Vibe coding and Claude Code: hype, or how you actually build software?

Vibe coding became the word of 2025 and has only grown bigger in 2026. And the idea is seductive: you type what you want in plain language, the AI writes the code, and you don’t need to understand any of it. I build software with AI every day myself, from our own CRM to this website, so I get the excitement. But there’s a wide gap between a fun weekend experiment and software your business runs on. In this article I explain what vibe coding actually is, where it does and doesn’t work, and what makes Claude Code, the serious tool behind it, different. No hype, just an honest framework for anyone considering building software with AI.

What is vibe coding?

Vibe coding is letting an AI write software based on what you describe in plain language, without reading or understanding the code yourself. You steer on the result and on your gut feeling, not on the lines of code underneath. The term is deliberately playful: you go “with the vibe” and trust the AI to figure it out.

The expression comes from Andrej Karpathy, co-founder of OpenAI, who launched it in a tweet on 2 February 2025:

“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, co-founder of OpenAI (via Wikipedia)

An important detail that often gets lost: Karpathy immediately added that it’s “not too bad for throwaway weekend projects”. The inventor himself positioned it as something for disposable side projects, not production. That nuance is the thread running through this article.

Why is vibe coding suddenly everywhere?

Because AI models crossed a threshold in 2025: good enough to conjure a working app out of a single sentence. That suddenly made software building accessible to people without programming knowledge, and that explains the explosion. In Flanders, search volume for “vibe coding” grew roughly sixfold in eight months.

The term broke through so hard that Collins Dictionary named it word of the year 2025 on 6 November 2025, with a striking description: “programming by vibes, not variables”. Meanwhile the tools popped up like mushrooms: Bolt.new, Lovable, Replit and v0 all promise “an app without code in minutes”. That promise is real, but it only tells half the story.

Can you build real software with vibe coding, or only prototypes?

For prototypes and internal experiments: absolutely. For production software that customers and revenue depend on: not just like that. Vibe coding is fantastic for quickly testing an idea, making a small internal tool or showing an MVP. But as soon as something needs to be reliable, secure and maintainable, “forgetting that the code exists” becomes a problem instead of a feature.

The sharpest distinction comes from Simon Willison, a respected developer and co-creator of the Django framework:

“If an LLM wrote the code for you, and you then reviewed it, tested it thoroughly and made sure you could explain how it works to someone else — that’s not vibe coding, it’s software development.”

— Simon Willison, co-creator of Django (simonwillison.net)

In other words: the moment you do read, test and understand the code, you’re simply doing software development with AI as an accelerator. Willison keeps a golden rule: he doesn’t commit a single line of code he couldn’t explain to someone else. Exactly that difference, review and responsibility, separates a fun experiment from something a business can build on. That’s why ClickForest uses AI as an accelerator for Flemish SMEs in Mechelen and Antwerp, not as an excuse to stop understanding the code.

What are the risks of vibe coding?

Code nobody understands is insecure and hard to maintain. If nobody in your business knows what’s actually running, security flaws creep in, bug hunting becomes a nightmare and technical debt piles up unnoticed until it catches up with you. That’s not doom-mongering, it’s measurable.

Security firm Veracode tested code from over a hundred AI models across eighty tasks in its GenAI Code Security Report. The outcome: in 45% of cases the AI-generated code introduced a security flaw from the OWASP Top 10, and the Veracode spring 2026 update confirms the number isn’t improving. Note the framing: this covers AI-generated code in general, not specifically vibe coding, but the risk only grows when nobody reviews the code. Veracode CTO Jens Wessling puts his finger on it:

“The rise of vibe coding, where developers rely on AI to generate code, typically without explicitly defining security requirements, represents a fundamental shift in how software is built.”

— Jens Wessling, CTO at Veracode (Veracode)

Speed also has a price you only see later. Security firm Apiiro analysed the code of Fortune 50 companies and saw that AI-assisted developers produce three to four times more commits, but introduce ten times more security findings (Apiiro). Those figures cover AI-assisted development in general, not vibe coding specifically, but the direction is clear: more code, faster, is not the same as better software. Developers themselves remain critical too. The Stack Overflow Developer Survey 2025 of some 49,000 respondents shows that 84% use AI tools or plan to, but trust is falling: 46% distrust the accuracy of the output, up from 31% a year earlier. The biggest frustration, cited by 66%, is “AI solutions that are almost right, but not quite”. And “almost right” is often more dangerous in production software than clearly wrong, because it slips through more easily.

What comes after the vibe coding hype? Agentic engineering

In 2026 the conversation has already moved on from vibe coding to “agentic engineering”. The core of that shift: the AI executes, but the human provides the architecture, the review and the responsibility. Even Andrej Karpathy, who coined the term vibe coding, now positions it as the baseline, not the endpoint.

