Three catalogs undercount one Claude skill by as much as 40,111 stars

All three catalog figures for ui-ux-pro-max sit below what the GitHub API returned. The API gave 128,854 stars for nextlevelbuilder/ui-ux-pro-max-skill ^[4]. One catalog listed 117,079, short by 11,775 ^[3][1]. Another listed 119,920, short by 8,934 ^[2]. The third listed 88,743, short by 40,111, or 31 percent of the API figure ^[3][4]. The three sites also disagree with each other by 31,177 stars ^[5].

Nokka writes that the figures he pulled change every day and should be read as a single-day comparison ^[6]. Three numbers below the live count is what stale caches look like. The post also discloses its own provenance: drafted by deepseek-v4.1-flash through Hermes Agent, checked and edited by Nokka ^[2].

One claim runs above GitHub's count, and it measures something else. A catalog advertises 316,581 installs for the same skill, about 2.46 times its star count ^[3][6]. Nokka set a single rule for the piece: every number direct from the GitHub API, and where he could not confirm a figure he would say so instead of guessing ^[5]. The post leaves that installs number unsourced.

mattpocock/skills was created on 3 February 2026 and showed 265,304 stars on 19 September ^[9], an average of about 1,164 stars a day across 228 days ^[7]. obra/superpowers showed 288,565 ^[7], which is 156,203 more than excalidraw/excalidraw's 132,362 ^[8]. Two entries, and stars are already hard to use as a ranking. Nokka writes that excalidraw is a hand-drawn diagram tool used alongside skills, and that its count is high because it is a large worldwide project ^[19]. For these counts to rank adoption, skills repos published within the last year would have to be drawing stars faster than a drawing tool in daily use. The API is authoritative about what GitHub recorded that day. Bulk-added stars and stars added one at a time look the same inside that number.

JuliusBrussee/caveman cuts filler out of agent output and keeps code, commands and error messages byte-identical, because garbled code or a mangled error message breaks the work downstream ^[13]. It is the most careful design in the list. The repo describes itself as "why use many token when few token do trick" and claims a 65 percent token cut ^[12]. Nokka writes that he has yet to find an independent test of that number ^[14]. Its license field came back NOASSERTION where the other repos returned MIT ^[11], which means GitHub could not match a standard license, and he advises reading the LICENSE file before commercial use ^[14].

tohmsc/caveman-mode does similar work with 70 stars, about 1,523 times fewer than JuliusBrussee/caveman ^[15][9]. humanizer-tools/slop-humanizer has 5 stars against blader/humanizer's 49,966, a factor of roughly 9,993 ^[16][17][10]. Name collisions are the cheapest way to install the wrong thing. The smallest entries have the clearest ownership: addyosmani/web-quality-skills at 2,804 stars, aimed at Lighthouse and Core Web Vitals ^[18], and Vercel's own vercel-labs/skills at 31,970 ^[20].