GLOW is a chat-based practice environment in which Graduate Teaching Assistants (GTAs) rehearse office-hour conversations with AI student personas and receive feedback aligned to a human-authored, behaviorally anchored rubric. Deployed during GTA onboarding across two recent cohorts at a large R1 university, GLOW enabled authentic, repeatable practice at scale, showed consistent improvement across retries, and was associated with higher self-reported confidence for difficult cases than discussion-only training. Persona-level analytics – especially lower success with Aggressive students – pinpointed where coaching should focus. We argue that rubric-anchored simulation is a pragmatic, scalable way to support GTA readiness without sacrificing pedagogical grounding.