The weekly star-growth ranking: how it is computed and the data behind it

ShowLiangke Li3 replies

3
Liangke LiMember#1

/weekly lists the ten GitHub repositories that gained the most stars in each complete week. Because "trending" lists are usually opaque, here is exactly how ours works, and the data is downloadable.

Method. We take daily snapshots of star counts for the projects tracked here. For each week we compare the snapshot nearest the Monday boundary with the one nearest the Sunday boundary; both must fall within 36 hours of the boundary. If either one is missing for a project, the project is dropped for that week rather than estimated. Archived, mirrored and unlisted repositories are excluded. A week only enters the archive once it has ten qualifying projects. Details are on /ranking-methodology.

What nine weeks showed (2026-07-27 through 2026-09-21, published as CSV at /data/star-growth):

  • 90 ranking slots were filled by 47 distinct projects.
  • 30 projects, 64%, appeared exactly once.
  • Mean overlap between consecutive weeks was 4.9 out of 10, ranging from 2 to 6.
  • The longest runs were seven weeks, by three projects: mattpocock/skills, stablyai/orca and DietrichGebert/ponytail. All three dropped out in the week of 2026-09-14; stablyai/orca was back the week after.

The overlap number is the interesting one. Half the list turns over every week, which is a lot of churn for "what developers care about". We read it as: star growth measures attention, not adoption, and it says more about launch weeks than about usage.

Mistakes in the data or the method belong in Site Feedback; questions about the numbers belong here.

3 replies

ValerieRegular#2

Why not simply mirror GitHub's own Trending page? It is what most people already look at.

Liangke LiAuthorMember#3

In reply to Valerie (#2)

Three reasons. GitHub does not publish how Trending is computed, so we could not explain a position. It is a point-in-time page with no archive, so you cannot ask "what was on the list in August". And it is not reproducible from data anyone else can download. Ours is slower and less exciting, but every row can be recomputed from the two snapshots that produced it.

IcarusRegular#4

The pattern in the long-run projects is hard to miss: two of the three seven-week runs are "skills" repositories, collections of task instructions for coding agents. The 09-14 list added addyosmani/agent-skills and Tencent/BrowserSkill. Whatever one thinks of the format, that is the category that held attention over the whole nine weeks, more than any model or framework release.