Skip to content
uux
UX Design7 min read

What v0, Lovable, and Bolt Output Still Needs Before Production

A concrete punch list of what v0, Lovable, and Bolt prototypes still need before they can ship to paying users without regret or rework.

Timothy Carter

Timothy Carter

Chief Revenue Officer ·

What v0, Lovable, and Bolt Output Still Needs Before Production

Every output from v0, Lovable, and Bolt ships with the same four production problems: accessibility shortcuts, a backend with the locks off, performance regressions the preview hides, and a design system that remembers only the current screen.

None of these are hard to fix once named, and all of them are expensive to find after launch.

The problem is never that the preview looks wrong. It is that the preview was built to look right, not to run. Each tool skips different parts of the production stack, so the handoff work is different for each.

Where Every Vibe-Coded Output Falls Short

Three findings frame the common ground. First, the developers using these tools do not trust the output. Stack Overflow's 2025 survey found that 46% of developers distrust AI tool accuracy, up from 31% the previous year, even as 84% report using or planning to use AI tooling in their workflow.

Second, AI-assisted output requires more rewriting than it replaces. GitClear's analysis of 153 million lines of code found that code churn (lines reverted or updated within two weeks of being written) was on pace to double in 2024 versus the 2021 pre-AI baseline, alongside a measurable decline in refactored and moved code.

Third, the accessibility floor is low across the modern web, and generators do not raise it. The chart below sets the WebAIM Million benchmark against the standard it is meant to meet.

For a vibe coded prototype, that floor is usually the ceiling. Generators default to styled divs over semantic HTML, inputs without labels, and contrast that fails against brand palettes the moment a founder overrides the defaults.

Home pages free of detected WCAG 2 A/AA failures (2025)
Home pages free of detected WCAG 2 A/AA failures (2025)Top 1,000,000 home pages: 5.2%Pages with no detected failuresWCAG 2 A/AA targetTop 1,000,000 homepages5.2%WCAG 2 A/AA target 100%
The accessibility floor the generators inherit: 5.2% of top-million home pages pass, against a 100% WCAG 2 A/AA target. Source: WebAIM Million, 2025

v0 Ships a Design System, Not a Product

v0 writes React and Next.js components in shadcn/ui and Tailwind, and it writes them well in isolation. The problem is that isolation. Routes do not share tokens, spacing drifts between runs, and the composite system that holds a product together has to be assembled after the fact. A route-by-route audit usually finds brand colors crashing through contrast thresholds on buttons the generator thought were fine.

Backend is the second shortfall. A v0 scaffold ships without auth, data persistence, or background work. For a landing page that is not a problem. For a vibe coded prototype that needs to accept sign-ups or write to a database, a backend has to be added, and the token system that keeps the Figma to production handoff from becoming a redo tax has to be imposed on the generator's output rather than trusted inside it.

Before v0 to production is a real claim, the output needs shared design tokens, a real backend, and an accessibility sweep on every custom composite: modals, comboboxes, menus, toasts.

A movie-set building facade seen from the side, revealing wooden bracing and empty space behind a convincing front wall.

Lovable Opens the Backend Before Launch

Lovable produces a full-stack app backed by Supabase. The database is real on day one, which is the appeal, and which is the problem. The default posture is open.

CVE-2025-48757, disclosed in May 2025, documented that Lovable-generated apps shipped with Row Level Security weak enough to let unauthenticated remote attackers read or write arbitrary tables. A later scan of more than 5,600 vibe-coded applications across Lovable, Base44, Create.xyz, Vibe Studio, and Bolt.new identified over 2,000 vulnerabilities, more than 400 exposed secrets, and 175 instances of exposed PII including medical records and IBANs. Wiz Research's January 2026 investigation of Moltbook, a vibe-coded AI social network, found a Supabase API key sitting in client-side JavaScript, RLS missing, and roughly 1.5 million agent API tokens plus plaintext OpenAI keys available to anyone who looked.

Before a Lovable production ready claim stands up, every table needs an RLS policy, every key has to move out of the client, admin and user boundaries need to survive a real adversarial session, and a PII inventory has to exist on paper.

