Nobody can tell you how many skills exist
We ran two full crawls of every public .claude/skills/SKILL.md a day apart. They disagreed on the count by about a third — a single crawl even misses ~a third of the discoverable repos.
Discovered-repo overlap between the two: . So we publish rates (which are stable to ~±2 pt), never an absolute count.
About a third of skills are copies of each other
of files are byte-identical to another author's — but only of that is a repo duplicating itself. of the redundancy is cross-repo: genuine cross-author copying. And much of it is wholesale: of files sit in repos that are near-clones of another repo (template & starter repos re-uploaded under a new name).
And ~1 in 9 aren't real skills
of published SKILL.md files have no valid manifest — templates, placeholders, empty scaffolds. Of the real ones, here's the anatomy:
Most skills are short prose (body median ). Proper tool-scoping and hooks are rare — a low-cost way to make a skill better and safer.
The most-copied skills
Identical files that appear across the most independent owners — the community's de-facto building blocks.
What you can do with it
- Before you write a skill, check if it exists. A third are copies and the de-facto standards are identifiable — adopt or improve the best existing version, or confirm you've found a genuine gap.
- Don't mistake volume for validation. Neither repo stars (author fame) nor raw copy-count means a skill is good — much high-copy content is just cloned templates. Read the content, not the spread.
- Differentiate to stand out. Most skills declare no tools and use no hooks — proper scoping is a low-cost way to make yours better and safer.
Search the distinct skills
A random, representative sample of valid-manifest skills with their content fetched. Search by name or description; filter by structure.
Why it matters
The duplication, the junk, and the count-instability all point to one gap: the ecosystem has no trustworthy, deduplicated index. People clone whole repos because there's nothing better to search. This is both the evidence that the gap exists and a first working version of the fix.
Explore it yourself
Everything here is open — the method, code, and the checked-in data. Check the numbers, challenge them, or fork the crawl. Want the next study? Watch the repo on GitHub.