The card shows two objects separated by a not-equal sign. On the left, a folder labeled Agent Skill contains a hatched SKILL.md row and a resources row. Below are three numbered pills linked by arrows: discover, activate, and load files, with callouts for name and description first and instructions when relevant. On the right, an Agent Card document contains one hatched row labeled AgentSkill, with a callout for remote capability listing. The bottom line reads local procedure files are not sent over A2A.
Agent Skills package instructions and supporting files on disk for a host agent. The intended flow is progressive disclosure. At discover, the agent reads only metadata such as name and description to decide relevance. At activate, it loads instructions when the skill applies. At load files, it opens bundled resources only as needed. That keeps context small compared with dumping every skill into the prompt up front. Skills are a local packaging pattern for procedures; they are not a network protocol.
An A2A AgentSkill is a different object with a similar name. It is an entry on a remote Agent Card that advertises an ability the server offers to clients. It describes what the remote agent claims it can do, not the text inside a local SKILL.md folder. Nothing in the Skill directory is automatically transmitted over A2A when a client connects. Tool retrieval cards cover selecting relevant tool definitions from a large catalog; this card resolves the naming collision between local Skills and remote AgentSkill listings.
Use Agent Skills when you author reusable playbooks for an agent you control. Use AgentSkill entries when you publish capabilities to external clients. Do not assume a remote AgentSkill implies the client can read your private skill files, and do not treat a local Skill folder as a substitute for an Agent Card entry when crossing process boundaries.