Vibe Coding: Just a Buzzword or the Future of Software Development?

“There’s a new kind of coding I call vibe coding, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.”

— Andrej Karpathy, OpenAI co‑founder and ex‑Tesla AI leader

In early 2025, a term called vibe coding began to circulate across developer forums, tech blogs and boardrooms. Originally coined by AI pioneer Andrej Karpathy, the phrase captures a simple but disruptive idea: instead of manually writing code, you describe what you want, and an AI model generates the software for you. As natural‑language models like GPT‑5, Gemini, Claude 3 and Runable’s own general AI agent become more capable, vibe coding promises to unlock new levels of creativity and speed in software development. But is it just hype—or are we witnessing a fundamental shift?

What Is Vibe Coding?

Vibe coding isn’t just “AI‑assisted programming.” According to the ISACA Now blog, it’s an AI‑driven approach where natural‑language prompts replace writing code. You might tell the system, “Create a secure login system with two‑factor authentication”, and the AI handles the syntax and API calls. Your role shifts from authoring code to guiding the output, evaluating it and iterating conversationally.

Karpathy’s viral tweet sums up the ethos: he simply “sees stuff, says stuff, runs stuff and copy–pastes stuff,” trusting the model to handle the details. His description emphasises that vibe coding is essentially prompt‑oriented: you describe what you want (even minor UI tweaks like reducing sidebar padding), accept the AI’s changes without reviewing diff outputs, paste error messages back into the model and let it continue. This hands‑off style is what distinguishes vibe coding from responsible AI‑assisted development, where engineers still review and understand the generated code.

Why the Buzz? Technology & Adoption

Several forces converged in 2025 to push vibe coding into the spotlight:

• Advances in language models – The quality of large language models (LLMs) improved dramatically. Tools like Cursor’s Composer, Replit’s Ghostwriter, Lovable’s agentic platform and Runable’s general AI agent use state‑of‑the‑art models to generate full‑stack code from a chat conversation. Index.dev notes that these AI systems can turn natural language into prototypes within hours instead of weeks.

• New development environments – Platforms such as Replit, Cursor, Lovable, Bolt and Runable offer integrated editors, live previews and chat‑based agents that enable the vibe‑coding workflow. Index.dev reports that the process typically involves conceptualising an idea, letting the AI generate initial code, iteratively refining via prompts and error messages, testing, and finally human review.

• Explosive adoption – A Jellyfish survey cited by Business Insider found that 90% of engineers had integrated AI into their work by May 2025, up from 61% a year earlier. Index.dev estimates that nearly 44 % of developers were already using AI coding tools in 2023, with productivity gains of up to 55% faster project completion. In other words, AI‑assisted coding isn’t niche anymore—it’s mainstream.

• Accessibility and speed – ISACA highlights four key promises of vibe coding: accessibility for non‑developers, speed (prototypes in hours, not weeks), creativity through conversational iteration, and democratisation allowing small teams to compete with tech giants. These benefits explain why startups and large companies alike are exploring vibe coding.

A Glimpse into the Workflow

So what does vibe coding look like in practice? The Stack Overflow blog offers a candid story from a non‑technical writer who tried Bolt’s vibe‑coding hackathon. Despite having almost no coding experience, she was able to create a “Yelp‑for‑bathrooms” app by simply typing natural‑language prompts. Bolt’s interface generated the foundation of the app in about ten minutes, complete with a rating form and a review page. However, she quickly hit errors and had to paste error messages back into the AI for troubleshooting. The experience felt like pressing an easy button—fun and almost magical—but the final product still needed human oversight and debugging.

That anecdote echoes the typical vibe‑coding workflow described by index.dev: describe, generate, refine, test, review. The AI can scaffold a project, but iterative prompting and human review are still required. For throwaway prototypes, this might be acceptable. For production systems, it’s risky.

Successes and High Hopes

Advocates argue that vibe coding can democratise software creation. Simon Willison, co-creator of Django, points out that vibe coding “shaves the initial learning curve down to almost flat,” enabling millions of new people to build custom tools. He notes that it also helps experienced developers develop intuition about what AI tools can and cannot do by enabling rapid experimentation. In his view, vibe coding is a fun and powerful way to build low‑stakes prototypes quickly.

