Vibes Don't Maintain

A restaurant owner built a working task-management web app for his staff using Anthropic's Claude Code without writing any code, but the prototype lacked responsive layout, authentication, settings, translations, error states, and edge-case handling, persisting data only well enough to survive a page refresh. The anecdote illustrates a broader gap in how non-engineers judge AI-generated software: they see the demo's happy path and assume the product is nearly finished, when most of a real codebase consists of safeguards for failures that may never occur. The author argues the prototype is a valuable proof of concept, but hosting and operations are not the remaining 10% of the work. The piece does not cite user counts, funding, or revenue figures. What to watch next is whether AI coding tools close the gap between demo-quality prototypes and production-grade systems, or whether demand grows for engineers to harden and operate what those tools produce.