evaluation-methodology

The composite score sums each dimension score multiplied by its weight, scales by 100, then applies an anti-pattern penalty of max(0.5, 1.0 minus 0.05 per flag). Dimension scores blend static, judge, and Monte Carlo layers using per-dimension weights.

Rewrite the frontmatter description to include the exact phrase "Use this skill when" followed by three or more specific, comma-separated trigger contexts. Test with five prompts that should trigger and five that should not to check discrimination.

Quick depth runs only static analysis in under two seconds. Standard adds the LLM judge, taking 30 to 90 seconds. Deep depth adds Monte Carlo simulation with 50 invocations and Elo ranking, taking 15 to 20 minutes via the certify command.

Additional content likely triggered anti-pattern flags such as OVER_CONSTRAINED or BLOATED_SKILL, and the penalty multiplies the whole score. Run with JSON output and inspect layers[0].anti_patterns, then move detail into a references directory.

Platinum requires composite 90 or higher and Elo 1600 or higher; Gold needs 80 and 1500; Silver needs 70 and 1400; Bronze needs 60 and 1300. Below 60 earns no badge, and the Elo check is skipped when Elo has not been computed.

Quick depth only runs static analysis, so it cannot judge output quality or detect orchestration problems, and rewrites that move content to references may score lower statically. Use standard depth for fair comparisons that include the LLM judge.