Source: Equipping agents for the real world with Agent Skills (Anthropic Engineering)
Anthropic's launch essay for Agent Skills: folders of instructions, scripts and resources loaded by progressive disclosure, later published as an open standard.
Contents
- What: engineering blog post
- Publisher: Anthropic (Barry Zhang, Keith Lazuka, Mahesh Murag)
- Date: 2025-10-16; updated 2025-12-18 with the open-standard announcement
- Original: Equipping agents for the real world with Agent Skills
Summary
Introduces Agent Skills as a way to turn a general-purpose agent into a specialist by giving it an onboarding-guide-like folder it can discover and read on demand. The design relies on the agent having a filesystem and code execution, so a skill can bundle far more material than fits in context and load only what the task needs.
Key claims
- A skill is a directory whose core file is
SKILL.md; YAML frontmatter must carrynameanddescription, the body holds instructions. - Three disclosure levels: (1) every installed skill’s name and description is preloaded into the system prompt; (2) the full
SKILL.mdis read when the skill looks relevant; (3) linked files are read only if needed. - Because files are read on demand, the amount of bundled context is described as effectively unbounded.
- Skills can ship scripts the agent runs as tools; code gives deterministic, cheaper results than token generation for tasks like sorting or form-field extraction. Instructions should say whether a script is to be run or read.
- Authoring advice: start from evaluations that reveal capability gaps, split large files, keep mutually exclusive contexts in separate files, watch real usage, and pay most attention to
name/descriptionbecause they drive triggering. - Security: malicious skills can introduce vulnerabilities or exfiltrate data; install only from trusted sources and audit scripts, dependencies and any instructions to contact external networks.
- 2025-12-18 update: Agent Skills published as an open standard (agentskills.io) for cross-platform portability; supported in Claude.ai, Claude Code, the Agent SDK and the Developer Platform.
- Stated future direction: complementing MCP servers, and agents that create and evaluate their own skills.