Which AI is best at your job?

Every month, each model does the same professional tasks twice: on its own, and with the matching skill from this library. A pinned judge scores both on a fixed 1–5 rubric. Pick your job to see who does it best, and whether the skill helps. All of it is open: the tasks, the harness and the raw results.

How to read it: differences under 0.25 are noise. The judge is a Claude model, which can favour Claude-style answers. A job with one tested model is not ranked. Models appear only once someone has run them; to add one, add its provider key to the monthly workflow, or run node skillbench/run-skillbench.mjs --models <id> yourself and open a pull request. See also SkillBench by domain.