How we collect and assess resources

A reproducible source snapshot, with explicit signals and limits. Quality is an editorial heuristic, not a user rating or a promise of safety.

Discovery and inclusion

GitHub Repository Search uses claude-skill, codex-skill, agent-skill, mcp-server topics and SKILL.md in README. A candidate needs 25 stars, a push within 90 days, a recognized license, a README of at least 250 characters and a SKILL.md of at least 350 characters. Forks, archived repositories, truncated trees and duplicate file hashes are excluded. The first preview includes skill files only; MCP-only repositories remain discovery candidates.

Quality formula · 100 points

• Documentation, 40: skill ≥500 characters 10 (otherwise 5); Markdown headings 10; example or fenced code 10; README ≥800 characters 10 (250–799: 5).
• Maintenance, 25: repository pushed within 30 days 25; 31–60 days 20; 61–90 days 15. Push activity is not proof that this individual skill was maintained.
• Stars, 20: min(20, round(5 × log10(max(1, stars)))), using repository stars, not skill popularity.
• Recognized license, 15. Skill frontmatter overrides repository metadata. A license is not an endorsement or a complete nested-file rights review.

The initial inclusion threshold is 75/100, an unvalidated editorial starting rule. Lower scoring entries are excluded from the initial snapshot and remain noindex if retained in a later review snapshot.

Static risk formula

Scan SKILL.md and up to four companion scripts ≤60 KB each for shell, network, file read/write and credential references. High = shell plus credentials or file write. Medium = any signal or unscanned companion scripts. Low = no signals in the inspected files and no known unscanned scripts. Remote dependencies, dynamically constructed code, linked references, malicious intent and actual execution are not assessed. Even low risk can be dangerous.

Updates and provenance

Refresh is a manual script with caching and rate-limit backoff; no hosted cron is enabled. Source-specific editorial notes are bound to the inspected file hash. Changed or unknown sources need another editorial review. Author avatars are linked GitHub identifiers, attributed in CREDITS.md.

Editorial review and publication

Offensive and forensic resources are withheld from the initial publication. Retained high static-risk entries stay noindex, show a warning before setup and are excluded from recommendations. The initial search cohort includes existing curated entries and up to 50 reviewed automatic entries, with a maximum of 100 resources. A score alone never opens indexation. Every Preview page stays noindex.