Origins
In AI, people often use skill / capability to mean “what a model or agent can do.”
Here, we need to distinguish general “capabilities” from Agent Skills, which have a specific organization and loading mechanism.
For example, an Agent Skill usually lives as its own directory, with a structure like this:
my-skill/
├── SKILL.md
├── scripts/
├── references/
└── assets/SKILL.md describes what the Skill is for, how to use it, and the instructions to follow. Scripts, references, and assets can work alongside it to complete a specific task.
So Agent Skills here do not mean “whatever an agent can do” in general. They mean a reusable capability package with a clear file structure that can be installed, discovered, and loaded on demand.
Equipping agents
In Anthropic’s October 16, 2025 article Equipping agents for the real world with Agent Skills, they noted:
As models keep improving, general-purpose agents can already interact with full computing environments. To equip domain-specific knowledge in a way that is more:
- composable
- scalable
- portable
they introduced Agent Skills.
Agent Skills organize instructions, scripts, and resources into structured folders that agents can dynamically discover and load when working on specific tasks, improving performance.
By packaging specialized knowledge into composable resources for Claude / AI agents, Skills extend what those agents can do and turn general-purpose agents into specialized ones for particular needs.
On December 18, 2025, Anthropic published Agent Skills as an open standard, aiming for the same Skill to work across platforms.
In January 2026, Vercel launched the skills CLI and skills.sh, further addressing:
- how to find Skills;
- how to install Skills;
- how to distribute Skills;
- how to update Skills;
Rough timeline:
2025-10
Anthropic introduces Agent Skills
↓
2025-12
Agent Skills → open standard
↓
SKILL.md becomes the core format
↓
2026-01
Vercel launches skills CLI + skills.sh
↓
More agents add support
↓
Today’s Agent Skills ecosystem takes shapeDefinition
An Agent Skill is a discoverable, on-demand, reusable package of procedural knowledge for AI agents.
The core problem it solves: modularize specialized knowledge and load it on demand.