My AI coding agents read two kinds of instructions. A skill is a page of instructions an agent reads before a particular kind of task. The guidance is one page of rules every agent reads at the start of every session — nearly every rule on it was learned from a real mistake. This page shows, for each skill and each section of the guidance, how the agent does on the same task with it and without it, scored by fixed pass/fail checks and by a second model acting as judge. Switch between skills and guidance at the top of the page.

The numbers update on their own: every new test run lands on the results branch and the page reads it fresh each time it loads. How the tests work is written down in How it works.