From Wow to Oops: The Vibe Coding Reality Check

This content originally appeared on DEV Community and was authored by Arnaud Denoyelle

AI is having its “gold rush” moment.

Every week, there’s a new tool promising you can “build an app in minutes” — just type a prompt and watch the magic happen.

Non-technical folks see it and think: “Wow, developers must be 100x faster now.”
But here’s the thing: the magic trick only works from the right angle.

In this article, I’ll unpack why the reality is messier and how to use AI productively without falling for the hype.

Note : What I call vibe-coding is creating an application almost without writing any code, relying solely on prompts. This is in contrast to LLM-assisted coding, where the developer creates the project and occasionally uses prompts to complete it.

The illusion of application generators

I fully agree with the description given here:

Your manager reads a few LinkedIn posts about AI, tries an application generator, and suddenly expects you to work 100 times faster.

But that’s not how it works.

A developer often spends more time reading code than writing it

especially when the code was generated by AI.

So, no—you won’t code 100 times faster with AI. At least, not today.

Many have tried this: go to an application generator (Replit, Lovable, Bolt, V0), enter a prompt, and receive a “working” app.

It feels like “AI built it”. In reality, it’s closer to “AI finished it”.

Most generated apps are 80% boilerplate, 20% actual features — and that boilerplate is almost always pre-written by humans

Think of the boilerplate as the foundation of a house—the technical groundwork before construction begins. In software, that means setting up tools, a database, a server, basic page structures, etc.

When developers code manually, this boilerplate is what takes the most time at the beginning. Generators just clone a boilerplate and sprinkle AI-generated features on top.

That’s why the first prompt often delivers impressive results, but subsequent prompts fail to produce the same value.

Right now, you can’t realistically expect AI tools to generate large, production-ready applications from scratch.

They’re still best suited for:

• Ideation
• Design team support
• Small personal tools

Typically, the design team can test prompts with style constraints, color palettes, fonts, etc., and see the results live while generating multiple mockup proposals.

The Hidden Complexity

You ask the AI to make a login/password form.

It gives you a form with two fields and a button. You think that’s it? That’s the “happy path,” the nominal case.

But alongside that, you also have to handle all the other cases:

• Data validation (show the field in red if it’s empty, etc.)
• Extra features (auto-fill, global styling, etc.)
• Is it responsive?
• Handle error messages
• If the password is wrong, allow another attempt (or not)
• If the address doesn’t exist, suggest signing up
• If an unexpected error occurs? “Please try again later”
• Handle abuse

What about someone trying to brute-force a password? Just for that, you have to go through a whole process of setting up a captcha, with Google API calls, etc.

And what about someone attempting an SQL injection?

Extra features:

• What if I forgot my password? You need a separate form, with an email being sent. That’s another whole process in itself.
• And what if I want to log in with a Google account instead? That’s another separate process to set up OAuth.

You need to think about handling all these cases (and actually know they exist in the first place), and it takes time to deal with them, even with AI — if only to simulate/test that you’re handling them correctly.

The Real Story Behind Productivity Curves

• Junior – Starts quickly but maintains a slow overall pace.
• Senior – Invests time upfront to build a solid boilerplate, then accelerates and eventually outpaces everyone.
• Vibe coding – Bursts out of the gate with impressive speed, but soon crashes into the vibe coding wall — where lack of structure and growing complexity drain momentum.

There’s a zone where the developer has to explain to their manager that, no, vibe coding won’t actually make us faster. This is tricky, because from the manager’s perspective, things currently look very different. Navigating this moment requires clear communication and strong teaching skills.

The Dangers of Vibe Coding

Sometimes, managers just want to ship code quickly, regardless of quality.

There are two major risks with AI-generated code: security and skyrocketing costs.

Application generators often allow users to write code without understanding it—creating high security risks. I expect many generated apps will leak user data simply because their creator doesn’t truly understand how they work.

If you can’t explain your code, you can’t secure it.

There’s also the financial risk: AI can produce inefficient code that consumes vastly more resources than necessary — sometimes even getting stuck in infinite loops — leading to bills in the hundreds of thousands.

Rule of thumb: never run code you don’t fully understand.

The AI Moore’s Law

Source : https://theaidigest.org/time-horizons

There’s a so-called AI Moore’s Law, suggesting that AI’s “time horizon” doubles every 7 months:

• Initially, AI could do in 30 seconds what a coder could do—success rate 50%.
• Seven months later, it could handle 1-minute tasks.
• Today, it can handle 1-hour tasks (still with a 50% success rate).

But remember—a typical software project can represent decades of development (5–10 developers over 2–3 years).

If the trend continues, AI could match one month of coding by 2030. Even then, that won’t be enough for fully automated software creation for non-developers.

For developers, though, this is still a huge advantage—AI will be able to take on increasingly large chunks of work.

LLM-Assisted Development

// Detect dark theme var iframe = document.getElementById('tweet-1953531950137815374-797'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=1953531950137815374&theme=dark" }

We don’t need to wait for fully autonomous AI.
If AI can do an hour’s worth of coding with a 50% success rate, it can also:

• Do 30 minutes with ~70% success (70%² ≈ 50%)
• Do 15 minutes with ~84% success (84%⁴ ≈ 50%)

That’s already a massive time saver.

Developers are also learning to distinguish between tasks AI can and cannot handle.

Right now, AI is especially effective for:

• Spotting bugs
• Writing and updating documentation
• Creating automated tests
• Developing small features
• Drafting development plans

These use cases will only expand.

It’s also becoming easy to “train” AI to code in a specific style.

You can give it a file describing:

• How you build forms (with examples)
• How you handle database requests

And it will follow your style for repetitive tasks, leaving you to connect everything together.

How to Work with AI Today

AI offers both opportunities and risks. The key is understanding its current limits to avoid unrealistic expectations.

There won’t be an “AI big bang” but rather a steady, continuous adoption.

AI won’t replace developers tomorrow. But developers who know how to work with AI — and know its limits — will replace those who don’t.

So my advice is simple:

• Experiment daily
• Learn what AI can and can’t do
• Never ship code you don’t understand
• Regularly test whether past limitations have improved
• Invest in communication — Learn to clearly explain AI’s limits and trade-offs to non-technical stakeholders

Vibe-coding can be a great creative tool — just don’t confuse the demo for the delivery.

Have you tried vibe-coding yet? What worked for you, and what blew up in your face?

This content originally appeared on DEV Community and was authored by Arnaud Denoyelle