The Vercel AI SDK makes it easy to call a model and stream a response, but the quickstart skips the parts that matter for a real application: forwarding the user's identity to your internal services, writing a system prompt that holds up under pressure, and building tools the model can't misuse. Here's how I approach all three.

How I built a secure sandbox for running untrusted player scripts using V8 isolates, a two-layer API architecture, and runtime code generation.