Starlight Skills Companion Files

When authoring a skill, there is often a limit to how much information should be written directly into the main .md document.

If an AI agent needs to understand a massive API schema, a JSON configuration format, or an entire script template, pasting hundreds of lines of code into your main document makes it incredibly difficult for a human to read.

The subfiles Field

A skill can possess supplementary text files. You define these via the subfiles array inside your frontmatter using glob patterns.

---
title: System Configuration Skill
description: Instructions on how to update our internal system configuration.
skill: true
subfiles:
  - "schemas/config-schema.json"
  - "examples/**/*.yml"
---

What are Companion Files?

In the context of AI Agents, companion files specified in subfiles are exactly analogous to Support Files (e.g., in Claude Skills). They allow you to pass external knowledge, strict schemas, or necessary reference material to the agent without cluttering the prompt itself.

Concrete Example: API Schema

Imagine you are creating a skill that writes API requests. Providing the entire OpenAPI schema inside the markdown file would make the document thousands of lines long.

Instead, separate it:

src/content/docs/api-client-rules.md (The Skill)

---
title: "API Client Rules"
description: "Rules for writing new API queries to our internal system."
skill: true
subfiles:
  - "schemas/internal-api.yaml"
---

When generating API requests, ALWAYS reference the exact endpoints and required parameters defined in `schemas/internal-api.yaml`. Do not invent endpoints.

src/content/docs/schemas/internal-api.yaml (The Companion File)

openapi: 3.0.0
info:
  title: Internal API
paths:
  /v1/users: ...

By separating them, the AI agent receives both the clear text rules and the strict structured data, while human readers only need to read the short markdown rules.

Why Separate Companion Files?

Separating your skill into the main document and companion subfiles offers massive benefits:

1. Context Window Efficiency: The agent reads the concise instructions in the main .md page, and only references the strict subfiles (like the JSON schema) when explicitly needed, saving its token window.
2. Human Readability: By moving gigantic datasets or strict mechanical schemas out of the prose, human developers can easily read the document's concepts without scrolling past walls of JSON.
3. Perfect Formatting: Exporting your mechanical datasets completely intact as secondary files avoids tricky Markdown string-formatting issues.

Usage Rules

• They must be text files. Including binary files (like images, zip files, or PDFs) in the matched glob pattern will purposefully trigger a build error to prevent corrupting the AI's string-based payload.
• Unmatched glob patterns are simply ignored.
• Over on the doc's website, the plugin automatically generates sub-pages for these files. Markdown/MDX files render normally, while scripts and schemas display as formatted code blocks.
• When outputting for agent consumption (SKILL.md), the subfiles are copied alongside the skill, preserving their relative directory structures. See the Deployment chapter for details on the output file format.