A Skill in 📝Claude Code is a file-based capability with YAML frontmatter that the harness loads on demand, exposing reusable workflows the agent can invoke by name when their description matches the task.
Skills live as markdown files in ~/.claude/skills/ (or as part of a 📝Claude Code Plugin) with frontmatter declaring a name, description, and optional metadata. The harness reads each skill's frontmatter at session start, builds a catalog, and surfaces relevant skills to the model when descriptions match the task. Invocation happens two ways: explicit, via typing /skillname as a 📝Slash Command, or implicit, when the model calls the Skill tool because it judges the skill applies.
When triggered, the skill's full content is injected into the conversation as instructions — system-prompt-level guidance scoping how the model approaches the task. This decouples reusable workflow patterns (deploy checks, PR reviews, brainstorming sessions, design audits) from the conversations that use them. The user packages each workflow once, then invokes it by name from any future session — no re-explanation required.
Skills handle the judgment layer of customization — workflows that need deliberate triggering when context calls for them. The complement is the 📝Hook (Claude Code), which handles the mechanical layer, firing on lifecycle events without user action. See 📝Claude Code Hooks and Skills for the side-by-side comparison and code examples.