I want this to be a distinction too, vibe coding with benchmarks and evals and making good architectural decisions for the LLM coder based on them takes a lot of mental energy and engineering insight. I've been making a voice agent and Claude's initial attempts were all based on best practices from two years ago. I had to do a lot of paper reading and research consolidation to get any kind of consistent, useable quality results.