Vibe coding isn't enough

Vibe coding has gone mainstream: anyone can spin up a service, a script, or a frontend in an afternoon. The floor for “have something running” has dropped to zero, but the ceiling didn’t. Reviewing code critically, judging whether the architecture will survive contact with real data, anticipating where it will break at scale: none of that has been solved with GenAI. If anything, those skills are scarcer relative to the volume of code being produced.

The same split runs through creative AI tools: image models nail the look but fail on details; music models like Suno make tracks that sound great but resist tweaks; LLMs draft fluent papers that are wrong in the specifics. None of those issues are easy to fix, and with code, it’s at its worst.

A prototype from a single prompt is impressive, but if the architecture underneath is wrong (wrong data model, coupling, or assumptions about scale), it can’t be fixed by prompting. The output isn’t a draft that can be refined later; it’s a foundation you’re already standing on.

None of this is theoretical: we’ve used AI assistants across dozens of projects, and the same failure modes keep surfacing:

• Code that works on the demo input and falls apart on real data shapes.
• Modules generated together and so tightly coupled that refactoring one means rewriting the rest.
• Plausible abstractions that don’t hold up against the actual domain.
• Performance that’s fine at POC scale and catastrophic in production.
• Tests that validate the implementation instead of the spec.

Because the model is locally consistent with what it generated before, additional prompts double down on the bad architecture instead of questioning it. Recognizing this and knowing when to throw the first attempt away and rebuild is engineering judgment, not prompting skill.

This is why teams need experienced engineers more than ever: the new floor of “writing working code” is accessible to anyone with an AI assistant, but the ability to understand code, find what’s wrong with it, and decide what to do next still requires expertise. For the last several years, we’ve constantly used coding assistants across dozens of projects. We can see the progress in the quality of the working solutions they produce. But we still need to think carefully about a project’s architecture and review the resulting code, because the truth remains simple:

GenAI produces artifacts. Engineers build systems.