AI App Builders From Prompts: Why Users Still Hit Walls

Prompt-to-app tools like Figma Make, Lovable, v0 and Bolt are being sold as the future of software creation. Describe an idea in plain language and get a working interface back. The people actually using them describe something messier. These tools are useful for quick demos, but getting a result someone is happy with takes far more effort than the pitch suggests.

What users are running into

The sharpest criticism comes from designers testing Figma Make, Figma's entry into AI prototyping. One widely discussed critique says the tool produces "generic and bland" web prototypes. The poster argues those prototypes are hard to tell apart from what Lovable, v0, ChatGPT or Claude already generate, and that Figma has drifted from its core identity as a place for precise, expressive design 1. Other commenters called Make a "low hanging fruit" and said Figma's feature releases often feel unfinished 1.

The hands-on accounts are more concrete. One designer built a fairly simple side panel: text-based cards, a few icons, a fake search input and a toggle filter. It took roughly 160 to 170 prompts to finish 2. They described constantly fixing elements that broke or shifted, rolling back to earlier results, and eventually editing code by hand while wishing for a proper IDE 2. They also hit a usage cap. They noted that Figma's documentation allows organization users around 60 to 70 prompts per month, which is well below what that one panel needed 2.

An agency comparison of Figma Make, Lovable and v0 found similar problems. Figma Make, still in beta at the time, tended to forget earlier design decisions and overwrite them with new prompts. It also produced visual defects such as poor contrast between buttons and their labels, and it only supported web designs 3.

Lovable: more power, more burden

Lovable fares better on capability in the same comparison. It can wire in Stripe payments, databases and user authentication, offers a Figma-like layout editor, and can help debug its own output, which the agency says Figma Make couldn't match 3. The tradeoff is who can use it well. The reviewers note that Lovable leans toward developers who can review, edit and fix generated code, so non-technical users may face a steep learning curve 3.

That tension runs through the whole category. The tools promise that anyone can build an app from prompts. But the more an app can do, the more it seems to require someone who understands code to keep it working.

The optimistic case

Not everyone is gloomy. Product writer Aakash Gupta, who tested Figma Make against Bolt, v0, Cursor and Lovable, reports teams adopting Make since it launched in May and became generally available in July 4. His argument is that the underlying AI has become a commodity. In his view, everyone is building on Claude and everyone can produce working code, so the real differentiator is the surrounding workflow, and Figma's position inside existing design processes is hard for rivals to match 4. He still concedes that Make is focused on UI components rather than full applications, works best for web interfaces, and has fewer framework options than the standalone tools 4.

A roundup from Banani, itself an AI prototyping vendor, declares 2026 "the year of AI prototyping tools." It slots Lovable and v0 as the picks for developers and Figma Make for prototype-to-code handoff 5. Its top overall pick is Banani's own product, so that ranking is best read as marketing rather than independent assessment 5.

Where the sources agree and diverge

The sources disagree on tone but agree on the main facts. Even the most enthusiastic account accepts that these tools generate interfaces and lightweight apps rather than production software 4. The most critical accounts accept that they're fine for fast mockups or usability-test prototypes 12. One Reddit commenter made that point directly, arguing Make isn't meant for production and that serious work belongs in a professional IDE connected to a design system 2.

The real disagreement is whether "good enough and already in your workflow" is a winning formula 4, or whether generic output and forgetful iteration undercut the value designers pay for 13.

Our read

Building apps from prompts works best when expectations are set correctly. These tools are fast sketching engines, not replacements for development or careful design.

The friction users report is not a minor bug. Prompt caps, regressions and the need to drop into code all point to the same limitation. Natural language is an imprecise way to control software once a project moves past the first draft. Gupta's commoditization argument looks right, but it cuts both ways. If every tool runs on similar models, then reliability, memory of past decisions and fair usage limits will decide which tools last.

Until those improve, a 160-prompt side panel says more about the state of the category than any launch announcement or funding headline.