Skillselion Research · Edition of 2026-08-12
The Agent Skill Clone Census
We content-matched all 79,848 listings in the AI agent skill ecosystem: 3,443 are verified copies of another listing, and they hold 26,061,541 installs, 15.8% of every install counted. The surprise is who does the copying.
On 2026-08-12 we dumped the whole Skillselion production catalog and checked every listing’s SKILL.md against every other. Opportunists ripping off hits explain little of the mass. Publishers re-listing their own skills account for 35.0% of copy-installs, one generative-media product under three names accounts for 53.7%, and the people copying Anthropic captured 3.15% of the audience the originals earned. You can reproduce every number from the dataset; the method and its limits are in the methodology.
Finding 01
15.8% of all installs land on copied content
A copy here means proof, not suspicion: the listing shares its name with another listing and its SKILL.md is byte-identical to it, or at least 90% identical for forks frozen on an older revision of the original. By that test, 3,443 listings (4.3%) are copies, and they carry 26,061,541 installs. So 4.3% of the catalog collects 15.8% of the installs. A handful of heavyweight copies produce that gap, and the next two findings name them.
- Share of all installs on copies15.8%
- Share of all listings that are copies4.3%
Finding 02
The heaviest copiers are publishers copying themselves
1,262 copies, holding 9,130,783 installs (35.0% of all copy-installs), come from the same publisher as the original: the identical skill, listed again under a second repository or namespace. The largest single case is ByteDance, whose larksuite GitHub org re-lists 15 of its open.feishu.cn Lark skills for 5,571,004 installs. Each install is real. It is the listing count and the per-skill totals that double, which matters to anyone quoting ecosystem numbers, including us.
- One product cluster under three handles14,005,596
- Same publisher, second listing9,130,783
- All other third-party copies (2,124 listings)2,925,162
Finding 03
One product, three names, 14.6% of the whole census
The single biggest block of copied installs belongs to one family of RunComfy-branded generative-media skills (image generation, face swap, video tools). The same skills appear under three publisher handles: agentspace-so, which holds the canonical listings, plus runcomfy-com and doany-ai carrying byte-identical copies. The copies alone hold 14,005,596 installs, 53.7% of every copy-install in the census. Count the whole family, originals included, and it reaches 24,049,825 installs, 14.6% of the entire ecosystem, from one product family listed under three names.
- agentspace-so (canonical listings)10,043,689
- runcomfy-com (copies)7,126,177
- doany-ai (copies)6,879,419
We can verify the files are identical; we cannot verify from the outside whether the three handles are one company or resellers of the same product. The dataset flags every row in the cluster, so you can weigh it either way.
Finding 04
Anthropic gets copied most. The copies get almost nothing.
Rank originals by how many copies exist and the platform vendor leads its own ecosystem: 374 listings are copies of Anthropic skills, with frontend-design alone copied 59 times. Copying is a bad growth strategy. All 374 Anthropic copies together hold 112,444 installs, 3.15% of what the originals earned. The 59 copies of frontend-design hold 14,818 installs against the original’s 767,871. Outside the big clusters, the median copy has 29 installs. Adoption stays with the canonical listing.
- frontend-design (Anthropic, original: 767,871 installs)59 copies
- web-design-guidelines (Vercel Labs, original: 535,426 installs)36 copies
- canvas-design (Anthropic, original: 96,830 installs)28 copies
- skill-creator (Anthropic, original: 348,833 installs)27 copies
- vercel-react-best-practices (Vercel Labs, original: 625,739 installs)23 copies
- xlsx (tfriedel, original: 2,479 installs)22 copies
One exception: the 22 copies of xlsx out-earn its 2,479-install original with 3,432 installs between them. Everywhere else in the top six, originals win by two orders of magnitude.
Method
Methodology
The census is a complete dump of the Skillselion production catalog taken on 2026-08-12: 79,848 listings and 164,841,042 cumulative installs, aggregating public registry data, chiefly the skills.sh registry, alongside GitHub and other public sources. Copy detection is content-gated and never name-only: rows that share a name qualify only when their stored SKILL.md is byte-identical to the group’s most-installed row, at least 90% trigram-similar to it, or byte-identical to each other in a leftover cluster. MCP server duplicates (50 of the 3,443 copies) use a same-repository rule instead. Installs count registry install events, not active users.
Same-publisher classification is conservative: it requires the identical author account, plus the one namespace pair we verified as a single company (ByteDance’s larksuite org and its open.feishu.cn listings). 148 copies point at 78 originals that have since left the catalog; they stay in the copy count and carry no canonical join in the dataset. Copies are one measurable slice of what gets called AI slop. Low-effort originals are real too, and this census does not attempt to count them.
A copy flag is a statement about file content, not about a publisher’s intent or quality. Skillselion is an independent project and is not affiliated with Anthropic, OpenAI, Cursor, Claude, Claude Code, or Codex.
Data
Download the dataset
The full census runs 79,848 rows, one per listing, with the copy flag and a canonical join per copy: the original’s id, name, publisher and installs, so you can rebuild every family and dedupe any total yourself. It is free to use with attribution under CC BY 4.0: cite “Skillselion Agent Skill Clone Census, August 2026” and link this page.
FAQ
Questions journalists ask
How much of the AI agent skill ecosystem is copied?
In this census, 3,443 of 79,848 listings (4.3%) are verified copies: they share a name with another listing and their SKILL.md is byte-identical to it, or at least 90% identical for forks frozen on an older revision. Those copies hold 26,061,541 installs, 15.8% of every install counted. Low-quality originals exist too; this census counts only proven duplicates, so treat it as a floor for the wider slop question.
Are install counts in this ecosystem inflated?
Some double counting is measurable. At least 9,130,783 installs (5.5% of the census total) sit on second listings from the same publisher, and one product family listed under three handles accounts for 24,049,825 installs, 14.6% of the census, on its own. Ecosystem install totals, including the ones Skillselion publishes, count each listing separately; the dataset attached here carries a copy-to-original join so you can dedupe them yourself.
Is copying a skill even allowed?
Mostly unanswered: 98.2% of the 58,766 skills in this census declare no license in their SKILL.md manifest. Code with no declared license grants no reuse rights by default, though a repository may grant one elsewhere. We count the copies; we do not police them.
Can I reuse these numbers or the dataset?
Yes. Cite "Skillselion Agent Skill Clone Census, August 2026" and link this page. The raw dataset (79,848 rows with per-row copy flag, canonical name, canonical publisher and installs) is downloadable on this page under CC BY 4.0.