The Repo Average Hides the Spread
The seven repositories in this batch have averages from 4.67 to 9.07. The more useful number is the range inside each one.
| Repository | Skills | Mean | Range | Spread |
|---|---|---|---|---|
| anthropics/k12-teacher-skills | 4 | 9.07 | 8.7 – 9.3 | 0.6 |
| genspark-ai/genoffice | 4 | 8.50 | 8.4 – 8.6 | 0.2 |
| NEON-Rutger/B2B-revops-skills | 6 | 8.47 | 7.7 – 9.0 | 1.3 |
| jdrhyne/agent-skills | 2 | 8.20 | 6.9 – 9.5 | 2.6 |
| gemini-cli-extensions/workspace | 5 | 7.00 | 5.9 – 8.3 | 2.4 |
| claude-office-skills/skills | 9 | 5.84 | 4.9 – 7.5 | 2.6 |
| nexscope-ai/eCommerce-Skills | 10 | 4.67 | 3.2 – 9.1 | 5.9 |
The e-commerce repository has a six-point spread. Its best skill, dynamic-pricing-ecommerce at 9.1, is better than the average of every other repository except the top two. Its worst skill, ecommerce-ab-testing at 3.2, is the lowest in the batch.
Stars are not the answer either
The correlation between repository stars and mean score across these seven repositories is weak. genspark-ai/genoffice has 11,420 stars and averages 8.50. nexscope-ai/eCommerce-Skills has 1,305 stars and averages 4.67. anthropics/k12-teacher-skills has 606 stars and averages 9.07.
At the skill level, across all forty, the correlation between stars and score is r = 0.203.
Stars measure whether a repository was noticed. They do not measure whether the skills inside it were written to be executed.
The practical rule
Use the repository as a prior. Then evaluate the individual skill. A strong repository raises your expectation; it does not replace the read. A weak repository can still contain one excellent file, and a strong repository can contain one weak file.
This is why the directory evaluates at the skill level, not the repository level. The methodology is public; the scores are per file; the reader can check the reasoning.