Most skills need only a name, a description and good instructions. Skill Builder keeps everything else in an Advanced section at the bottom of a skill’s details, closed by default. Its heading lists what is inside and shows how many settings or tests are in use, so nothing set is hidden without a sign. This guide explains each setting and who reads it.
What is under Advanced
| Setting | What it does | Read by |
|---|---|---|
| Metadata | Extra key: value details, such as author or version |
Any tool (ignored if not known) |
| Argument hint | Text shown after /name to say what to type |
Claude Code, Copilot |
| Run only when asked | disable-model-invocation: the skill runs only when you call it |
Claude Code, Copilot |
| Show as a command | user-invocable: list it in the / menu |
Claude Code, Copilot |
| Other fields from a file | Fields Skill Builder does not edit, kept from an imported file | Whatever tool reads that field |
| Tests | Requests and tasks to try the skill with (guide) | Your AI tool, through Connect AI |
The first four are written to the frontmatter of SKILL.md. Tests are exported separately as evals/.
Metadata
The open Agent Skills specification has a metadata field for anything else you want to record: an author, a version, an owner. Write one key: value per line. Tools that do not know a key ignore it, so it is safe in any tool.
Metadata does not change how the skill behaves. Use it to keep track of skills in a team.
Argument hint
Claude Code and Copilot show the hint after the command name in the / menu, for example /review-pr [number]. It says what to type; it does not check it. Use $ARGUMENTS in the body to read what was typed (see slash commands as skills).
Run only when asked
Turn this on for a skill with side effects, such as deploying or sending a message. The AI tool then never picks the skill by itself: it runs only when you type /name. It is written as disable-model-invocation: true.
Leave it off for knowledge or routines the AI should reach for on its own.
Show as a command
On by default. Turn it off for background knowledge that the AI loads when a request matches, but that makes no sense to call by hand. It is written as user-invocable: false, and only when it differs from the default.
Turning off both toggles would leave a skill nothing can start, so keep at least one on.
Other fields from the file
When you import a skill, fields that Skill Builder has no input for are kept and written back unchanged instead of being dropped. They are listed here, and you can remove one to leave it out of the export.
Which tools read these
Only Claude Code and GitHub Copilot read the argument hint and the two toggles. Other tools ignore them, but some (claude.ai uploads and the Skills API) refuse a file with fields outside the specification. The checks note this, so look at Checks before you upload a skill there.
Connect AI and slash commands
These settings are written to the file and apply once the skill is installed in a repository for a tool that reads them. They do not change how Connect AI serves the skill, and Connect AI does not rely on / commands for skills in your project.