Google updated Stitch with a “vibe design” workflow and Figma shares fell almost 12 percent over two days. The market reaction was immediate, but the product boundary matters more than the stock chart.
Stitch accepts natural-language direction, generates multiple UI approaches, and offers a Voice Canvas where the agent asks questions and changes the layout during a conversation. It exports to Figma and HTML/CSS, connects to coding agents through MCP, and offers an SDK. The free tier includes 350 generations per month.
That is a strong zero-to-one tool. A product manager or designer can explore several directions and put something concrete in front of stakeholders within minutes. Production design lives in the remaining distance: a system with hundreds of components, variants, breakpoints, accessibility rules, content extremes, and accumulated decisions. Generating a plausible screen does not maintain that system.
Where Stitch fits without owning the system
Google Labs rebuilt Stitch as an infinite canvas: several visual directions, clickable prototypes, an agent that tracks the project's evolution, and Voice Canvas to critique the layout while you talk. You can drop in image, text, or code as context. You can extract a design system from a URL or load rules in a DESIGN.md file and reuse them on another project.
That solves the blank page and the meeting where nobody can picture the product. A team with a living library already has a different owner for production. Tokens, variants, breakpoints, and component states already live in Figma or in code. If five generated screens become a handoff on the spot, the system drops out of the path.
MCP, the SDK, and HTML/CSS export speed the trip to engineering. They do not check contrast, focus order, long copy, or the button the team maintains in 14 variants. Whoever receives the export needs to know whether they are looking at exploration or a contract.
What to measure when a prototype arrives in ten minutes
Vibe design changes expectation at the same speed as the tool. If a PM generates five directions in ten minutes, the next request tends to be "pick one and close it." Without an explicit calendar block for the path from 1 to 100, refinement becomes slack nobody named.
Record the interval between the first shared prototype and the moment the screen enters the system with tokens, empty states, errors, mobile, and real content. Count how many generations die in exploration and how many enter design review. If the second count grows and the first stays flat, the team is accumulating visual debt that looks like speed.
The free plan with 350 generations a month covers an ideation trial. Generation volume is an exploration metric. Consistency, accessibility, and hours spent fitting the screen into the system remain the production metrics. Keep the two families apart in the stakeholder report, or the meeting only sees a full canvas.
The checklist Voice Canvas does not inherit
"I want something premium and minimal, like Stripe" describes atmosphere. It does not describe the spacing token, the disabled state, the form with a validation error, a 40-character label on a button, or the screen reader path through a modal. The agent can interview and critique in real time. The team's checklist stays with the team.
Before promoting a direction, run what you already require in the production file: contrast, touch target, keyboard, extreme content, i18n if the product has it. A screen that only works with short copy and one breakpoint is still a prototype, even if it is pretty and clickable.
DESIGN.md is context for the agent, useful so you do not reinvent color and type on every Stitch project. If the product's source of truth is Figma plus code, the markdown needs to be born there. Reverse the order and the team gets two systems that diverge by the second sprint.
Exporting a frame does not move governance
The market read Figma export as a land grab. Exporting a frame is cheap. Maintaining 200 components, three breakpoints, and the library engineering consumes is the work that shows up in cost.
A policy you can sustain: Stitch for discovery and alignment; Figma for components, handoff, and visual QA. Anyone wiring Cursor or Claude Code through MCP still needs an acceptance criterion in code. Pixels that resemble the generated mock do not prove the product fits the system.
What worries me is the design deadline shrinking on the calendar while the 1-to-100 work remains. If the team adopts Stitch, kickoff has to carry the number of days (or hours) still reserved for refinement. Without that number written down, fast generation only brings the date fight forward.
If your company is weighing something similar, reach me on LinkedIn or WhatsApp.
