Google AI published a hands-on demonstration this week showing what "vibe coding" looks like in practice: a developer talks through what they want to build while Gemini 3.1 Flash Live writes and iterates on code in real time, producing a working website in under 10 minutes.

What Was Demonstrated

The demo, recorded in Google AI Studio, shows a developer speaking their requirements out loud — no keyboard typing, just conversational prompts — while the model generates HTML, CSS, and JavaScript on the fly. The result is a functional, styled web app built entirely through voice-driven iteration.

The demo is available as a remixable template, meaning anyone can fork the starting point and try the workflow themselves inside AI Studio.

Why Gemini 3.1 Flash Live

Gemini 3.1 Flash Live launched earlier this week specifically for real-time voice and vision agent use cases. Google says the model achieves response latency close to natural dialogue speed, which is what makes interactive vibe coding sessions feel fluid rather than staccato.

The model also shows improved instruction-following in noisy environments — a deliberate design choice for agents that need to parse conversational input rather than clean typed text.

Broader Trend

The vibe coding demo sits alongside a larger shift in how developers interact with AI models. Rather than writing a detailed prompt and reviewing output in a loop, conversational iteration lets developers shape code incrementally, correcting in real time as they see what's being built.

Google notes the workflow "keeps up with brainstorms" — meaning the bottleneck is now the developer's thinking speed, not the model's generation speed.

The template is live in Google AI Studio now.