How Skillselion Research Is Done

Three report shapes, kept apart

Merging them would cannibalize both the live snapshot and the dated citations. The hub at /research collects all three.

• Edition. A frozen, dated census under /research/<slug>-<yyyy-mm>. Numbers never change after publication. Each edition ships its raw CSV. Example: The Agent Economy Census, August 2026.
• Evergreen. The State of AI Agent Skills 2026 recounts the live catalog. Cite it for "what is true now", not for a figure that has to survive a fact-check years later.
• Reference. AI Agent Statistics is one line per statistic, each sourced to a frozen edition. The page updates when a new edition lands; the editions stay frozen.

The dump

An edition starts with a complete pull of the production catalog on a stated date (the listings API with full detail, walked by keyset). That file is the only input. We do not mix a fresh dump with leftover numbers from an earlier one, and we do not invent counts the file does not contain.

Install counts are cumulative registry install events, not active users. Bundle installs (one action installing several skills, common for suite publishers) count each bundled skill separately; editions that this affects say so next to the finding. MCP-server rows currently carry no install signal, so we do not rank MCP servers by installs.

For skills.sh, the public number is the all-time leaderboard and search total, the same figure the HTML skill page shows. We do not rank or quote from the per-skill detail endpoint (GET /skills/{id} installs): on 16 Aug 2026 that field was tens of thousands of times the public count for some publishers (typically 12 on the page vs 200k+ on the detail API). August 2026 editions dumped the catalog as stored that day, so a few publisher totals in those CSVs still reflect the detail figure. Those editions keep their published numbers and carry a dated errata note rather than a silent rewrite.

The freeze law

Once an edition is public, its published numbers do not move. A later dump is a new dated route. We do not silently patch a live census, because that would make every existing citation point at a different arithmetic. Corrections, if one is ever needed, are a dated note on the same URL plus a new edition. They are not a rewrite of the original figures.

The fact-check gate

Every number a report page publishes must recompute from the shipped CSV alone. A standalone script asserts the page constants against that public file before the edition ships. A second reviewer independently recomputes three to five of the headline figures from the same file and fails the page on any mismatch. That gate has already caught a real error (a copy-count that named the wrong unit) and forced a missing column into a CSV. Editions without a re-runnable script do not ship.

Reuse and citation

All Skillselion Research is free to reuse with attribution under CC BY 4.0. Name the report and edition, and link the report page. Each edition has a "Cite this report" box with the exact string. Journalists and researchers who need a cut we have not published can email info@skillselion.com; we answer with the numbers and the method together.

What we will not claim

A research page that cannot defend a number is worse than no page. Standing refusals:

• Empty security-scan fields. Catalog columns for a first-party Verified-Safe scan and a trust score are not populated. We do not publish pass rates from empty fields. Audit coverage in the security census comes from third-party verdicts the registry already publishes, and the edition says so.
• MCP install rankings. MCP rows have no install signal today. Star counts measure the host repository, not the server. Any MCP census must say that in the method, or it does not ship.
• Unexplained cross-edition deltas. Catalog-wide install totals have jumped in ways that daily velocity cannot explain (likely upstream re-syncs). We do not headline those gaps as growth until they are root-caused.
• Numbers nobody can recompute. If it is not in the named snapshot, it is not on the page.

Who publishes this

Skillselion Research is published by Skillselion, an independent directory and research organization operated by Ellelion LLC (Wyoming, United States). Pages attribute the work to that organization. Skillselion is not affiliated with Anthropic, OpenAI, Cursor, Claude, Claude Code, or Codex. More on how the catalog is run.