PluginEval

Claude Code: /plugin marketplace add wshobson/agents → /plugin install plugin-eval@claude-code-workflows

Three layers stack, each more expensive than the last: a static pass that parses SKILL.md and finishes in under two seconds, an LLM judge that scores triggering, orchestration, output and scope in about thirty seconds, and a Monte Carlo layer that runs the skill fifty to a hundred times to estimate reliability. The static checks are opinionated in a way worth knowing before writing skills — progressive disclosure is measured against a 200 to 600 line sweet spot, and every detected anti-pattern multiplies the layer score down by another five percent, flooring at half. Results feed an Elo ranking against a corpus, so a skill is placed relative to its peers instead of against an absolute bar. It ships as a real Python package with a test suite, not just prompt files.