AI glossary · Basic · Agents
Skills
Skills (Agent Skills) are 'work manual bundles' that package instructions, scripts and reference materials into a single folder to help an AI agent do a specific task well. The agent pulls out and reads a bundle only when the task needs it.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Perception & action
- Fields
- Agents
- Added
- 2026-10-08
What it is
A skill is a folder centered on a single document file called SKILL.md. SKILL.md contains at minimum the skill's name and description, followed by instructions on how to do the task. If needed, materials such as executable scripts, reference documents and templates are included as well. Anthropic first introduced it as a feature for Claude and later released it as an open standard that can be used in tools from multiple companies.
A skill folder at a glance
- my-skill/One skill = one folder
- SKILL.mdRequiredName · description + task instructions
- Name · descriptionOnly this is read normally
- InstructionsRead when the task matches
- scripts/OptionalPrograms that can be run directly
- references/OptionalReference documents
- assets/OptionalTemplates
How it works
The key idea is 'reading only as much as needed, a little at a time (progressive disclosure)'. First, at startup the agent reads only the names and descriptions of all skills (discovery). When a user's task matches a skill's description, it then loads that skill's full SKILL.md (activation). Next, it follows the instructions to do the task, running the bundled scripts or opening additional reference files as needed (execution).
Why it is getting attention
The amount of text an AI model can hold and work with at once (the context window) is limited. Because skills normally take up only their name and description, keeping many skills available adds little burden. Company- or team-specific procedures and know-how can also be bundled into a single folder, version-controlled and reused. Another commonly cited advantage is that a skill built once can be used as-is in multiple agent products that support the same format. Because they are simple to create, some say even non-developers can easily edit and improve them.
Things to know
A skill does not retrain the model itself; it is only instructions and materials loaded when needed. So how a skill's description is written affects whether the agent pulls it out at the right time. Skills that contain scripts work properly only when there is an environment where code can be run safely. Skills obtained from outside can contain executable code, so it is best to check their source before using them.
Example
The feature in the Claude app that creates Excel, PowerPoint and Word files is implemented with skills, and users can also create and add their own skills, such as their organization's brand guidelines. Among the news covered on this site, AWS released 'aws-ai-ml', a SageMaker inference optimization skill for coding agents. When a coding agent loads this skill, it can handle model deployment tasks by following the procedures and tools AWS has put together. OpenAI's Codex developer documentation also has a separate guide on how to create skills.
Where it came up on this site
- Papers EviSkill: evolving agent skills based on re-executable evidence 2026-10-08
- briefing AWS releases 'aws-ai-ml', a SageMaker inference optimization skill for coding agents 2026-10-06