Your line that the description "has to say both what the skill does and when to reach for it" matches what I see at scale. Disclosure: I run Skillselion, a directory that indexes tens of thousands of public SKILL.md files, and vague descriptions are the most common defect I see in the wild, exactly as your linter ranking suggests. Two candidate rules for skillmd-lint's list, drawn from that corpus: first, a license check. The overwhelming majority of public skills declare no license at all, which quietly blocks the fork-and-adapt loop your "read it as if it were code" advice depends on. Second, a trigger-phrase check: descriptions that state what but never when, since those register fine and then never fire. Your injection-shaped-content story about the comedy skill rejected over the word "secretly" is a genuinely useful data point about marketplace scanners. Is the seven-mistake ranking published anywhere with frequencies? That distribution would be worth citing.
Your line that the description "has to say both what the skill does and when to reach for it" matches what I see at scale. Disclosure: I run Skillselion, a directory that indexes tens of thousands of public SKILL.md files, and vague descriptions are the most common defect I see in the wild, exactly as your linter ranking suggests. Two candidate rules for skillmd-lint's list, drawn from that corpus: first, a license check. The overwhelming majority of public skills declare no license at all, which quietly blocks the fork-and-adapt loop your "read it as if it were code" advice depends on. Second, a trigger-phrase check: descriptions that state what but never when, since those register fine and then never fire. Your injection-shaped-content story about the comedy skill rejected over the word "secretly" is a genuinely useful data point about marketplace scanners. Is the seven-mistake ranking published anywhere with frequencies? That distribution would be worth citing.