Skip to content

Agent Skills

Summary

A minimal form factor for packaging procedural knowledge that agents can dynamically load. Organized folders containing a SKILL.md instruction file, optional scripts as tools, templates, and reference materials.

Origins

Introduced by Barry Zhang and Mahesh Murag at Anthropic in December 2025. The industry quickly converged on this standard — adopted by Claude Code, Pi, Tolgee, and others.

What Skills Are

Skills are folders. That simplicity is deliberate: - SKILL.md — core instruction file (markdown) - Scripts (Python, bash, etc.) — executable tools the agent can run - Templates — fill-in-the-blank documents - Reference docs — examples, documentation, guidelines

Progressive Disclosure

The key innovation that makes skills scale: 1. At startup, agent sees only metadata (name + short description) — like reading book titles on a spine 2. When agent decides it needs a skill, it reads the full SKILL.md 3. Rest of folder is organized for ease of access 4. Allows hundreds or thousands of skills without context overload

This contrasts with prompts, where everything is crammed into context at once and gets messy fast.

Types of Skills

Type Creator Purpose
Foundational Platform builders New general/domain capabilities (document creation, scientific analysis)
Third-party Partners Product-specific skills (Notion workspace, Browserbase automation)
Enterprise Organizations Company best practices, internal workflows, code style guides

Skills vs MCP

MCP Skills
Role Connection to outside world Expertise and procedural knowledge
Analogy The hands The experience
What it does Pull data from APIs, connect to tools Tell the agent what to do with the data
Together Agent that can connect to anything AND knows how to handle what it finds

The Emerging Agent Architecture

Model (CPU) ↓ Agent Runtime (OS — orchestrates context, tokens in/out) ↓ MCP Servers (Network/IO — external tools and data) ↓ Library of Skills (Applications — domain expertise)

Future: Agent-Created Skills

Skills are designed as a concrete step toward continuous learning: - Anything the agent writes down can be used efficiently by future versions - Already possible: Claude creates skills via "Skill Creator" - Goal: Claude on day 30 is much better than Claude on day 1 - Vision: collective, evolving knowledge base curated by people and agents

See Also