How ranking works
Search starts with your task. A skill matching more of your search terms ranks ahead of one matching fewer. Within that group, explicit matches in the skill name come first, followed by relevance, maintenance history and documentation evidence.
Relevant to your task
Field-weighted BM25 favors matches in the name and description over incidental mentions in a preview. Specific task words in the skill name are prioritized over incidental mentions, with rarer technology terms carrying more focus than generic task words; exact names and phrases receive an additional boost. We recognize narrow aliases such as PostgreSQL/Postgres and common Chinese technical phrases; React and Vue, or competing database engines, are not synonyms. Chinese support is a limited technical vocabulary, not general translation.
Task forms such as “validate,” “validation” and “validator” match each other. Document processing is distinct from technical documentation. Common Chinese phrases for document analysis, spreadsheet extraction and validation are supported; Excel, XLSX, CSV and JSON remain separate terms so results do not imply unsupported file formats.
File history, not crawl time
The date shown is the latest change to the indexed skill file on the repository’s default-branch history. A README update elsewhere does not refresh every skill. A Git commit records an edit; it does not prove that APIs or dependencies were tested.
Among equally relevant candidates, the maintenance weight decreases continuously with age: a 180-day decay scale with a 35% floor. Useful older material stays discoverable. This applies to the last file change, not the original publication date.
We inspect up to 2,000 commits per source and exclude artificial shallow-clone boundaries. Unknown dates receive a conservative weight; limited history may establish only “unchanged since.” Neither is presented as a recent update. First observation by our index is not a creation date.
Evidence and popularity
The base assessment gives up to 65 points for useful description and document structure, examples, verification guidance and references; 20 for traceable source metadata and a stated license; and 15 for logarithmically scaled repository stars. Raw length, keyword repetition and author celebrity do not earn unlimited points.
Maintenance decay applies to the resulting assessment. Archived sources are reduced further. A small source-diversity adjustment limits one large collection crowding the results. Stars belong to the repository, not the individual skill; license metadata is not a legal review.
What scores mean
These are transparent ranking heuristics, not measured task-success rates or security certifications. Search scores are relative to a query. The discovery leaderboard uses the same maintenance and documentation assessment without query relevance. Safety indicators are separate metadata heuristics.
Only matching names with equivalent file contents are merged; revised versions remain independently rankable. Historical IDs resolve to their current entries. The website, REST API and remote MCP share this policy and a daily catalog snapshot. API results include ranking components and timestamp status so clients can inspect why a skill was suggested.
Policy version 2026-11-18.1. We will refine these initial weights using relevant-task evaluations and actual repeat usage.
Explore ranked skills →