LinklyAI/best-skills

Best Agent Skills is a continuously updated leaderboard that tracks the popularity of agent skills — reusable capability packages that give AI agents (like Claude Code, Cursor, or OpenClaw) specific abilities, such as web browsing or UI design. The project merges signals from several separate skill registries plus social media into nine rankings, and publishes the underlying data as open CSV files.

Why it exists

Each skill registry only sees its own corner of the ecosystem. skills.sh counts Claude/Vercel CLI installs, ClawHub counts OpenClaw downloads, Tencent SkillHub counts installs from China — and none of them capture social chatter. This project merges those views into one cross-ecosystem picture, showing global installs, China installs, and social mentions side by side for each skill.

What the rankings show

The project produces nine lists, refreshed daily:

• best-100 — overall "Worth-Installing Score" (popularity + momentum + buzz + maintenance + trust)
• top-installs — most installed across all ecosystems
• trending-7d — fastest-growing over the last week
• social-buzz — most discussed on X, Hacker News, Bluesky, and GitHub (generic-name skills are excluded because their mentions can't be reliably attributed)
• most-active — popular skills with the most frequent updates
• official-100 — top skills from verified publishers
• official-vendors — verified companies/organizations ranked by total installs
• top-repos — the GitHub repositories behind skills, by stars
• rising-stars — best newcomers first seen within the last 30 days

Around 10,000+ skills are tracked, with Top 10 previews in the README and full Top 100 lists in CSVs.

How the numbers are handled

Two principles stand out:

1. Raw numbers preserved — every CSV keeps each platform's original counts, so anyone can verify or re-rank the data.
2. No apples-to-oranges summing — install counts from different platforms are never added together. Cross-platform scores are based on a composite rank percentile rather than raw sums, so a million installs on one platform doesn't get compared directly to a million on another.

The project also uses an AI decision model (jev from TypeSafe) rather than keyword matching to judge whether social posts are actually about a skill, whether a listing is a real maintained skill, and which category it belongs to. Placeholders, deprecated, and harmful listings are excluded, and every judgment probability is published.

How you can use it

• Read it directly — browse the README tables, the full CSVs, or the interactive web version at linkly.ai/skills.
• Give it to an AI agent — paste a single line (Read https://linkly.ai/skills/llms.txt) into any agent that can fetch a URL; the file tells the agent where the CSVs are, what each column means, and how to pick the right file.
• Install as a skill — npx skills add https://github.com/LinklyAI/best-skills --skill best-skills makes the rankings part of an agent's standing knowledge.

Data structure

The repository stores one folder per day (e.g. data/2026-10-23/) containing:

• raw/ — untouched per-platform counts: skills.sh, ClawHub, SkillHub, GitHub repos, social buzz, X posts, AI judgments, and more.
• rankings/ — the nine computed lists.

A data/latest/ folder always mirrors the most recent day, and an index/first-seen.csv tracks cumulative first-seen dates to power the rising-stars list.

Sources and license

Install/download counts come from public endpoints of skills.sh, ClawHub (per its third-party directory policy), Tencent SkillHub, and the GitHub API. Social signals come from X, Hacker News (Algolia), Bluesky, and GitHub search. The project is not endorsed by any of these platforms.

All data and documentation are released under CC BY 4.0 — free to use, republish, and build on, including commercially. The only requirement is attribution with a link back to the repository.