do not require API keys, fine-tuning, or running a server. The format is no longer Claude-specific. Anthropic created it and released it as Agent Skills, an open standard specified at agentskills.io, whose overview page lists more than 40 compatible agents and products as of September 2026, among them VS Code, GitHub Copilot, Google's Gemini CLI, OpenAI's Codex, Cursor, JetBrains Junie, Amp, Roo Code, and Block's Goose. A SKILL.md folder written for Claude Code can be reused in any of them. Compared to MCP servers, which expose tools and require running code, a skill is a folder of markdown plus optional scripts, so it is easier to author, share, and version-control. In practice teams use skills for: codifying internal style guides ("write a PR description like our team does"), wrapping common workflows ("triage a bug report"), and embedding domain knowledge ("explain this FERC regulation"). They are a low-friction way to make an LLM consistently good at something specific.