Vibe coding : un mode, pas une méthode
Two ways of using AI code generation, and they answer different questions. A **mode** is a behaviour you adopt for a specific context, then drop. A **method** is a system you apply all the time. Vibe coding, describing what you want and letting the AI write the code without reading all of it, makes a great mode: the prototype, the proof of concept, the two hours spent checking that an idea stands up. Run as a method, on a product that lives for months with real users, it always hits the same wall: rewrite after rewrite, bugs that keep coming back, and fear of touching the code because nobody knows any more what holds what together. The distinction comes from Willy, who publishes on AI agent governance. His picture for it: teaching yourself to swim feels like real freedom in a pool, and gives out the day you cross a lake with a current. What takes over is not dropping AI, it is a switch to governance: decisions written down, context kept in order, rules added as real friction shows up. Vibe coding is not the opposite of that governance, it is the exploration phase that comes before it. That wall has a mechanism, which Willy describes in another video ([Instagram reel DdKE1NKIHH-](https://www.instagram.com/reel/DdKE1NKIHH-/)) as three debts feeding into each other. A spec written too fast leaves gaps: that is **documentation debt**. The agent runs that gappy spec and ships a product in its image: that is **technical debt**. You then spend your evenings fixing what keeps breaking, and that mental load is **cognitive debt**. The symptom shows from far off: you work for your agents instead of working with them.