Score Predicts Downloads. Star Count Predicts Nothing.
If you are choosing between two skills, the score is a weak signal and the star count is no signal. The useful part is narrower than that: only the top band separates itself.
We have 591 skills with a complete six-dimension evaluation (the other 174 in the directory are imports whose dimensions were never scored — a separate problem I will write about separately). For those 591 we also have a download count and, for most, a GitHub star count for the source repository.
Two correlations, both computed on that set:
| Relationship | Pearson r |
|---|---|
| Our score ↔ downloads | 0.47 |
| GitHub stars ↔ our score | 0.02 |
The second number is not a rounding artifact. Stars tell you a repository was noticed; they do not tell you whether the skill inside it is any good. We wrote about that separately, and 765 skills later it is still true.
The first number is more interesting, and the way most people will read it is wrong.
It is a cliff, not a slope
Correlation suggests a gradient. The actual data is closer to a step:
| Score band | Skills | Average downloads | Median |
|---|---|---|---|
| 9.5 and above | 18 | 32.1 | 32.5 |
| 9.0 – 9.5 | 165 | 8.3 | 7.0 |
| 8.0 – 9.0 | 293 | 6.4 | 6.0 |
| 7.0 – 8.0 | 97 | 5.7 | 5.0 |
| Below 7.0 | 18 | 5.8 | 5.0 |
Read the bottom four rows again. A skill at 7.3 and a skill at 9.4 get essentially the same number of downloads — 5.8 versus 8.3, with medians of 5 and 7. The difference between "mediocre" and "very good" is invisible in the download data.
Then the top band is four times higher than everything below it.
So the honest reading is: the score is not a ranking device. It is a filter with one useful threshold, somewhere around 9.5. Below that line, 390 of 591 skills cluster in a band where the number stops carrying information about adoption.
Why this might be an artifact
I would not publish the correlation without the caveat that undercuts it.
Downloads on this site include automated traffic. We have separately found bot sweeps that download the top-rated skills in coordinated batches — the same set of skills, seven times each. If an automated client scrapes a "top rated" list and downloads everything on it, that produces a positive score-download correlation mechanically, without any human preference being involved.
So the 0.47 is an upper bound on how much real users prefer higher-scored skills, and possibly a substantial overestimate. What survives the objection is the cliff: even if bots are doing the downloading, they are downloading the top band disproportionately, which means the top band is the one that gets surfaced — by us, by directory listings, by whatever list the client is following.
That is still actionable, just for a different reason than "users prefer quality."
What to do with it
- If you are publishing a skill: the goal is to clear 9.5, not to move from 8.0 to 8.7. The data cannot distinguish those two, and neither can a reader.
- If you are choosing a skill: use the score as a floor filter. Above 9.0, read the rationale and decide on fit; the numbers below that are not telling you which one to install.
- If you are building a directory: publish bands, not decimals. A 0.2-point difference in a ranking is noise — we measured that directly by re-running the same evaluation twice.
The one thing stars are good for
They measure a repository's reach, which is a real thing — just not the thing you are trying to decide. A high-star repository is more likely to have maintainers, issues, and a future. That is a durability signal, not a quality signal, and the two should be kept separate in your head.
The methodology behind the score is public at methodology, and every skill page carries the written rationale that produced its number.