You built something with AI in an afternoon. It worked — honestly kind of magic — and then someone asked you to ship it. Suddenly there are no tests, no clear ownership, the architecture drifted three times, and the PR review is a disaster. The problem isn't the AI. It's the workflow.
In this talk, we'll frame AI coding as three modes: Sketch (Vibe Coding), Blueprint (Context Engineering), and Contract (Spec-Driven Development). Each has a moment where it's the right tool — and a moment where it'll quietly wreck your codebase.
You'll see one real feature built three ways in a live demo: first with raw prompting, then with injected context and tool permissions, and finally with a spec-first CLI flow that produces reviewable, testable code. We'll cover when to switch modes, how to write a lightweight "constitution" for your repo, and how to adopt spec-driven practices on an existing codebase without going full waterfall.
Leave with a clear playbook, templates you can use the next day, and a simple rule for knowing which mode you're in before you start.