Large language models have become surprisingly capable coding partners - not just for autocomplete, but for turning a rough problem statement into runnable code, tests, and iterative improvements. In this talk I'll give a short, technical overview of why coding assistants work (internet-scale code + documentation pretraining, instruction tuning, and feedback loops), and where they predictably fail (ambiguous specs, hidden requirements, brittle dependencies, and "confidently wrong" outputs).

Most of the session is live, terminal-based demos using Claude Code. We'll start from plain-English requirements and iteratively produce working implementations: scaffolding a small project, generating tests, debugging failures, refactoring for clarity/performance, and adding documentation. Along the way I'll highlight practical workflows: "spec-first prompting", test-driven prompting, using the assistant to explain unfamiliar code, and using constraints to keep changes small and reviewable. The goal is for faculty and students to leave with repeatable patterns for using coding assistants effectively - without outsourcing understanding or sacrificing correctness.