Skills: teaching AI agents to act consistently

Skills or Agent Skills, initially introduced by Anthropic, are a concept within the AI ecosystem that enhance the capabilities of AI agents by teaching them to perform specific tasks consistently. They allow users to write instructions once, which the AI follows automatically, thus eliminating the need for repetitive prompts. A skill comprises a folder containing a SKILL.md file with metadata and detailed instructions, and may include additional resources like scripts, references, and assets to support complex tasks. Agents use a three-step process of scanning, matching, and loading to efficiently utilize skills without overloading memory. To address the common issue of AI agents not reliably triggering skills, an AGENTS.md file can be used to provide global context and ensure retrieval-led reasoning over pre-training-led reasoning. Skills are considered foundational for building reliable, repeatable behaviors in systems powered by large language models, as they promote consistency and reduce errors, making them indispensable for those seeking predictable and evolving AI agents.