metadata.md

Skill metadata contract

Create one UTF-8 JSON object:

{
	"name": "example-docs",
	"title": "Example Documents",
	"description": "Answer questions using the indexed Example documents.",
	"source_summary": "The source set covers product setup and operating procedures.",
	"topic_groups": [
		{
			"name": "Product usage",
			"summary": "Setup, configuration, and common workflows.",
			"query_terms": [
				"product name",
				"configuration"
			],
			"document_ids": [
				"001-example.pdf"
			]
		}
	],
	"aliases": [
		{
			"canonical": "Example Product",
			"aliases": [
				"Example"
			]
		}
	],
	"coverage_notes": []
}

Validation limits:

• name, title, description, source_summary, and topic_groups are required.
• name: lowercase letters, digits, and single hyphens only; maximum 50 characters.
• title: maximum 80 characters. description: maximum 500. source_summary: maximum 1,200.
• topic_groups: 1-12 objects. Each requires name (maximum 80) and summary (maximum 300), with at most 10 query_terms (maximum 80 each) and 5 document_ids (maximum 255 each).
• Every document_ids value must exactly match a document ID returned by the completed batch iterator.
• aliases: at most 20 objects. Each requires canonical (maximum 80) and at most 10 aliases (maximum 80 each).
• coverage_notes: at most 10 values, maximum 300 characters each.
• The completed iterator proportionally samples its bounded outline across documents and includes deterministic preservation-only descriptions for image content that was not analyzed.
• Keep every field grounded in the completed iterator's document list and outline. Omit uncertain aliases or coverage claims rather than inferring them.
• Add a coverage note only when the returned outline explicitly states that boundary. Missing topics, versions, or examples do not prove that the source set excludes them; use an empty array when no explicit boundary is available.