What vibe coding can (and can’t) do for software engineering

The pros and cons of AI-powered dev: From prototype magic to production risk.

I’ve been coding for 30 years. But this year, I watched a founder ship a SaaS product, get paying users, and get hacked, all within 72 hours. Welcome to the world of vibe coding.

In March 2025, everyone on Twitter, YouTube, and Discord was talking about vibe coding. Ex-OpenAI researcher Andrej Karpathy coined it:

“English is the hottest new programming language” — Andrej Karpathy

That quote kicked off more than a trend. It sparked a new identity in software development, one that lets you build apps without writing code line by line.

Advantage or liability: What is vibe coding?

Vibe coding is the practice of describing what you want in plain English and letting AI generate the code.

It’s fast, fun, and full of false confidence. You speak or type prompts, and AI tools handle the heavy lifting. Want to create a SaaS app? Prompt it.

This means trusting AI so completely, that we’re encouraged to stop coding manually or even reviewing code carefully. 

In theory, anyone with an idea can describe their vision to an AI model and see it come alive in code. The philosophy is supposed to be democratic and experimental, lowering barriers for non-engineers, helping them create software.

Vibe coding also encourages rapid prototyping by experienced developers who want to push the limits of what AI can do. It aligns with the “no-code/low-code” movement of empowering “citizen developers” the only difference is that AI helps the creator write custom code for any requirement rather than using pre-built templates or modules.

It’s also the most interesting development in programming since open source. But we’re reaching a point where the hype may outpace the safety.

Read more: 6 limitations of AI code assistants

What vibe coding gets right

The best thing about vibe coding is that it lowers the barrier to building. You’ve got an idea. If you want to see if it works, you open up an AI tool, speak your requirements, and in minutes, you’ve got something functional. 

It’s perfect for 

• Minimum viable products/prototypes (MVPs) and side projects
• Early demos for investors and internal teams
• Product ideation

This is brilliant for product managers, creatives, and junior devs exploring new spaces. Journalists like Kevin Roose have developed AI-powered apps without a coding background. It’s fast. It’s expressive and it lets you build wide. The barrier to entry has never been lower.

Just describe what you want, “A website that emails daily quotes” or “A form that tracks student attendance” and boom, you’ve got something tangible enough to test. It feels like magic, but like most magic tricks, it works best in short bursts until you need it to perform under pressure.

Andrej said it best: “It’s not coding; I just see things, say things, run things, and copy and paste things,” and it mainly works.

That’s the key: mostly.

I’m not against the concept. I just think we need to be more realistic about where the vibes should stop.

Read more: Your 2025 guide to AI, no-code, and developer-led software

Where the vibes should stop (The 60% rule)

Vibe coding ends at 60%. Beyond that line, you’re no longer prototyping, you’re building. The following 40% belong to real engineers. It’s where things like authentication, payments, security, and scale reside, and hard-won judgment that AI simply doesn’t possess.

It’s also where your product stops being a prototype and becomes a responsibility.

The moment you start preparing to scale, by serving customers or handling sensitive data, you’ve crossed into the other 40% of the software lifecycle, the part where vibe coding stops being an advantage and starts becoming a liability.

You can’t AI your way through secure authentication or payment logic without consequences. I’ve seen devs vibe code their way into success and then vibe code their way into a data leak. The founder mentioned earlier? His API keys were exposed in plain text.

That’s not a startup story, it’s a cautionary tale.

If vibe coding gets you 60% of the way there, stop. Beyond that, it gets dangerous.

Read more: What is prompt engineering?

What Startups and CTOs should know

Companies need to be smart about where vibe coding is allowed.

“It’s great for web toys. Not great for systems with stakes.” — John Campbell, Director of Engineering, Security Journey.

The false narrative about AI replacing devs

It has become fashionable to say that AI is replacing developers. AI didn’t cause the recent wave of engineering layoffs. It’s a correction from the pandemic hiring spree. Blaming AI is a great way to mitigate the PR blow of layoffs. Meanwhile, companies like Meta continue to hire engineers and offer 30-40% raises to top performers.

What junior engineers need to know

I hear a lot of panic about AI replacing junior devs. Tools like ChatGPT are getting better by the day. But here’s what you need to know: AI can generate, but it can’t reason yet.

Your job is to learn the craft before you try to optimize it. Don’t treat vibe coding as a shortcut to avoid understanding the fundamentals. Use it as a companion.

If you’re coding because it pays well, it’s going to get harder. If you can’t sleep until you fix that annoying bug, don’t worry. You’re fine.

Read more: Why AI won’t replace developers

Our stack needs real thinking

At Knolli, we build copilots. We help creators and companies build domain-specific AIs that talk to users, answer questions, and sometimes even write code.

Do we believe in vibe coding? Yes, but only for prototyping. No copilot, no matter how advanced, should ship production apps unsupervised. A lot of emerging vibe coding stacks feel like IKEA furniture kits: A bunch of parts, promising quick assembly. But software isn’t packed. Integrations are complex. Security is non-negotiable. 

If you’re using vibe coding to create a mockup, do it. If you’re trying to replace your engineering team with it, you’re asking for trouble.

Vibe coding isn’t stupid, but it’s not enough

The value of vibe coding is undeniable. It has made software creation more inclusive, democratized the creation of prototypes, helped freelancers build $ 750-a-day businesses, and let non-technical professionals build tools for themselves.

But it’s not production-grade engineering.

When something breaks in a vibe-coded app, the same non-technical creator will need to call someone who understands the code. The AI won’t explain why your subscription logic failed or why a webhook misfired during checkout. Use AI to get started and explore, but don’t mistake a prototype for a product. The hottest new programming language might be English, but the best engineers still know when to go back to code.

More from We Love Open Source

• What is prompt engineering?
• Why AI won’t replace developers
• 6 limitations of AI code assistants
• Your 2025 guide to AI, no-code, and developer-led software
• How I use AI agents to automate my workflow and save hours