Agent Skills are folders of instructions, scripts, and resources that agents can discover and use to do things more accurately and efficiently. Each folder has a SKILL.md file which is the main brains - this has metadata and instructions that tell an agent how to perform a specific task. Additional folders and files can be used for scripts, references, etc.
Distilling it down really far, skills (especially just the SKILL.md) file is just a saved prompt.
It's an open standard, meaning the goal is to be supported by any and all AI development tools, rather than bespoke. The goal is to make reusable instructions that are tailored towards a specific task, giving agents new capabilities and repeatability.
A main advantage over other [[20260306121823-ai-instructions]] or [[20250811104331-model-context-protocol-mcp]] approaches is that skills use Progressive Disclosure to better manage (read: reduce) context costs [[20260318091849-ai-context-tokens]]. Agents load only the name and description of each skill. It's recommended to split up large skill files into references or similar directories, and then tell the agent when to load them. https://agentskills.io/skill-creation/best-practices#structure-large-skills-with-progressive-disclosure
Agents load skills through progressive disclosure, in three stages:
SKILL.md instructions into context.When crafting a skill, ensure that the skills contain domain-specific context. Avoid generic and vague instructions. Generate skills from real world artefacts and knowledge.
Skills are similar to tools in [[20250811104331-model-context-protocol-mcp]], but tools are more defined capabilities. Think of like an API request, or a specific function. Skills are more guidance on how to do something.