Anthropic's implementation of the Agent Skills pattern: a directory-based system where each skill is a folder containing a SKILL.md instruction file, optional scripts, and reference documents. The model loads a short index of available skills into its context, then loads a specific skill's full contents only when triggered. Skills work across Claude apps (Claude Code, Claude Agent SDK, Claude.ai) and are portable between projects.
Frequently asked questions
How do Claude Skills differ from custom system prompts?
A custom system prompt is one fixed block of instructions loaded on every call—great when behavior is the same across all tasks. Skills are conditional: dozens of focused instruction sets, only loaded when the model decides one applies. For an agent that handles 10 distinct task types, a skill per type is dramatically cheaper and more reliable than a 10× longer system prompt that covers them all up front.