Against Vibe Coding (suhas.org)

A curated roundup of recent essays argues that “vibe coding” — the practice of letting LLMs spew code without rigorous specs or human scrutiny — is creating serious long-term problems. Writers from multiple corners of the web describe the same pattern: fast, large volumes of generated code that often look plausible but are wrong, inconsistent, or unmaintainable. The collection surfaces cultural pushback and evolving consensus that hobbyist prompt-driven coding doesn’t scale; instead it produces “comprehension debt,” “AI spaghetti,” and a market for expensive cleanup services. Technically, contributors point to concrete failure modes: LLMs rarely ask clarifying questions, first attempts can be mostly garbage (one post cites a ~95% failure rate), generated code introduces redundant functions, mismatched styles (classes vs functional patterns), and misused global configuration — all of which multiply maintenance costs. The debate also highlights a role shift: successful teams are moving from laissez-faire prompt-coding toward “vibe/context engineering” — formalizing inputs, specs, and orchestration — and valuing developer judgment and taste over raw promptcraft. The implication for AI/ML teams is clear: LLMs can accelerate output but only when paired with rigorous verification, domain knowledge, and architecture-level stewardship to avoid turning short-term productivity into long-term technical debt.