Agent Skills

Anthropic published Agent Skills on 16 October 2025: folders of instructions, scripts and resources that Claude scans during a task and loads only when one matches. It pulls in the minimal files needed rather than the whole set.

Anthropic named four properties. Skills stack together, and Claude coordinates them without human intervention. One format runs on Claude.ai, Claude Code and the API. Loading stays on demand. A skill can also carry executable code, for work where a program is more reliable than generated tokens.

Distribution follows the surface. In Claude Code they install as plugins from the anthropics/skills marketplace, or by hand into ~/.claude/skills. The API adds a `/v1/skills` endpoint and requires the Code Execution Tool, in beta, for the sandboxed environment. skill-creator interviews the user about the workflow, then writes the directory and formats SKILL.md.

Addy Osmani, Shubham Saboo and Sokratis Kartakis made Agent Skills the reference pattern for dynamic context by progressive disclosure in their May 2026 Google whitepaper, set against the always-loaded cost of AGENTS.md, CLAUDE.md and GEMINI.md. Katelyn Lesse at Anthropic pairs skills with MCP: tools, context, and the expertise to use both.

On 6 August 2026 Google joined Amazon, Cursor, Microsoft, OpenAI and Vercel as Core Maintainer of Agent Plugins 1.0.0, a vendor-neutral packaging format that carries one skill per subdirectory of `skills/`. Anthropic publishes both of the components being packaged, and is absent from that maintainer list.