Bolt Previews a Demo, Not a Deployment

Bolt.new runs full-stack code inside the browser using WebContainers. The preview is functional, which flatters the output and hides the shortfall. A Bolt.new app shipping to production usually needs its environment configuration rebuilt from scratch, because the browser runtime does not deploy anywhere real users will hit it. Environment variables, secrets, build pipelines, and test coverage are all out of scope for the preview.

Performance is the other blind spot. The in-browser preview is network-local, memory-local, and competing with no real workload, so Largest Contentful Paint readings look reasonable until the app sits behind a CDN and meets a cold user on a mid-tier phone.

Mobile vs desktop Core Web Vitals pass rates, 2023 to 2026
Mobile vs desktop Core Web Vitals pass rates, 2023 to 20262023: 36%; 2024: 44%; 2025: 48%; 2026: 49.7%Mobile pass rate (%) →Desktop pass rate (%) →2023202420252026
Mobile and desktop CWV pass rates move together, and mobile has not closed the gap. Source: shno.co, 2026

The scatter above plots mobile Core Web Vitals pass rates against desktop across four years. The two measures move together, and mobile never catches up. A vibe coded prototype deployed without a performance budget lands on the mobile line, not the desktop one, because that is where the audience actually is.

Before a Bolt.new app shipping is ready for users, it needs deployment and environment configuration, a bundle and LCP pass, test coverage on the paths real users will take, and a working error state on every integration the generator wired happy-path only.

What the Handoff Must Carry

The output of all three tools is a fast start on a surface. What they do not produce is the connective tissue a product needs: a design system that holds across every route, auth and permissions that survive adversarial sessions, states for empty and loading and error and partial data, copy that reads as if a human wrote it, and the instrumentation that tells the team whether any of it is working.

A design queue that owns this handoff keeps the work out of the engineer's week and off the founder's inbox. The live vibe code sessions model takes generator output and finishes it in-session, which is the shortest path from prototype to production-ready. The alternative, asking the generator to finish its own output through more prompting, is a recurring scramble dressed up as agility.

The economics are explicit once the loaded numbers come out: a subscription-based design queue closes shortfalls the generator reliably misses without adding a hire, and routes AI prototype to production work through a single intake rather than three separate contractors. That is the whole case for a managed design queue sitting between the generator and the ship.

When to Keep Iterating and When to Hand Off

Keep iterating in the generator when the surface is still the question: landing page variations, dashboard shell exploration, onboarding flow prototyping before the backend exists. The feedback loop inside the tool is faster than any handoff, and the output is cheap to throw away.

Hand the output off when any of three conditions holds. Real users will hit the app, which is a security question. The app writes to a database the business cares about, which is a correctness question. Or the generator has been re-prompted more than three or four times on the same screen without converging, which is a cost question. Any one of the three is enough. Two together is overdue.

Timothy Carter

About the author

Timothy Carter

Chief Revenue Officer

Timothy Carter is the Chief Revenue Officer at UUX.co, where he leads client strategy and revenue growth across a portfolio of design, software, and digital marketing brands. With more than 20 years in enterprise search, content, and demand generation, he has spent his career on the commercial side of the web — which is where most design decisions ultimately get judged.

That perspective shapes what he writes here. Timothy is less interested in design as decoration than in design as a revenue lever: how a landing page earns a click, why a pricing table converts or stalls, and what actually happens to a funnel when the UX debt finally comes due. He has worked with everyone from pre-seed startups to Fortune 1000 brands, and has seen the same handful of interface mistakes cost all of them real money.

Tim lives in Orlando, Florida with his family. When he is not deep in a conversion audit, he is usually on a golf course or chasing a fly-fishing trip somewhere with better weather.

Keep reading

Figma to Production Without the Redo Tax
UX Design7 min read

Figma to Production Without the Redo Tax

How to structure Figma files with Dev Mode, design tokens, and a handoff spec that developers can ship without rebuilding half the UI from scratch.

Samuel Edwards
The Real Loaded Cost of Three UX Design Models
UX Design7 min read

The Real Loaded Cost of Three UX Design Models

A side by side look at loaded monthly costs for a UX subscription, in-house designer, and design agency, with numbers product teams can actually use.

Timothy Carter