Vibe coding to build a website: what works and what doesn't

Vibe coding is everywhere right now. And as often happens when something rises fast, opinions are either very enthusiastic or very sceptical. We'll try to be neither.

We've been using AI in our workflow for a while, including Webflow MCP v1.2 since this quarter. What we write here is our reading of the situation: not an absolute truth, and certainly not a judgement on people who make different choices. We may be wrong on some points. But this is how we see things, and why.

‍

What is vibe coding?

The term comes from Andrej Karpathy, an AI researcher and OpenAI co-founder, who used it in early 2025 to describe a way of programming that differs from the traditional approach. Instead of controlling every line of code, you describe what you want to an AI, it generates, you iterate. The idea is to "give in to the vibes" with the AI rather than master everything technically.

In web design, it works like this: you open a tool, describe the website you want, the AI generates a first version, and you keep going through successive exchanges until you get the result you want. No hand-written CSS, no manually configured components. You prompt, you refine, you deploy.

The tools that popularized this approach are Lovable, Bolt.new, Base44, Cursor with assistants such as Claude or GPT-4, and to some extent the AI features built into platforms like Framer or Webflow.

‍

What does vibe coding do well?

There are obviously undeniable advantages.

‍

Speed of execution

A landing page, a portfolio site, an offer page with a form: a motivated person can have something presentable in a few hours. Not in a few minutes as marketing demos suggest, but still in a few hours, compared with several days with a traditional approach. That's significant.

‍

Accessibility for everyone

Someone with no coding knowledge can build something that works, iterate on the design, test different wording and edit their page without depending on a developer. It's a real shift in access to web creation.

To validate an idea or test an offer, it's often enough. A capture page for a beta launch, a quick presence for a project that is still uncertain: vibe coding meets that need without over-engineering. And for that, it doesn't need to be technically perfect.

‍

Low cost

The entry cost is much lower. Comparing prices with an agency means comparing two different products (more on that below). But for someone with a limited budget and a clearly defined need, the financial argument is real.

‍

Who is vibe coding good for?

Asolopreneur launching an offer, a freelancer building their online presence, someone testing a concept with few resources: building a website with AI is probably a very good decision. Fast, sufficient, affordable. It's exactly the right tool for that situation.

‍

The question gets more complicated when the website is no longer a personal presence or a test, but a growth tool shared across an organization. When a marketing team has to work on it regularly. When the CMS has to be managed by people who weren't there when it was built. When the website changes often, runs A/B tests and generates leads for a sales team.

‍

It's not that we think vibe coding is bad in that case. It's that we think the questions that matter are different, and vibe coding tools don't naturally raise them. What happens after launch, who can edit the website, how it evolves, what happens when someone else has to work on it: that's where things get complicated, in our view.

‍

The real problem with vibe coding: it's a closed system

Visual quality isn't really the problem with a vibe-coded website (although, left to its own devices, it will look like 80% of vibe-coded websites, and people are starting to recognize them).

‍

The problem is mainly that it's a closed system.

When you vibe code a website, you build in a logic where the AI controls the structure. You can ask for changes and the AI makes them, but you don't really have a hand on what happens underneath. You don't know how the components are organized, you don't control how styles are named, and you can't make a precise change to a specific element without going through the AI.

‍

For personal use or a test project, that dependency is often acceptable. You're the only person involved, you know the tool, and if something goes wrong you can start a new prompting session to fix it.

For an organization with several people working on the website, it's another story. You become dependent on the AI for every change, even a minor one. And if the tool evolves, changes its terms, or your subscription changes, you have no clean codebase to fall back on. It's a form of control debt that builds up.

‍

The problem with who vibe codes inside a company

There's something we often see in discussions on the subject that seems underestimated to us.

In a company, the people who turn to vibe coding to create or edit a website are often marketers, designers and project managers. Rarely developers. And that makes sense: the whole point of vibe coding is precisely not needing technical skills to build something.

‍

The problem is that the AI can make technical decisions whose consequences these profiles have no way of assessing. Will the way the components are structured cause performance issues in six months? Will the way the CMS is organized make it impossible to integrate the marketing tool we'll want to connect next year? Are the choices made in session 1 and session 5 consistent with each other?

A developer would see these questions. A marketer who vibe codes, in most cases, won't. Not because they're less skilled in their field, but because it isn't their field. And the AI won't ask these questions for them unless it's asked to.

‍

Technical debt can therefore build up with no visible warning sign, because the website keeps working visually for a while. That's the scenario we find most likely in organizations that bring vibe coding into their processes without technical oversight.

‍

This remains our hypothesis, not a certainty. We don't know exactly how these situations play out in practice for teams that make this choice. But it's the structural risk we see.

‍

Design systems and AI in production

There's a link between this question and how effective AI is on a website in general.

A structured design system, well-defined variables, consistent components, understandable naming: that's what makes AI effective and reliable on a website. When we use Webflow MCP v1.2 in our workflow, we see the difference concretely. On a well-organized website, the AI can work consistently with what exists. On a website with a vague or inconsistent structure, it improvises, and high-speed improvisation produces results that are hard to predict.

That's not an argument against vibe coding, though. It's rather the observation that AI in production is all the more useful when it has a structured foundation to rely on. And vibe coding, by nature, rarely produces that structured foundation.

‍

The tools: an honest overview

Lovable is probably the most used tool right now for building websites and apps through prompts. It understands natural-language requests well and produces something visually decent quite quickly. Like all these tools, the underlying structure is what it is: you have little visibility into it and little direct control.

Bolt.new (by StackBlitz) works in a similar way. Very useful for prototyping. For a website that will evolve over several months with different people, the same questions arise.

Base44 is starting to position itself for slightly more complex projects, with better component management. It's an interesting direction, though it doesn't fundamentally solve the question of control.

Framer with its AI features follows a different logic, closer to AI-assisted no-code than pure vibe coding. Generally more stable and predictable, at the cost of less flexibility.

The AI assistants from Wix, Squarespace or Hostinger are designed for the general public and quickly produce something functional for simple needs. They are probably the clearest options for the uses they're intended for.

In every case, the limit is structurally the same: the tool generates what you ask for. If you don't know what to ask in terms of organization and maintainability, it won't ask those questions for you.

‍

What questions should you ask before using vibe coding?

Two practical questions, not meant prescriptively, just because they seem useful to us.

The first: who will maintain this website, and with what level of technical oversight? If you're on your own or have a developer available to keep an eye on things, vibe coding can work with some rigour. If the website will be managed exclusively by non-technical profiles, the question of control deserves to be asked upfront.

The second: is this website meant to exist or to grow? A website designed to validate an idea quickly and a website designed to generate leads sustainably don't have the same requirements. It's not a value judgement, just two different uses that probably call for different approaches.

‍

Our agency's take on vibe coding, in short

Vibe coding is useful. For the right use and the right profile, it's often the best option available, and we're not trying to make it controversial.

We believe that its main benefit—letting anyone build quickly without technical skills—is also what creates the biggest limitations for organizations with requirements for maintainability, collaboration, and longevity. Using these tools for a personal project is not the same as building them into the workflow of a marketing team running a growth website.

We may be wrong about this. The tools are evolving fast, and some of the problems we identify may be solved within 12 months. But that is our assessment today.