On March 18, 2026, Google Labs product manager Rustin Banks announced a full overhaul of Google Stitch with an unfamiliar workflow: don’t start with wireframes. Start by describing the business goal, the emotion users should feel and reference examples, then explore many directions quickly and narrow down. He called it vibe design, and outlets including The Register, TechRadar and Fast Company put the phrase in headlines within a day.
Practice was already heading there. In a 2024 survey of about 10,000 freelance designers by 99designs by Vista, 52% said they used generative AI tools, up from 39% a year earlier. The term descends from “vibe coding,” coined by Andrej Karpathy in February 2025. Code has an instant referee — the compiler or a failing test — while design quality still has to be judged by people, which is why the idea took another year to reach design.
Stitch launched as a Google Labs experiment in May 2025, turning text or sketches into UI screens and front-end code. The overhaul centers on an infinite canvas where screenshots, code snippets, text and competitor screens serve as context. A design agent reads the whole project and generates several directions at once; static screens become clickable prototypes; and an MCP server and SDK let other agents call Stitch.
At Google I/O in May 2026, Stitch changed again: generation now streams onto the canvas as it happens. Designers can interrupt and redirect mid-generation, use voice as a constant input, co-edit with others and publish directly to the web through Netlify.
When three mockups took two days, meetings debated which three to make. When twenty take minutes, meetings debate which to keep. Design studio Eleken reported that some phases that once took five to ten days now take an hour with AI — while stressing that what compresses is execution speed, not expertise.
Strengths: escaping the blank canvas, generating variations in color, density and tone, drafting multi-screen flows like onboarding or checkout, and persuading stakeholders with clickable screens.
Weaknesses:
Many teams now split the work in three: exploration in Stitch, building in tools like v0, Lovable or Bolt, and final polish by people.
Vibe design doesn’t take the job; it reallocates time from making to choosing. Choosing requires explainable criteria rather than taste — what this user needs at this moment, what emotion the brand promised, which metric the screen should move. The same criteria produce better prompts and better selections, echoing the argument that a designer’s value in the AI era comes from judgment, not production.
There’s a risk, too. The wireframing grunt work that trained junior designers’ eyes is exactly what got automated. Spend saved time generating more and you only get more to choose from; spend it on user research, flow redesign and critique and it converts into judgment. Real-time generation also removes the pause where second thoughts used to happen. Just as verification is the missing step in vibe coding, design teams need to build deliberate checkpoints into the flow.
A workflow where designers describe goals, emotions and references in natural language, then select and refine from AI-generated options instead of drawing wireframes first. The term arrived with Google Stitch’s March 2026 overhaul.
Precise element-level edits, brand-specific character, edge states like empty or overflowing content, and guaranteed accessibility still need human work.
Clear, explainable criteria for what to keep and why. The same criteria drive better prompts as well as better selections.
To apply what you just read to your own site, start with a free audit of where things are now.
A strategist replies within 24 hours on business days.
Buzzvil's BD strategy team has zero developers but runs 32 AI workflow automations — from 7-minute revenue drop analysis to 90% fewer manual steps. How they scaled from personal hacks to team systems.
Meta’s Muse Wonder and Muse Wonder Full add Active Memory and Active Focus so its AI agent can handle tasks that take days. But a Toronto user’s Marketplace mishap shows the risks of agentic AI.
Pew cloned thousands of panelists as AI digital twins and found a 12.4-point average error. Five ways synthetic respondents fail — and where marketers can still use them in research.