“Vibe coding is about raising the floor for everyone in terms of what they can do in software. Agentic engineering is about extrapolating the ceiling.”

— Andrej Karpathy, co-founder of OpenAI (Sequoia AI Ascent 2026)

Raising the floor means building becomes accessible to everyone; extrapolating the ceiling means turning it into professional software, with the human responsible for quality and security. Simon Willison coined his own term for this too, vibe engineering: deliberately disciplining AI assistance with tests, review and understanding. Exactly that mature way of working is what a tool like Claude Code is made for, and why ClickForest uses it to build production software for Flemish SMEs in Mechelen and Antwerp instead of throwing away prototypes.

The difference between the two ways of working at a glance:

What is Claude Code, and why is it different from vibe coding?

Claude Code is Anthropic’s agentic coding tool that searches your codebase from your terminal, reads the relevant files, makes changes across multiple files, runs tests and commits the code. It’s not a chat window suggesting loose snippets, but a system that makes a plan, executes it and corrects itself on failures: agentic at the project level instead of autocomplete line by line. Anthropic itself positions it as “an agentic coding tool that lives in your terminal”, not smart autocomplete.

That’s a fundamentally different way of working from the vibe-coding tools that turn one prompt into one app. Claude Code launched in February 2025 and has been generally available since spring 2025. That it resonates with professionals shows in the Stack Overflow Developer Survey, where 9.7% of developers use Claude Code, alongside 17.9% for the comparable Cursor. It asks for your approval by default before applying changes, so at every step you can see the code, steer it and decide what gets committed. That makes it easier to follow Willison’s golden rule, code you understand and can explain, but it doesn’t guarantee the discipline: you can switch that approval off, and conversely you can work just as carefully with tools like Cursor or Codex. The tool facilitates the discipline; applying it is still on you. Vibe coding is the concept and the hype; Claude Code is one of the tools that lets you do it in a mature, responsible way. Where the differences between AI tools per task lie is covered in AI tools per task; how Claude compares to ChatGPT and Perplexity in practice is in the practical guide to ChatGPT, Claude and Perplexity, and how Claude Code fits as one of three modes in the wider Claude landscape in Claude Desktop explained.

What do you actually build with it in practice?

Full production software, provided you know what you’re doing. ClickForest builds with Claude Code not only this website, but also its own CRM, quoting module, client portal, site chatbot and its own MCP servers that connect AI to our data and tools. That’s no throwaway weekend project, it runs in production every day for Flemish SMEs in Mechelen and Antwerp.

The difference with pure vibe coding is the approach: every change is read, tested and understood, with the same discipline as classic development, but much faster. That’s how we built clickforest.com on the Astro framework, with Claude Code compressing the build time considerably without sacrificing quality; the full story is in Astro or WordPress for your SME website. For businesses this means custom software, which used to require an expensive development team, is now within reach through the building custom software with AI service or a full website development project. If you’d rather understand the approach before investing in it, our AI consultancy for SMEs will get you started. (About those MCP servers, the “USB-C for AI” that connects Claude to your own data, a separate post is coming soon.)

Do you need programming knowledge, and who is this for?

Not for a prototype; for production you do, or someone who safeguards it on your behalf. That’s the core for an SME owner: vibe coding lowers the barrier to testing an idea to almost zero, and that’s pure gain. But as soon as it needs to become something that handles customers, payments or sensitive data, you need someone who reads, tests and understands the code.

That doesn’t have to be you. It can be a technical partner who uses AI as an accelerator and carries the responsibility for quality and security. That’s how ClickForest uses AI to build faster for Flemish SMEs in Mechelen and Antwerp, but with the discipline of real development underneath. If you’re considering AI for more than software, for example AI agents or broader automation, the AI for growth hub gives you the overview, and you can dig deeper into the Claude ecosystem via Claude Skills.

Conclusion: vibe coding is a great tool in the right hands

Here’s how I see it, after a year of building with it daily:

1. For prototypes and experiments: do it. Vibe coding is the fastest way ever to test an idea. Enjoy the vibe.
2. For production software: only with review. Code nobody understands is a time bomb of bugs and vulnerabilities. “Almost right” isn’t good enough.
3. Know the difference between the hype and the tool. Vibe coding is the concept; Claude Code is one of the tools to do it responsibly, agentic instead of guesswork.
4. Responsibility stays with a human. AI writes, a human understands, tests and stands behind it.

The biggest thinking error I see is mixing up vibe coding and professional AI development. It’s not “AI builds everything by itself”, it’s “AI builds faster, provided someone steers it”. Not coincidentally, the whole industry is shifting from vibe coding to agentic engineering in 2026, exactly that approach. ClickForest builds custom software with Claude Code that way for Flemish SMEs in Mechelen, Antwerp and across Flanders: the speed of AI, with the craftsmanship underneath. Curious whether your idea is feasible? Book a video call or get in touch and we’ll look at it together.