evaluation-rubrics — Claude Code skill: scoring criteria and scales

What evaluation-rubrics does

The skill walks the agent through six steps: define purpose and scope, pick criteria, design the scale, write descriptors for each level, calibrate on sample work, then iterate. Templates for each step sit alongside a methodology file covering scale design and calibration.

Criteria are meant to be observable rather than impressionistic — the test is whether two independent reviewers score the same item the same way. The skill suggests a 1–4 or 1–5 scale, with 4–8 criteria covering the dimensions that matter. Descriptors have to separate adjacent levels concretely, not just say "good" and "very good".

Five patterns are laid out: analytic, holistic, single-point, binary checklist and standards-based. Calibration needs several reviewers on the same samples; if their scores drift apart, the descriptors are the thing to rewrite.