Two "Standards" That Are Really One Author's Template
A heading that appears in 15% of a corpus looks like a convention. Check the author column before believing it.
Two of the more distinctive headings in the SKILL.md corpus:
| Heading | Files | Distinct authors |
|---|---|---|
## Skill Contract | 119 | 1 |
## Data Sources | 125 | 2 |
The Skill Contract heading is used by 119 skills, all from a single author (aaron-he-zhu). Data Sources appears in 125, of which 118 are from the same author and 7 from one other (agiprolabs).
If you sampled randomly and found a "Skill Contract" section in most of your sample, you would reasonably conclude it is a common practice. You would be describing one publisher's style guide.
Why this keeps happening
The directory has 65 authors. Six of them supply 70% of the skills. That concentration means any structural pattern originating in a large catalogue will show up as a corpus-level statistic.
We have now hit this pattern four times in a row:
- Category averages → author averages.
- Trigger quality by category → one marketing catalogue's writing style.
- Chinese-language share → one bilingual catalogue.
- Section conventions → one template.
Each time, the apparent ecosystem-level finding dissolved at the author level. The general lesson is not about any specific number. It is that a corpus of 765 items from 65 authors is not 765 independent observations — it is closer to 65, with wildly unequal weights.
What a real convention would require
A genuine convention would show up across many independent authors. In this corpus that is rare:
## Overview— universal, but because our importer adds it.## When to Use— 178 files across many authors. This one is real.## Output/## Output Templates— 106 files, split between two names. Behaviourally a convention; terminologically not.
Everything more specific than that tends to be one catalogue.
What to do with a heading
If you are writing a skill, copying a good catalogue's structure is fine — it was probably designed carefully. But do not treat any single heading as required, and do not assume a reader has seen it before. The only structure a router actually depends on is the description field.
If you are reading a corpus, always group by author before drawing a conclusion. The heading counts look like a map of the ecosystem. They are mostly a map of who shipped the most files.