# Vibe Coding Term coined by [[Andrej Karpathy]] in February 2025 to describe a coding approach where you let LLMs generate all the code while providing goals, examples, and feedback in natural language. The programmer shifts from writing code to guiding, testing, and giving feedback. Karpathy described it as "fully giving in to the vibes, embracing exponentials, and forgetting that the code even exists." ## How it works - Describe what you want in natural language - Let the LLM generate the code - Accept the results without necessarily reviewing every line - Iterate by describing problems or desired changes - Karpathy used tools like Cursor Composer with Sonnet and SuperWhisper for voice input ## Simon Willison's distinction [[Simon Willison]] drew an important line between vibe coding and responsible AI-assisted programming. His view: - **Vibe coding**: Building software without reviewing the code the LLM writes. "Irresponsibly building software through dice rolls, not caring what code is produced" - **Vibe engineering**: Engineers at the top of their game using AI tools responsibly to accelerate their work - His rule: never commit code he couldn't explain exactly to someone else. If you've reviewed, tested, and understood all the LLM-generated code, that's not vibe coding — that's using an LLM as a typing assistant ## Risks - Code quality and maintainability suffer without human review - Security vulnerabilities may go unnoticed - Technical debt accumulates quickly - Difficult to evolve or debug code you don't understand ## Context Builds on Karpathy's 2023 claim that "the hottest new programming language is English." Karpathy later suggested vibe coding is fading as "agentic engineering" becomes the next phase, where AI agents handle more structured, multi-step development workflows. At Sequoia AI Ascent 2026, Karpathy sharpened the contrast with a "raises the floor / raises the ceiling" framing: vibe coding *raises the floor* (anyone can produce working software in English), while [[Agentic Engineering]] *raises the ceiling* (skilled engineers coordinate fallible agents to ship faster while preserving quality, security, and taste). Both are facets of the broader [[Software 3.0]] paradigm, where prompts ARE the program. The everyday experience of using LLM-driven tools that "live on your computer" (Claude Code as a spirit/ghost; voice-driven coding) is what Karpathy calls *ambient programming* in his 2025 LLM Year in Review. ## References - Karpathy's original post: https://x.com/karpathy/status/1886192184808149383 - Willison's take: https://simonwillison.net/2025/Mar/19/vibe-coding/ - Wikipedia: https://en.wikipedia.org/wiki/Vibe_coding ## Related - [[Vibe Engineering]] - [[Agentic Engineering]] - [[AI Engineering]] - [[Cognitive debt]] - [[Unreviewed AI code anti-pattern]] - [[Code is cheap, quality is not]] - [[AI Agents]] - [[Claude Code]] - [[BMad Method]] - [[Simon Willison]] - [[Andrej Karpathy]] - [[Slopsquatting]] - [[Software Supply Chain Security]] - [[Software 3.0]] - [[Menugen Architecture Pattern]] - [[MenuGen Deployment Gap]]