The most dangerous pull request in your company might contain zero lines of code. It's late, and the last PR waiting for review contains a handful of markdown files, part of some skills and rules in your IDE. This huge pile of nearly unreviewable text... How do you even know whether this change badly hurts the team's workflow?

Skills are becoming the new R&D workhorse, yet the science of testing them remains vague. Drawing from production systems, we'll explore state-of-the-art techniques for testing and improving skills: spec linting, activation rates, outcome verification, several TypeScript-based evaluation frameworks, including some that specialize in skills, seven types of fundamental judges, and a handful of practical best practices for building high confidence in agent skills (exemplified with JavaScript code)