AI agent skills are a way to give AI agents procedural knowledge — the step-by-step know-how for specific tasks — using a simple skill.md markdown file format. Each skill has a YAML front matter with a name and description (used as a trigger condition), followed by instructions in plain markdown, plus optional folders for scripts, references, and assets. To handle large numbers of skills efficiently, a three-tier progressive disclosure system loads only metadata at startup, pulls full instructions when a matching task is detected, and loads resources only when needed. Skills complement other agent knowledge mechanisms like MCP (tool access), RAG (factual lookup), and fine-tuning (baked-in model knowledge). The skill.md format is an open standard (Apache 2.0) published at agentskills.io and adopted by platforms including Claude Code and OpenAI Codex, making skills portable across platforms. A key security warning: skills can include executable scripts with file system and API key access, so publicly available skills should be reviewed carefully before use, as audits have found prompt injection, tool poisoning, and malware in the wild.
Sort: