In February 2025, a tweet that went viral from Andrej Karpathy introduced the term “vibe coding”. Karpathy describes an experience where code becomes invisible, a natural dialogue with an LLM.
With vibe coding, you no longer need to type lines of code: you see, you say, it’s built. While this initial definition popularized the concept, it does not distinguish the two main types of vibe coding that have since emerged.
Two faces, two uses
The first, which we could call technical vibe coding, relies on tools like Cursor, Claude Code or Codex. It is aimed at technical profiles who maintain complete visibility over the generated code, can take control at any time and themselves calibrate the level of autonomy granted to the AI. The AI acts like a pair of peer programming: extremely talented and fast, but capable of making erroneous choices if left in control unsupervised.
The second, product vibe coding, is based on platforms like Lovable or Base44. It is aimed at non-technical profiles, founders, product owners, innovation teams, who wish to transform an idea into a functional prototype without going through a traditional development cycle. The promise: drastically lower the entry barrier and realize a vision in just a few clicks.
Two very different approaches, sometimes complementary depending on the phases of the project. But in both cases, without a coherent and holistic approach, the limits quickly become apparent.
Design: consistency cannot be improvised
In vibe coding mode, the starting point is often a simple conversational prompt. But precisely adjusting the screens requires multiple iterations. Without a design system formalized upstream with a logic of reusable and coherent components, the process becomes time-consuming, imprecise, and produces interfaces visually inconsistent with the brand identity. A return to traditional design tools or at least a design system logic is necessary to structure this step.
The prototype that turns into a product
This is the most insidious trap. The product is built through sometimes contradictory injunctions, added without a global vision of the architecture. Result: apparently functional code, but fragile, difficult to maintain and expensive to evolve. A “100% vibe coded” application, in just a few prompts, often constitutes a good prototype, rarely a good product. Nothing new in reality, in the world before, prototypes launched into production were rare.
Refactoring, the eternally unloved step
As in traditional production modes, breaks are necessary to review the code, restructure the architecture and optimize processing. This step is often neglected in the excitement of vibe coding. However, it can also be achieved with an LLM. Technical debt exists even with AI, dealing with it over time remains a good practice.
Safety: the one-eyed man in the land of the blind
An LLM generates functional code, not necessarily secure code. Without expert proofreading, which can also be assisted by AI, vibe coding can take shortcuts. Examples of hacked coded vibe applications are multiplying. Compartmentalize data, do not expose tokens and secrets, principles that remain valid.
A production organization to rethink
Native AI companies (Cursor, Anthropic, etc.), the new kings of this new era, are well aware of these limits. AI-augmented IDEs integrate with design tools like Figma, and develop rapid prototyping capabilities (Google Stitch, Claude Design from Anthropic). General public platforms (Lovable, Bolt, etc.) now provide access to the generated source code and allow it to be edited. Anthropic launches products dedicated to security. Etc.
But the issue goes beyond the tools: end-to-end orchestration remains a human problem.
In organizations and production processes designed with yesterday’s tools, how can production models be adapted? How to link design, code and other AI layers? What is the distribution of roles in a world where everyone can generate code? How to structure teams to get the best out of them?
The first adaptations are implemented, often by companies with a strong technical culture. Teams are tightening up to increase efficiency… one pizza is already too much.
At Doctolib, POs push code generated with AI, then reviewed and validated by a lead developer.
Ultimately, the solutions may be found in yesterday’s concepts revisited in the light of these new tools.
One thing is certain: the subject is no longer solely technical. Organizations must evolve radically or risk massive dropout.
