Vibe Coding: Karpathy's Term, Its Evolution, and What It Means in 2026
Karpathy coined 'vibe coding' in February 2025. It became Collins' Word of the Year. By 2026, the term had split into two very different practices. Here's where it stands.
In February 2025, Andrej Karpathy posted a description of a new mode of software development he'd been practicing: "I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works." He called it "vibe coding."
Merriam-Webster listed it as a "slang & trending" expression the following month. Collins English Dictionary named it Word of the Year for 2025. The term spread from developer circles into general business discourse faster than almost any technical coinage in recent memory.
By 2026, "vibe coding" had accumulated enough conflicting definitions that it was nearly unusable as a precise technical term. Which is interesting, because Karpathy's original framing was quite precise — and the drift tells you something important about both the concept and the field.
The Original Definition
Karpathy's coinage was specific about both the method and the scope. Vibe coding meant handing the implementation layer to the model and staying at the level of intent. You describe what you want. The model generates code. You run it, observe the result, describe what needs to change. You do not read the code in detail. You do not own the implementation — you guide it.
The original context was casual software development: personal tools, scripts, prototypes. Karpathy's framing explicitly didn't claim this approach was appropriate for production systems with safety requirements or complex architectural constraints. It was a description of what he personally did for a certain class of problem.
The core claim was: for simple software that mostly works, staying at the intent level and delegating implementation to the model is faster than writing code yourself, even if you can write code well.
The Bifurcation
By mid-2025, "vibe coding" had split along two fault lines.
The first meaning — the one that traveled to non-developer audiences — meant roughly "using AI to generate code without understanding it." This interpretation was heavily amplified by critics arguing that vibe coding produced insecure, unmaintainable, and unreliable software. The critics weren't wrong about the risks; they were describing a real failure mode of the approach when applied without judgment.
The second meaning, adopted by working developers who took Karpathy's original framing seriously, meant something more nuanced: staying at the level of intent for appropriate tasks, with judgment about which tasks qualify. This version acknowledged that vibe coding is a spectrum and that "appropriate tasks" is doing a lot of work in that description.
Karpathy himself helped clarify the distinction later in 2025 by drawing a sharper line between vibe coding (intent-level delegation, appropriate for casual/prototype contexts) and what he called agentic engineering (structured use of AI agents for complex software, requiring deep technical discipline). The two practices involve similar tools but very different levels of oversight and engagement.
What Happened to the Term by 2026
The New Stack ran a piece in 2026 titled "Vibe Coding is Passé" — capturing the arc neatly. Among professional developers, "vibe coding" as Karpathy originally described it had become neither cutting edge nor controversial. It was just... how you built prototypes. The interesting questions had moved to agentic engineering: how do you structure complex multi-step tasks for AI agents, how do you verify their output, how do you catch the failures that compound across sessions.
In business contexts, "vibe coding" still got invoked as a somewhat alarming term for "AI writes your code without proper oversight" — but even this usage was fading as AI-generated code became normalized enough that the term for it didn't need to be borrowed from a casual developer coinage.
The term did real conceptual work in 2025 by naming something that had been happening but hadn't been named. Once named, it could be discussed, critiqued, and bounded. That's what good coinages do.
The karpathy-guidelines Skill as a Corrective
The karpathy-guidelines skill can be read as a partial corrective to naive vibe coding. If vibe coding at its worst means "generate code, run it, don't read it, iterate until it works," the karpathy-guidelines constraints — simplicity first, surgical changes, goal-driven execution — are exactly the discipline that separates useful vibe coding from reckless vibe coding.
"Surgical changes" prevents the model from rewriting large sections of code when you've described a small modification. "Goal-driven execution" prevents the model from iterating on vague success criteria. "Think before coding" prevents the model from guessing about ambiguous requirements.
These aren't rules against vibe coding — they're rules that make vibe coding less fragile. The behavioral constraints in the skill encode the judgment that separates appropriate AI delegation from abdication.
The Tension That Remains
The residual tension in vibe coding discourse is between two genuine positions that haven't been reconciled.
Position one: delegating implementation to the model is appropriate for the vast majority of what software engineers actually do day to day. Most code is not novel, not safety-critical, and not architecturally complex. For most code, the bottleneck is specification, not implementation — and vibe coding addresses that bottleneck correctly.
Position two: the cumulative effect of intent-level delegation is a codebase that no one understands deeply. Technical debt accrues not in bad code but in diffuse ownership. When something breaks at an architectural level, no one has the mental model to fix it.
Karpathy's own position, as best as it can be read from his public commentary, is that both positions are right for different tasks — and that the key skill is knowing which tasks fall into which bucket. The karpathy-guidelines skill is designed for situations where you're using the model for real coding work, not casual vibe coding. The two practices require different defaults.
For more on how Karpathy's views on agents connect to vibe coding's evolution, see Karpathy on Agents: His Public Takes on Autonomous Coding.
Part of the Karpathy on Claude Code series. Published 2026-05-23.