60,611 skills in the wild — what a full census of the public SKILL.md corpus shows

Insight · Mike Arbuzov · SkillFed Research · July 28, 2026

SkillFed walked all 6,177 repositories in its discovery queue end to end: 2.5× more unique skills than listings claimed, 13,122 per-agent variant files merged, and 86,956 vendored aggregator copies excluded — more copies than originals.

How many agent skills exist in public? Directory listings and awesome-lists will give you a number. We wanted the real one — so we walked every repository in SkillFed's discovery queue end to end, every tree, every SKILL.md, and counted.

The census

A typical skill body is small: 1,867 tokens on average, median 1,668, 90th percentile 4,505. The long tail is real — the largest single SKILL.md weighs 80,691 tokens — but only 5.1% of bodies exceed 6,000 tokens.

Finding 1: listings undercount the corpus 2.5×

Before the walk, the discovery queue — built the way most skill directories are built, from listings and search results — claimed 23,771 skills. The tree walk found 60,611 unique skills: 2.55× more.

The undercount has structure. 13.9% of listed entries turned out to be repo-level pointers rather than skills. Listings routinely named a handful of skills in repos that contained dozens. And the naive enumeration strategy — take the shallowest SKILL.md match per name — would have silently dropped nearly three-quarters of a multi-skill repository's contents. A skill count is only as good as the walk that produced it; listings are discovery hints, not enumerations.

Finding 2: the same skill ships many times

The single biggest counting hazard is that one skill is frequently many files. Two patterns dominate:

• Per-agent variants. The same skill shipped once per target agent — claude/, codex/, copilot/ directory branches carrying near-identical copies. One enterprise repo holds 148 SKILL.md files that resolve to just 32 actual skills.
• Vendored copies. The same skill re-committed under .claude/skills/, .agents/skills/, and versioned subtrees of the same repository.

Counting files triple-counts the ecosystem; merging them is where the signal is. Collapsing 13,122 variant files did more than fix the count — it produced a dataset: 7,026 skills now carry explicit which-agents-does-this-support evidence derived from the paths their variants shipped under. That field was empty on every record before the merge.

Finding 3: copies outnumber originals

64 aggregator repositories — bulk collections vendoring other authors' skills — contain 86,956 SKILL.md files between them. That is more than the entire unique corpus. One aggregator alone vendors 6,290.

Include them and every statistic about the ecosystem becomes a statistic about duplication; naive crawls that count them report a corpus half again larger than reality, with the excess concentrated in whoever got vendored most. We exclude aggregators above a uniqueness threshold and elect one canonical record per skill, so that authorship and provenance stay attached to the original.

Finding 4: a third of the corpus can't be reproduced

Only 64.6% of skills carry a license that permits storing and full-text-indexing their bodies. The remaining third of the ecosystem is discoverable — we know it exists, where it lives, and what it's called — but its content cannot be lawfully redistributed. Skill authors who want their work found: a LICENSE file is worth more reach than any amount of keyword tuning.

Why a search engine runs a census

Every number above changes what a skill search returns. Dedup decides whether three copies of one skill crowd out two alternatives. Variant merging decides whether "works with my agent" is answerable at all. Aggregator exclusion decides whose name appears on the result. License gating decides what can be shown rather than merely linked. This census is the ground SkillFed's skill search stands on — and alongside it we read the research: the field map of 364 agent-skills papers and the 191-note research directory are the same discipline pointed at the literature.

The same discipline, pointed back at ourselves: we later ran the SEO skills we recommend against this site and found 61 things, 19% of the criteria stale on arrival.

Method and limits

The population is SkillFed's discovery queue — 6,177 public GitHub repositories accumulated from listings, search, and crawl — walked completely, with per-repo checkpointing and a verified claimed-versus-stored reconciliation (zero rows lost). It is a census of that queue, not of all of GitHub: a repository nothing in our discovery ever pointed to is invisible here. Counts are a snapshot as of 2026-07-28; the ecosystem is growing quickly enough that the absolute numbers will age faster than the structural findings.

More from SkillFed Research

1. Insight · Aug 2026 Any AI chat can now run skill search — and you approve every request

   No install, no account, no connector. Your chat writes an abstract wish, you paste the link back, and it reads five security-swept skills. The whole request is a URL in plain English — the privacy boundary is something you check, not something you're asked to trust.

2. Field report · Jul 2026 61 findings on a site we built for SEO

   A site with build-blocking structured-data lints, machine-readable mirrors and an enforced internal-linking floor still failed 61 checks drawn from the SEO skills our own editorial recommends — including FAQPage markup that same post called retired. 19% of the skills' criteria were stale too.

3. Insight · Jul 2026 The largest direction in agent-skill research is spreading outward, not settling down

   Papers on agents that write their own skills land steadily farther from the direction's own semantic center month over month — the only trend in our analysis that survives multiple-comparison correction (BH p = 0.0016) — with no single axis carrying the drift.

4. Insight · Jul 2026 Zero of 184 recent papers connect skill self-authoring with skill security

   Five of the six research-direction pairs in the recent agent-skill literature are bridged by dual-topic papers. The pair formed by its two largest directions — agents authoring their own skills, and securing skill files — is empty, and three null models say that is not chance.

5. Field report · Jul 2026 Agent-skills research didn't exist before 2023 — and its fastest-growing direction today is security

   A SkillFed field map of 364 agent-skills papers, 2016–2026: none of this work existed before 2023, and skill security went from nothing to the second-fastest-growing direction in about three quarters.