Vibe Coding.

The term goes back to a post by Andrej Karpathy in February 2025. He described a way of programming in which you rely entirely on the model, try the result, and on errors simply hand back the error message instead of reading the code yourself. Collins made 'vibe coding' its word of the year in November 2025, and at that point the expression lost its precision: today it covers everything from careful AI-assisted development with tests and review to clicking an app together unchecked. For an honest conversation about budget and risk it is worth defining the term before using it.

In practice vibe coding runs as a loop. You describe in a sentence or two what should exist, the model produces files, you start the application, report back what does not work, and have it corrected. Modern development environments and coding agents run commands themselves, read error messages and change several files in one pass. The strength of this loop is speed: hours rather than weeks from idea to first running version. That is also its weakness. What emerges is a version that runs — not necessarily one that suits a second developer, a load peak, or a security review.

The risks are well documented and worth treating soberly. Code nobody has read typically contains credentials in the source, missing permission checks on endpoints, unvalidated input, and dependencies nobody consciously chose. A second, less obvious effect follows: without tests and without an understood architecture, fear of change grows. Every new feature can break something nobody can explain, and the model then repairs symptoms instead of causes. The usable rule of thumb is therefore not 'never' but: the greater the damage from a failure, the stricter the human control — up to review, automated tests and a security check before going live.

A concrete case shows the sensible boundary. An SME wants to know whether a digital booking form with automatic scheduling would be used at all. Instead of writing a specification, a prototype is built in a few days and tried out by twenty existing customers. The goal is a decision, not a finished product: if the form is used, the investment is justified; if it is ignored, little money is lost. That is exactly what we build prototypes and MVPs for, from CHF 10,000. If the decision is to proceed, the prototype is not dragged along but rebuilt with tests, clean data handling and review.