Vibe-Coded MVP: What Design to Keep, Fix or Redo
Redesigning a vibe-coded app from Lovable, Bolt or v0? What to keep, fix or redo before enterprise buyers arrive: states and roles first, then the look.
Your Lovable prototype passed the demo. The buyer liked it and asked for a trial. A week later their operations lead invited a colleague as a “viewer”, and the viewer deleted a customer record.
Nobody had designed what a viewer can and can’t do. The demo never needed it.
That’s the gap between a product that looks finished and one that holds up. You usually don’t need to rebuild. You need to harden the parts a demo never touches: what each role can do, and what the screen shows when something is empty, slow, broken or forbidden. Fix those first. Then deal with the look, which is the smaller problem.
This post gives you:
- the four places a demo hides design debt
- a keep, fix or redo table, layer by layer
- the first five changes, in order
- one test that shows whether a screen looks generated
How we approach AI products is on our AI-SaaS product design page. If an agency built your V1 and you’re not sure you even have editable files, start with who owns the Figma files when your agency leaves.
Why “it looks done” is the wrong test
AI builders are good at the screen you asked for. They’re much less reliable at the parts you didn’t ask for. A context window can’t hold what a designer learns in ten years of shipping products.
Security shows the pattern most clearly. In March 2025 a researcher scanned 1,645 projects built with Lovable and found about one in ten exposing data, because nobody had set the database rules that keep one user’s data away from another’s. It’s on record as CVE-2025-48757, which Lovable disputes. Veracode tested more than 100 AI models on 80 coding tasks in 2025. In 45% of the tests, the generated code had a known type of security flaw.
Security belongs to your engineers. The design lesson is the same, though. The path you showed in the demo works. Whatever you didn’t specify wasn’t built.
The design debt a demo hides
A demo has one user, one role, sample data and a good day. Real use breaks a generated product in four places.
States nobody designed are the first. Empty, loading, error, partial, expired, permission denied. A generated app usually has the full version of its main screen, and a spinner or a blank page for the rest.
Roles come next, which is where the viewer deleted the record. Admins and members see the same screen. People can do things their role shouldn’t allow. In a product used by several companies, one company’s data can show up in another’s search.
Then there are one-off components. Each prompt created its own button, table and pop-up. You end up with four shades of grey and three corner styles, none of which have names. Code drifts even with a design system in place, and linting only helps if it runs every day.
Last, there’s no single source of truth. The design lives in a chat history. If a Figma file exists, it hasn’t matched the code for weeks.
None of this is visible on a laptop in a meeting room. All of it shows in week two of an enterprise pilot.
What to keep, fix or redo
| Layer | Usually keep | Usually fix | Redo when |
|---|---|---|---|
| Core flows | The main job, if users already finish it | Extra steps the prompt added “just in case” | Users reach the end and don’t come back |
| Hierarchy | Navigation that matches how users talk about their work | Screens where everything looks equally important | A buyer can’t tell what the product does from the main screen |
| Components | The basic building blocks from your UI library | One-off versions: merge them into a small named set | There’s no shared library at all |
| States | Whatever exists for the main screen | Empty, error, loading and permission states everywhere else | You can’t list which states exist |
| Roles | The data model, if engineering signs it off | What each role sees and can do, screen by screen | Permissions are only enforced in the interface |
| Visual style | Your colours, if you chose them | Type sizes, spacing, density | It’s the library’s default theme |
Use the table as a starting point. The last column matters most. Redo is for layers with no foundation, not for layers you don’t like.
The first five changes, in order
- Write a one-page UI spec. What the product is for, who uses it, the roles, the three main flows and the components you allow. Founders who have been through this keep giving the same advice: don’t rewrite everything, but write down what the product is before the next prompt. The spec also stops the next AI-made screen from undoing your work.
- List every state. Put every screen in a row and every state in a column, then mark what exists. In a generated app, expect only the first column to be full.
- Map the roles. Put every role in a row and every sensitive action in a column. Agree it with engineering, because what the interface allows must match what the server enforces. This is the fix for the viewer who could delete.
