Sending a message involves more than choosing a send function. An agent also needs a sender, permission to send and a way to check the result.
A skill can describe that procedure so the agent has instructions for the whole task.
What does a skill contain?
A skill contains a SKILL.md file with metadata and instructions. It can include supporting files for the task.
The Agent Skills specification requires a name and description in the file's frontmatter. The description tells the host when the skill applies. The Markdown body carries the procedure.
Supporting directories can hold executable scripts, reference material and assets such as templates. They let the agent fetch detailed material when the task needs it.
For example, a message-sending skill could describe how to select a sender and check delivery. A reference could explain the response fields. The skill's instructions would connect those steps into one task.
How does the agent load it?
The host discovers the skill's metadata, then loads its instructions when the task matches.
The specification calls this progressive disclosure. The host can load supporting files later, keeping unrelated material out of the initial context.
Anthropic's skill overview illustrates this with separate levels for metadata, instructions and supporting resources. A host must implement skill discovery and loading for this behavior to happen. Placing a file on disk alone does not establish that an agent used it.
How is a skill different from a tool?
A skill describes a procedure. A tool exposes an action the agent's application can execute.
| Component | Role in a message-sending task |
|---|---|
| Skill | Describes the prerequisites, send step and outcome check |
| Tool | Sends a message or retrieves its recorded status |
| MCP server | Exposes tools and other capabilities through a protocol |
The MCP specification defines the client-server interface. A skill can teach an agent how to use that interface or a CLI.
Instructions do not issue credentials. The application still needs permission to execute the requested action. A successful tool response also needs interpretation: acceptance and delivery describe different outcomes.
How do I use skills with Bird?
You can install Bird's agent skills to give a coding agent procedures for the Bird CLI.
The general skill instructs the agent to check authentication before workspace operations. Its message workflow includes selecting a sender, sending and reading the resulting message back. The agent needs that result before it can report delivery.
Skills and credentials have separate installation steps. Installing a plugin cannot grant access to your workspace. Complete the documented sign-in before asking the agent to act.
For a host that invokes tools, connect Bird's MCP server. For a shell-capable agent, the CLI guide describes the command output the skill uses to check each step.