Skills: reusable capabilities
A skill’s files cooperate toward one useful result. Different workflows can reuse that result.
Start with a task and a result
A video-frame skill accepts a local video, an interval, and an output directory. It returns screenshots with timestamps. It does not also summarize the video, critique the UI, or write a storyboard. Those are workflows that can use this capability.
The bundle is one cooperating unit
SKILL.md tells the agent when the capability fits, what it needs, and how to run it. Bundled code performs mechanical work. Templates and assets can support that method. For video extraction, the guide supplies the method, the script produces the images, and the tests check the expected frames and failure cases.
The agent drives a guide-based skill
The loader discovers the guide, and run_skill prepares its environment and returns instructions with bundle paths substituted. The current agent then follows the method using its tools. A guide can direct the agent to run Python or other bundled code through the shell; that is different from declaring an executable entrypoint.
Reuse the result in different workflows
A video summary uses the extracted frames to identify important moments. A visual review inspects them for UI defects. A storyboard arranges them into a sequence. Each workflow uses the same extraction capability instead of carrying its own version of that code.
Keep the capability focused
The video-frame bundle defines its inputs, extracts the images, and returns a timestamped manifest. Summarizing or evaluating those images belongs to the caller. That separation is what lets the skill serve more than one workflow.