Single-Skill Repos Beat Mega-Repos on Score — Barely

The intuition is that a focused repository produces a more polished skill than a hundred-skill catalogue. The data agrees, weakly.

Across the 591 fully-evaluated skills, we can group by how many skills share the same source repository:

The direction is what you would expect: focused repositories score slightly higher and get downloaded slightly more. The magnitude is the story — a 0.2-point difference, which is below the resolution we have measured for our own evaluator (re-running the same evaluation twice moves scores by up to a full point).

So the honest conclusion is that the gap is real in direction and not established in size. Anyone quoting "single-skill repos are better" from this table is reading precision the data does not have.

Why mega-repos dominate anyway

66 repositories supply 591 skills. The six largest supply 535 of 765 in the full directory. If mega-repos were meaningfully worse, that concentration would be a problem. At a 0.2-point difference, it is not.

There are mechanical reasons mega-repos hold up:

• Shared conventions. A catalogue applies the same frontmatter, the same section order, the same references layout across every skill. Consistency scores well on structure and engineering.
• Batch evaluation effects. Skills from one catalogue are often evaluated in the same run, with the same context, which produces correlated scores. That compresses within-catalogue variance and can lift an entire catalogue.
• Selection. Authors who publish 100 skills have usually published a few and iterated. The catalogue reflects accumulated practice.

Where the difference actually shows

Looking at the tails rather than the means is more informative. Every skill scoring below 7.0 in the current directory comes from a multi-skill repository, and four of them are from a single family. The failure mode of a mega-repo is not a lower average — it is a longer tail of thin entries published to keep a catalogue growing.

That is the real trade-off: mega-repos raise the floor through shared conventions and lower the floor by publishing skills that did not need to exist.

What to do with this

• If you are publishing one skill: you have a small quality edge and a small distribution disadvantage. The edge will not carry you; distribution will.
• If you are evaluating a catalogue: sample the middle, not the top. The top of any catalogue is curated; the middle is where the variance lives.
• If you are building a directory: count repositories, not skills. A 119-skill catalogue is one publisher, and treating it as 119 independent data points will mislead every aggregate you compute.