Vibe Coding in Practice (zed.dev)

In a live Agentic Engineering session, veteran engineer Manuel Odendahl demonstrated "vibe coding" by building a Doodle-style calendar clone while running multiple AI models in parallel and treating each model as a specialized tool. Rather than fine‑tuning prompts to produce perfect code, Manuel uses a lightweight YAML DSL as an intermediate representation—formal enough to be interpreted as code, yet readable like natural language—and orchestrates model outputs into sandboxed features. He stressed managing expectations, learning from model failures, and guiding models with engineering judgment instead of micromanaging every line of generated code. For the AI/ML community this highlights a practical workflow for rapid prototyping and human–AI co‑engineering: model composition (parallel, specialized models), intermediate DSLs (YAML) for predictable, sandboxed outputs, iterative prompting, and on‑the‑fly documentation generation to understand system behavior. The approach can speed feature development and make tooling accessible to non‑programmers, but it remains probabilistic—outputs can be nonsense and require verification—so it demands skilled oversight, testing, and continuous adaptation as model behaviors evolve. Vibe coding is positioned as a complementary modality to pair programming or whiteboard design: high‑velocity and exploratory, yet needing strict validation before production use.