Skills
Claude Code ships `claude plugin eval`: every case runs with and without your plugin, and the delta is the only score that proves it did anything
Claude Code 2.1.269 (Sep 11) added `claude plugin eval`, which runs a plugin against a suite of prompt cases three times each, then repeats every case with no plugin loaded so you get WITH, W/OUT and Δ columns. Six grader types exist: regex, tool_used, tool_order and file_exists are computed from the transcript and cost nothing, while llm and baseline call a judge model. The docs name the most common first finding explicitly, a Δ near zero with the `tool_used: Skill` grader failing, which means your skill's description is not winning on natural phrasing.
↳ Follow the thread