这就是 Claude Code Skill 开发里最隐蔽的一类失效： 注册成功、语法合法、claude plugin validate 全绿，但在真实对话里从未被调用。它不像报错那样会跳出来提醒你，反而因为模型本身够强，产出看着还行，于是问题被长期掩盖。

Graded evals use a judge to score tone, faithfulness, and coverage. They cover the outputs assertions can't reach, such as the quote-drafter's register and the summarizer's selection of what mattered.