KG-Registry includes a small set of local agent skills to support curation and maintenance workflows in this repository, plus a set of read-only discovery skills for exploring what the registry knows about a resource or domain.
The curation and maintenance skills are designed for curators working in a checked-out copy of the repository. The discovery skills are different: they need no local clone and query the published Parquet files over HTTP, so anyone with network access can use them.
The two kinds of skill have different requirements.
The read-only discovery skills (search-resources, trace-upstream-sources, find-downstream-usages, find-similar-resources) require no local clone. They query the published Parquet files over HTTP at https://kghub.org/kg-registry/registry/parquet/ using DuckDB with the httpfs extension, for example via uv run --with duckdb --no-project python. All you need is network access.
The curation and maintenance skills (kg-registry-curation, curate-next, kg-registry-validation, kg-registry-product-url-update) must be run from a local clone of the repository. They rely on:
resource/reports/curation_problems.tsvuv run make validate-file FILE=...If you want to curate or maintain resources and are not working in a local clone, stop and clone the repository first.
The current agent workflows are defined in the repository’s .agents/ directory.
That directory includes:
Use the kg-registry-curation skill, or the matching curate command entry point when your agent environment supports commands.
This workflow is for expanding a stub resource or improving a specific existing resource page. It focuses on the source page under resource/<id>/<id>.md, expects web search when the current page is incomplete, and finishes with local validation.
Use the curate-next skill, or the matching curate-next command entry point when available.
This workflow selects the next stub resources from the local repository, skips resources already listed in reports/curation_problems.tsv, and dispatches per-resource curation.
Use the kg-registry-validation skill.
This workflow checks a curated file for schema validity and curation-specific quality issues such as placeholder stub content, missing provenance structure, missing products, or incorrect source associations.
Use the kg-registry-product-url-update skill.
This workflow is for replacing a stale product_url with the best current live URL for the same product. It is especially useful when an old version-specific URL has disappeared and the best replacement is a newer canonical page, release page, or current download endpoint.
These skills are read-only and need no local clone. They answer questions about what the registry already records by querying the published Parquet files over HTTP at https://kghub.org/kg-registry/registry/parquet/ (resources.parquet, resource_domains.parquet, resource_products.parquet, resource_taxa.parquet). They run DuckDB with the httpfs extension — for example via uv run --with duckdb --no-project python — and do not edit resource pages. Provenance, usages, and publications live inside the raw_data JSON column of resources.parquet.
Use the search-resources skill.
This workflow finds resources covering a subject area (for example, “everything about RNA interactions”). It maps the free-text topic to the controlled domain vocabulary, runs a keyword search over names and descriptions, sweeps the source files for matches the structured query misses, and returns a ranked, de-duplicated list noting each resource’s category and activity status.
Use the trace-upstream-sources skill.
This workflow determines the upstream data sources a resource (usually a knowledge graph) is built from, by walking each product’s original_source and secondary_source, resolving the referenced KG-Registry identifiers, and optionally recursing to build the full provenance tree back to primary sources.
Use the find-downstream-usages skill.
This workflow determines everywhere a resource is used downstream. It combines a reverse-provenance sweep (other resources whose products name the target as a source) with the target’s own curator-recorded usages field.
Use the find-similar-resources skill.
This workflow finds alternatives and near-duplicates of a resource by combining structured signals KG-Registry records (shared domains, taxa, category, graph scale, shared upstream sources) with heuristics such as description similarity and overlapping publication authors, then ranks the results with an explanation of why each is similar.
The skills are exposed through two complementary, independent standards. Either can be used on its own; supporting both costs nothing extra.
/.claude-plugin/marketplace.json defines a Claude Code plugin marketplace with two plugins:
kg-registry-discovery — the four read-only discovery skills (no local clone needed).kg-registry-curation — the four curation and maintenance skills (require a local clone).Add the marketplace and install a plugin from a Claude Code session:
/plugin marketplace add Knowledge-Graph-Hub/kg-registry
/plugin install kg-registry-discovery
This resolves against the repository’s default branch, so the plugins become installable once these files are merged to main.
/.well-known/ai-catalog.json is an Agentic Resource Discovery manifest (the ai-catalog format ARD builds on). It lists every skill as a catalog entry with a domain-anchored URN identifier, a short description, capabilities, and representativeQueries that registries use for semantic matching. Each entry’s url points at the raw SKILL.md, and its metadata links back to the marketplace plugin that installs it.
Because the site is served under the /kg-registry base path, the manifest is published at https://kghub.org/kg-registry/.well-known/ai-catalog.json rather than the bare domain root. An ARD registry can be pointed at that URL directly; placing or linking a copy at the kghub.org domain root is a separate hosting step outside this repository.
The agent skills are written around a few KG-Registry-specific rules:
resource/, not generated output in registry/src/kg_registry/kg_registry_schema/schema/kg_registry_schema.yamloriginal_source and secondary_sourceresource/.These skills are not a substitute for broader project setup or site deployment instructions.
For repository setup, development, and troubleshooting, see the root-level development documentation such as README-sitedev.md.