Agent Skills
An open, folder-based format for packaging procedural knowledge an AI agent loads on demand — instructions, scripts, and resources in a SKILL.md the agent reads only when a task calls for it.
Production AI is a system: context, tools, permissions, traces, evals, and feedback loops around the model.
What it is
Agent Skills is an open format introduced by Anthropic in 2025: a skill is a folder with a SKILL.md file (name, description, instructions) plus optional scripts and resources. The agent sees only the name and one-line description until a task matches; then it loads the full instructions. This progressive disclosure keeps the context window lean while the skill library grows.
Why it matters
Prompts scattered across application code are hard to version, review, or share. A skill is a reviewable artifact: it lives in a repository, goes through pull requests, gets eval cases, and ports across any agent that supports the format.
How we use it
Our production stack manages agent capabilities as folder-shaped skill packages — markdown instructions alongside tool declarations, schemas, and eval cases — synced from the repository into the runtime with content-hash change detection. Public product sites publish skill indexes under /.well-known/skills so external agents can discover what a service offers.
Related resources
How an educated AI agent — the one with the codebase, the tools, and the tacit context that lets it succeed — distills its competence into transferable skill documents a virgin agent can run from scratch.
A structured view of what an AI system can actually do — what skills it has, what tools it can call, what data it can read, what decisions it can make — so the team and the users share the same model of its scope.
Anthropic's open standard for exposing tools, resources, and prompts to AI models — released in late 2024, broadly adopted across the agent ecosystem, the connective tissue of modern AI tool integration.
The discipline of deciding what an AI model sees on every call — instructions, retrieved data, memory, tool definitions, examples — and how to assemble them reliably as the workflow grows.