- Set a type scale and design tokens. A few text sizes, a spacing scale and named colours. Everything after this step stays consistent.
- Merge the components. One button family, one table, one pop-up pattern. Replace the one-offs as you touch each screen, not in one big rewrite.
Only then change the look.
Why the generic AI look costs trust
In April 2026 a developer scored about 1,600 Show HN launch pages for the patterns AI builders default to. Roughly one in five had four or more. The Hacker News discussion drew 235 comments, and the giveaways were familiar: the purple gradient, the same Inter headline, a little badge above it, rows of identical cards with an icon on top. Many commenters read that look as low effort, and your buyers are likely to read it the same way.
Everyone builds with AI now, and buyers know that. What they react to is a product that feels generated: default type, default spacing, stock illustrations, and copy that could belong to any company in the category. In a Harris Poll in the UK, about seven in ten people trusted ads made entirely by people. About four in ten trusted ads where people used AI tools, and about two in ten trusted ads made by AI alone. That’s advertising, not software. Still, the direction is clear. Work a person steers earns about twice the trust of work AI does alone.
Steering means three things in practice. A person decides the type, density, colour and tone before anything is generated, and AI works inside those choices. Most of what AI produces gets thrown away, because its value is speed of options and the designer’s value is choosing. And the work goes through several rounds on type, spacing, copy and states, until the product looks like your company and nobody else’s.
One test works on every screen: could a careful designer defend each choice on it? If the answer is “the model picked it”, that’s the next thing to change.
For the visual layer, the fastest order is type and density, then colours, then the words on the screens, then one interaction people remember. Founders tend to skip the words. They’re what makes interchangeable screens sound like a specific company.
We use AI every day at Spirixx. It summarises recordings, drafts variations and speeds up the making. The decisions stay with the two principals on your project, who review every output and send it back until it’s right. At the end of every two-week cycle you get a recorded walkthrough, so you can see those decisions and the reasons for them.
What this looked like on our last project
Our most recent client brought us a vibe-coded app with 57 designed screens covering four states. The product needed about 150. We started by sorting every design choice into two piles: decisions the client had made, and AI suggestions someone had accepted without questioning. That showed us what the client wanted, and what still needed work.
From the outside, it doesn’t look like starting from scratch. It still takes more time than anyone expects, because we rework the parts nobody had thought about. That’s design debt you can’t see in a demo. The earlier you deal with it, the fewer problems it causes.
FAQ
Can I keep the code and only fix the design?
Often, yes. If engineering confirms the data model and permissions are sound, the design work sits on top. If permissions only exist in the interface, that’s an engineering fix first.
Do I need a design system first?
Not a full one. You need enough to stop new screens drifting: a type scale, tokens and a small named set of components. A full design system can wait until more than one team builds on the product.
Why does every shadcn app look the same?
Because most of them ship the library’s defaults: the same type, the same corner radius, the same neutral colours. The library is a good foundation, but its defaults belong to everyone who installed it.
Should I rebuild or harden?
Harden, unless a layer has no foundation. Rebuild only what sits in the “redo” column. A full rebuild throws away every habit your users have already learned.
How many of your screens would still let that viewer delete a customer record today? To find out, sign in with your lowest role and try every delete, export and invite button on three screens. It takes under an hour, and it’s the start of the role map in step 3.
Two weeks is enough to decide what to keep. A Clarity Sprint gives you the keep, fix and redo list in writing, with what each fix is worth. Booked on its own, the Sprint carries a written guarantee: if the report doesn’t surface at least ten times its fee in annual impact, we refund the fee. You can see what a written audit report looks like first. Book a free 30-min call, or read how the Clarity Sprint works.
Sources
- NVD, CVE-2025-48757
- Veracode, AI-generated code security research (July 2025)
- Hacker News, Scoring Show HN submissions for AI design patterns (April 2026)
- Hacker News, Ask HN: Vibe-coded prototypes in production (January 2026)
- Harris Poll UK, The authenticity trade-off
- Product Hunt, After vibe coding, how do you take your product to the next phase?
