Claude Code's 17 Official Skills: the Rule vs the Practice

Short summary

A deep analysis of all 17 official Anthropic Claude Code skills reveals that docs and practice diverge: the 500-line limit is a target not a law, triggering depends entirely on frontmatter descriptions (not body), and staged loading is the exception rather than the norm. Anthropic measures description quality with 20 eval queries run 3x each across 5 iterations, picking descriptions by held-out test scores. Exclusion clauses in descriptions are only useful when skills overlap; otherwise they reduce triggering, the worst flaw.

• •Skill triggering depends on frontmatter description, not body — write pushy descriptions
• •500-line limit is a target; context budget is the real principle
• •11 of 17 official skills are single SKILL.md files with no subfolder
• •Measure description quality with eval queries: 8-10 positives, 8-10 near-miss negatives, run 3x over 5 iterations

Generated with AI, which can make mistakes.