GitHub's November 12 guide presents agent skills as reusable instructions for repeatable engineering work. A skill is a versioned directory centered on a `SKILL.md` file that describes when a task applies and how to carry it out. GitHub's example keeps one set of skills as a source and packages client-specific files for Claude Code, Cursor and GitHub Copilot. The key caveat is in the guide's own summary: reusable instructions do not make every client share the same tools, project context or permissions.
What can move with a skill?
A skill can preserve a bounded workflow: the task it handles, its inputs and expected outputs, checks to run, references to consult, and where a human should review the result. In GitHub's Data Forge example, several skills share project facts through a manifest instead of hard-coding the same database and naming details in each instruction file. A packaging script then creates the client-specific files needed by the example setup.
That is different from saying one file is automatically understood by every agent. GitHub describes a packaging approach for its example, and its Copilot documentation defines which locations Copilot will load. Its Agent Plugins documentation also describes a portable plugin format for compatible surfaces, with `skills/NAME/SKILL.md` as the skills location. Compatibility depends on the client and plugin format being used.
What does the host still have to provide?
Instructions cannot install a database client, supply a project manifest, authenticate to a repository or grant an agent permission to alter a system. The running environment provides those tools, credentials, network access and authorization. A skill that explains how to query a database does not, by itself, create a safe or valid database connection. The project-specific values that the workflow needs should be supplied and reviewed by the host rather than assumed from the reusable instructions.
GitHub recommends verifying each installed skill in the intended client: confirm that it recognizes the skill for the right task, loads the referenced files, can access the required tools, and produces the requested evidence. For consequential changes, the guide's examples use checks and human approval points. Instructions can guide an agent, but required merge blocks and permissions need to be enforced by the repository and host system.
GitHub guide: SKILL.md structure, client-specific packaging and host-provided capabilities
GitHub Copilot documentation: skill setup and loading behavior
GitHub Agent Plugins documentation: portable skill paths for compatible surfaces
A useful portability test
To move a skill between clients, start with the job rather than the client configuration. Keep the instructions and review criteria in a clear source directory, place project-specific values in a manifest, then package only the files each supported client expects. In each client, check tool availability, credential scope, referenced-file loading and the evidence produced. Keep approval rules outside the skill where the host can enforce them.
GitHub's guide is an example of a repeatable workflow being made easier to share. It does not establish that all agent clients use identical formats or behavior. For developers, the practical benefit is reducing duplicated instructions while keeping each host's permissions and setup visible.