Deployment & Installation

Once you have authored your skills, the final step is making them available to AI agents.

Output Modes and CLI Behaviors

The way you configure the outputMode in your astro.config.ts dictates what installation methods are available to consumers. The skills CLI has internal specifications for resolving skills: it looks for a well-known endpoint if given a web URL, and looks for a root folder if given a code repository.

Public URLs (well-known-endpoint)

If your mode is set to well-known-endpoint (or both), the plugin generates skills into dist/.well-known/skills/.

• Deployment Requirement: You must deploy your dist folder to a public static hosting provider (e.g., GitHub Pages, Vercel, Cloudflare Pages) so that it is openly accessible on the internet.
• Installation Method: Provide the live URL to the skills CLI. The CLI internally understands that a regular URL implies it should check the .well-known/skills endpoint on that domain.

npx skills add https://your-starlight-site.example.com

Source Repositories (collocated)

If your mode is set to collocated (or both), the plugin generates skills into a skills/ directory and commits them to your repository.

• Deployment Requirement: You must commit the generated skills/ folder directly into your Git repository (e.g., pushed to GitHub). You do not necessarily need to host or deploy the actual website.
• Installation Method: Provide the repository identifier or repository URL. The CLI internally understands that a repository reference implies it should check the root skills/ directory of the codebase.

npx skills add your-name/your-docs-repo

The output directory defaults to skills/ relative to the Astro project root. In a monorepo, the Astro package is typically a subdirectory, so you will want to point the output to the repository root instead:

starlightSkills({
  collocatedDir: '../../skills', // relative to the Astro package, resolves to the repo root
})

Output File Format

Each skill is written as a SKILL.md file — a Markdown file with a YAML frontmatter block that conforms to the Agent Skills specification.

skills/
└── react-components/
    ├── SKILL.md               ← skill instructions + metadata
    └── schemas/
        └── component.json     ← companion files (if any subfiles are defined)

The SKILL.md frontmatter mirrors the skill’s frontmatter fields (name, description, license, compatibility, etc.), while Starlight-specific fields (skill, skill-name, title, etc.) are stripped. The body is forwarded verbatim from the source Markdown page.

For the well-known-endpoint mode, an additional index.json is written at dist/.well-known/skills/index.json listing all skills and their files, enabling automated CLI discovery.