There and back again: integrating vibe coding and programming foundations in a three-phase course structure

For decades, programming education has rested on a quiet assumption that one of the main skills being taught is the skill of producing code. Generative AI has unsettled that assumption. In what Andrej Karpathy termed vibe coding, a programmer looks to express intent in natural language and an AI system looks to help translate this into executable code. Such practice is no longer marginal, AI-assisted tooling is now embedded across professional software development, reshaping how programming work itself is valued and performed. This con- fronts educators with a question: if machines can help write the code, what is programming knowledge for?
The tempting answers are the extreme ones. One says foundational knowledge is obsolete, and curricula should focus on teaching students to direct AI systems. The other says AI assistance is but a crutch, to be banned until foundations are secure. Both treat programming competence as a single dimension along which AI either advances or impedes students. We argue that the reality emerging from early studies of AI-assisted development is more nuanced. Reliance on vibe coding alone can foster beliefs that AI substitutes for foundational knowledge and can produce an illusion of competence, leading to uncritical acceptance and helplessness when generated code fails, a pattern that has prompted explicit caution against over-reliance. Yet studies also reveal what we might call the Vibe Coding Paradox, where AI-assisted development has been shown to work best precisely for those who already know how to program. The machine that supposedly replaces programming knowledge in fact rewards it.