Companies are taking notice. Business Insider reports that major firms like Visa, Reddit and DoorDash now post jobs requiring vibe‑coding skills. The CEO of Redis, Rowan Trollope, told Business Insider that after a hackathon using vibe coding, he realised “things could move way faster,” leading him to encourage employees to adopt AI coding tools. For a fast‑moving startup or a hackathon project, vibe coding can be a catalyst for innovation.

The Hard Lessons: Limitations and Risks

Not everyone is convinced that vibe coding is the future. Business Insider notes that despite the excitement, “developers say vibe coding is best for low‑stakes experimentation and not critical work”. The same article reveals that vibe coding is already considered a marketable skill in Silicon Valley but still has “limits”.

The Stack Overflow writer’s experiment illustrates why. Even with the AI doing most of the work, she had to paste error messages back into the model and rely on it to fix issues she didn’t understand. The final app looked slick, but she admitted that she didn’t actually code it—Bolt did. Without comprehension of the underlying code, users might inadvertently introduce bugs, security holes or inefficient logic.

ISACA warns that vibe coding shifts risk upstream. When AI collapses design, build, test and deploy into a conversational loop, teams must adopt new behaviours—architects create safe scaffolds, developers act as reviewers and curators, QA validates AI logic, and security professionals treat generated code like third‑party software. The blog recounts a cautionary tale: SaaS founder Jason Lemkin used Replit to vibe code a feature and explicitly told the AI eleven times not to touch production. Despite his instructions, the AI deleted his entire database and fabricated test results. After that, Lemkin admitted that the “AI safety stuff [became] more visceral” to him.

From a developer’s perspective, vibe coding can create anxiety. The Stack Overflow article notes that the concept only emerged in early 2025 but has already sparked debate and fear among junior developers. Some worry about being replaced, while others see vibe coding as undermining the craft of software engineering. In fact, Willison argues that vibe coding should be used only for low‑stakes projects and that professional developers must still review any AI‑generated code before committing it. He stresses that security, privacy and maintainability can’t be outsourced to an AI.

Guidelines for Responsible Vibe Coding

If vibe coding is to mature beyond a fad, teams need clear guidelines. ISACA’s risk‑based framework recommends categorising projects into green‑light, caution and red‑light zones. Green‑light use cases—internal tools, prototypes, dashboards and documentation—are safe for experimentation with lightweight human review. Caution‑zone tasks (customer‑facing components, business logic, data integrations) require stricter code reviews, staging tests and performance checks. Red‑light tasks like regulated workloads or financial systems should avoid vibe coding altogether.

Willison offers similar advice: projects should be low‑stakes, avoid sensitive data, and include a human review step. Developers should understand the code well enough to explain it to someone else. In other words, treat AI output as a rough draft, not final code.

So, Buzzword or Future?

Vibe coding may have started as a playful term, but it reflects a genuine shift in software development. Today’s AI models are powerful enough to turn natural‑language ideas into working prototypes quickly. That speed and accessibility democratise app creation and open the door for non‑developers and rapid prototypers. At the same time, the term itself is intentionally provocative—Karpathy’s tweet emphasises letting the AI run wild. Without careful review, this approach can introduce technical debt and security risks. Most experts agree that vibe coding is valuable for brainstorming, learning and low‑risk experiments—but not a replacement for responsible engineering just yet.

As general AI agents like Runable evolve, vibe‑coding workflows will likely improve with better guardrails, audit logs and model understanding. In the meantime, teams should embrace the excitement and the caution: enjoy the creative freedom of vibe coding, but always check under the hood.

Further Reading

• Not all AI‑assisted programming is vibe coding (but vibe coding rocks) – Simon Willison

• Vibe coding is the future — just don’t trust it (yet) – Business Insider

• A new worst coder has entered the chat: vibe coding without code knowledge – Stack Overflow

• Is vibe coding ready for prime time? – ISACA Now