Compare AI coding skills head-to-head to find which one works best for your task. Run controlled A/B tests: same model (Claude Code), same tasks, different skills. Each candidate runs in an isolated worktree sub-agent. Output: a markdown comparison report with scores and a winner.