We moved away from an LLM-first approach and shifted toward a code-first architecture with bounded AI assistance.