best-skills: A Daily Ranking of the Top 100 Agent Skills, Built From Real Installs

Ninety-four days after Anthropic's Skills API left beta, there are now more agent skills circulating than any one person can realistically evaluate by hand, and a GitHub star count doesn't tell you which ones people actually installed and kept using. best-skills is a repository that tries to answer that question directly: a free, daily-updated ranking of the top 100 agent skills, built from real install and download numbers pulled across skills.sh, ClawHub, Tencent's SkillHub, GitHub, and social mentions on X, Hacker News, and Bluesky, not from stars or self-reported popularity.

The repository itself is small, just a README, a methodology doc, and a data/ folder of CSVs, but the numbers behind it are real: as of this week it's tracking installs across four separate registries and has grown from a standing start on August 11 to 424 GitHub stars.

What each ranking measures

The core problem best-skills is built to solve is that every skill registry only sees its own ecosystem. skills.sh counts installs made through the Claude and Vercel CLIs. ClawHub counts downloads inside the OpenClaw ecosystem. Tencent's SkillHub tracks China-region activity separately, and its numbers aren't a mirror of ClawHub's even when a skill is synced between the two, they represent genuinely additional local usage. None of the three sees social buzz on its own.

best-skills pulls all of it into nine separate rankings rather than pretending the numbers are comparable. The flagship list, "Best 100," is a Worth-Installing Score built from five weighted dimensions: popularity (30%, install percentile within whichever platforms a skill exists on), momentum (15%, position on skills.sh's own trending list), reputation (20%, a percentile blend of X, GitHub, Hacker News, and Bluesky mentions), maintenance (15%, update recency and version count on ClawHub), and trust (20%, whether a publisher is verified on skills.sh or ClawHub). Other lists split this back out for anyone who wants a single dimension instead of a blended score: Top Installs, Trending (7 days), Social Buzz, Most Active, Official 100, Official Vendors, Top Repositories, and Rising Stars for anything under 30 days old.

Running it day to day

There's no dashboard or connector to set up. The whole thing is a GitHub repository with a README that renders the top 10 of each ranking directly, plus a data/YYYY-MM-DD/rankings/ folder holding the full top 100 as CSVs for whichever date you want. Checking which skills are worth installing this week means opening the README, or one of the CSVs directly if you want the full 100 rather than the top 10 preview, since every ranking links straight through to the skill's actual page on whichever registry it came from.

The methodology document is unusually candid about how the numbers are actually built, which matters more than the score itself. Raw numbers from different platforms are never summed together, since a skills.sh install and a ClawHub download measure different populations under different rules; instead every number is converted to a percentile within its own platform first. Repository stars never rank an individual skill, since a monorepo bundling 17 skills would otherwise make each one look artificially popular off one shared star count, stars get their own separate "Top Repositories" ranking instead. A skill whose name is an ordinary English word, like "weather" or "github", is excluded from social-buzz scoring entirely, because there's no way to tell a mention of the skill from a mention of the word. And unmeasured dimensions are filled with a flat neutral score rather than dropped, specifically so a skill nobody has data on can't accidentally outrank one that was actually measured and scored zero.

The gaps the project owns up to

The project is upfront about several real gaps rather than papering over them. skills.sh's telemetry only counts CLI installs, so anyone installing a skill by cloning the repo directly is invisible to the count, and installs are bundle-level, meaning every skill in a popular multi-skill repository inherits the same inflated number, whether or not each individual skill inside it is actually used. Star-growth figures only go back to when the project itself started snapshotting, since GitHub restricted its own stargazers history API in July 2026. Social coverage is English-dominant across X, Hacker News, and Bluesky, so non-English adoption signals aren't captured yet. And the anomaly detector currently catches one specific pattern, identical install counts shared across three or more skills from the same publisher, without yet catching near-duplicate clusters that fall just short of identical.

None of this makes the rankings meaningless, but it does mean a high Worth-Installing Score is evidence of measured adoption within these five specific registries and time windows, not proof that a skill is well-built or safe to run unreviewed. There's no security scanning behind the trust dimension yet, only whether a publisher is formally verified.

A discovery tool, not a safety check

This is a genuinely useful reference for anyone deciding which of the dozens of skills circulating each week are actually worth the install, rather than guessing from a star count or a single enthusiastic post. It's a poor fit for anyone who needs a security or code-quality signal, since nothing here evaluates what a skill's code actually does, only how many people have installed it and how recently it's been touched. The data itself is published under a Creative Commons Attribution license, so it can be reused or re-ranked by anyone willing to credit the source, which is a different license than a permissive code license like MIT, worth knowing if you plan to build on the dataset rather than just reading the README.

Sources: best-skills, LinklyAI, GitHub; the repository's own methodology document, read directly against the live rankings and LICENSE file rather than the scan's snapshot numbers.