What I've learned building this

The biggest challenge was keeping her reactive — early versions hallucinated hero names and ability descriptions that were obviously wrong to anyone watching. The prompt now keeps her relational and reactive to what I report, rather than generative about things she can't know.

The voice stack was the hardest part. GPT-SoVITS had cascading dependency issues — torchcodec DLLs, a Gradio/Jinja2 conflict, and the FFmpeg build all needed sorting before the API ran stable. Took a while, but it sounds good now.