Pull the all-time install leaderboard for Claude Code skills today and count the authors. The ten most-installed skills come from exactly three names: Matt Pocock has five, Vercel Labs has four, and Anthropic, the company that created the skill format, has one.
We run Skillselion, a directory of Claude Code, Codex and Cursor extensions ranked by real installs. The numbers below are from our live catalog this morning, July 29 (refreshed daily from skills.sh, GitHub and MCP registries).
The top 10, with installs
| # | Skill | Author | Installs |
|---|---|---|---|
| 1 | find-skills (repo) | Vercel Labs | 2,723,091 |
| 2 | frontend-design (repo) | Anthropic | 715,763 |
| 3 | grill-me (repo) | Matt Pocock | 690,693 |
| 4 | agent-browser (repo) | Vercel Labs | 591,076 |
| 5 | vercel-react-best-practices (repo) | Vercel Labs | 587,496 |
| 6 | grill-with-docs (repo) | Matt Pocock | 585,310 |
| 7 | improve-codebase-architecture (repo) | Matt Pocock | 565,718 |
| 8 | tdd (repo) | Matt Pocock | 544,478 |
| 9 | web-design-guidelines (repo) | Vercel Labs | 497,786 |
| 10 | setup-matt-pocock-skills (repo) | Matt Pocock | 495,226 |
How three names ended up owning the board
Each of the three got there a different way.
Matt Pocock's five slots are an author subscription. Developers install his whole engineering workflow in one motion: the grilling interviews, the architecture pass, the TDD loop. The tell is slot 10. setup-matt-pocock-skills is a one-time per-repo configuration skill that scaffolds what his other skills assume (where issues live, triage labels, docs layout). A skill whose only job is to prepare a repo for his other skills has nearly half a million installs. You install that because you adopted the author, not because you needed a skill.
Vercel's four slots are a product surface. find-skills at number 1 is the discovery layer itself, the skill that installs other skills (we wrote about it two days ago). The other three ship Vercel's opinions about React, web design and browser automation straight into every agent that installs them. When agents write a large share of frontend code, the style guide the agent follows is worth owning.
Anthropic's single slot is the format itself doing its job. Anthropic defined SKILL.md and ships a first-party skills repo, and its one entry, frontend-design at 715,763 installs, sits at number 2. The community out-shipped the format's creator on its own format. That is what an open format is for, and install-ranked lists make it visible.
The 1,600-install cliff at position 10
Positions 11 through 25 explain the cliff. All fifteen are Lark skills covering Feishu/Lark docs, calendar, mail, spreadsheets and more, packed into an 863-install band between 492,763 and 493,626, the signature of a bundle installed as one unit.
setup-matt-pocock-skills holds position 10 by exactly 1,600 installs. If the bundle adds its next 1,600 installs before setup-matt-pocock-skills does, the bottom of the top 10 stops being individual skills developers chose and starts being whatever arrived attached to something else.
What this means when you pick skills
Install counts are a trust signal, but the top 10 shows the trust attaches to authors and bundles more than to individual skills. Practical reading:
- A famous author's fifth-most-installed skill rides on the trust his first one earned. Judge it on its own SKILL.md.
- Bundle skills carry near-identical counts by construction. Fifteen Lark skills at ~493k each is one distribution decision counted fifteen times.
- A standalone skill that earned six figures without an author brand or a bundle behind it is the strongest signal on the board.
The full live leaderboard, filterable by category, is at skillselion.com/skills.
We build Skillselion, a directory of Claude Code, Codex and Cursor extensions ranked by real installs. Live catalog, refreshed daily from skills.sh, GitHub and MCP registries.
Top comments (4)
The concentration is interesting because skills are supposed to be an ecosystem, but discovery can quickly turn into a power law. The next useful question is whether the top skills are winning on quality, naming, distribution, or simply being early enough to become defaults.
We got a partial answer to that this week. The catalog-wide picture points at distribution more than quality: the only blocks moving in lockstep are a 13-skill enterprise-suite bundle and a 5-skill media bundle from one author, every member rising the same +13% regardless of what it does - installs arrive per bundle, not per skill. Meanwhile the fastest accelerator in the top 100 right now is a browser-control skill at +62% daily rate while the median top-100 skill sits at 0.94x its late-July pace, and that one is hard to explain by naming or earliness - it tracks a capability shift, since agents got an in-app browser last month. Earliness still looks strongest at the very top: the #1 skill is a discovery skill that installs other skills, so every new user compounds its lead. So provisionally: discovery-layer position beats quality at the head, bundling beats both in the middle, and capability shifts are what actually move the rankings week to week.
That bundle effect is a useful signal. If installs move as a package, then ranking individual skills by install growth can overstate independent demand. I would want discovery metrics that separate suite distribution from skill-level pull, otherwise the catalog starts optimizing for bundle shape instead of actual utility.
Your overstatement concern matches what we see in the data, and the cleanest separator we have found so far is within-repo gain dispersion over a dated window. A bundle shows near-zero dispersion: over the last nine days, six of the seven caveman skills gained between 20,722 and 21,202 installs each, a spread of 480 on gains of about 21,000. That is package movement, not seven independent decisions. Independent pull shows a gradient: the same window inside the mattpocock repo runs from about 96,000 gained at the top to under 30,000 at the tail, a better than threefold spread, because each skill is being chosen at its own rate. One caveat cuts the other way: author subscription makes even genuine per-skill demand partially correlated, so dispersion can understate real pull for a trusted author's mid-tier skills. As a first-pass filter though, dispersion near zero is a strong bundle tell, and it is computable from public install counts alone.