Andrej Karpathy recently gave a talk called "From Vibe Coding to Agentic Engineering." One line stuck with me: "People have to be in charge of this spec, this plan. Work with your agent to design a spec that is very detailed." He is describing what happens when you stop treating AI as a magic text box and start treating it as something that builds from structured input. The casual approach works for small things. A quick script, a throwaway page, a one-off function. The moment the problem gets real, you need a spec, not just prompts. I have been building this way for a long time now. I want to talk about what I have learned from practicing spec-driven development. What it actually looks like in practice, and why it produces better work with less correction. Four layers It is common for projects using Claude Code or similar tools to have only one layer of configuration: a CLAUDE.md file with some instructions and context. Maybe some rules. That is the behavioral layer.…