The 1,000 to 2,000 Character Trap
We have measured the relationship between SKILL.md length and score before. Across 591 evaluated skills, the correlation was r = 0.24. The new batch shows a much stronger version of the same pattern.
| Description length | Skills | Mean score |
|---|---|---|
| 1,000 – 2,000 chars | 5 | 3.40 |
| 2,000 – 4,000 chars | 8 | 6.14 |
| 4,000 – 6,000 chars | 3 | 8.53 |
| 6,000 – 10,000 chars | 24 | 7.50 |
Across the forty new skills, the correlation between stored description length and score is r = 0.542.
The buckets are small — three skills in the 4,000 to 6,000 band — so treat the exact means as noisy. The direction is consistent: the shortest files are the weakest.
What the short files look like
The five skills under 2,000 characters are mostly e-commerce how-to pages: ecommerce-ab-testing, conversion-rate-optimization, customer-feedback-analysis, shopify-seo, and shopify-conversion-optimization. They describe a topic in a few paragraphs and stop.
The 6,000 to 10,000 character band contains the education, revenue-operations, and Jira skills. They are longer because they include procedures, examples, and boundaries.
The causal trap
It would be easy to conclude "write longer skills". That is not the lesson. A long file can be padded, and a short file can be excellent if it encodes a narrow, well-defined task. The correlation is a symptom: in this batch, the short files are short because nobody wrote the procedure.
Our earlier analysis of 591 skills found r = 0.24, not 0.54. The difference is likely composition: this batch deliberately includes a repository of thin marketing pages, which drags the short end down.
The right reading is:
- Length is not a quality target.
- A very short skill is a prompt to check whether the procedure is missing.
- A long skill still needs a trigger, a sequence, and a verification step.
For the earlier, larger analysis, see Does a Longer SKILL.md Score Higher?.