Last Tuesday at 1 AM, I was debugging a critical production issue in my AI dev tool. As I dug through layers of functions, I suddenly realized — unlike the new generation of developers, I was grateful that I could actually understand my codebase. That’s when I started thinking more about Karpathy’s recent statements on vibe coding.

For those who missed it, Andrej Karpathy recently shared his thoughts on what he calls “vibe coding” — essentially surrendering code comprehension to AI tools and hoping for the best. His exact words? “I ‘Accept All’ always, I don’t read the diffs anymore.”

I have learnt a lot from Karpathy and use AI tools daily, but there’s a world of difference between augmenting your capabilities and completely surrendering your understanding.

Two days ago, my friend Owen messaged me in a panic. He had built an impressive SaaS app using Bolt, but realized that his OpenAI API key was completely exposed. He was fortunate to have caught it early, but what if this had actually went into production?

Owen isn’t alone. Unfortunately, AI coding assistants often generate functional but insecure code unless explicitly prompted about security concerns.

After walking Owen through securing his application, I realized these lessons could help others. So I compiled this comprehensive security checklist for vibe coders — the same advice that saved Owen’s project.

I also wrote up a “vibe security prompt” at the bottom of the article. Give that to your AI of choice and secure your application!

As a software engineer experimenting with AI for the past 2 years, I’ve tested nearly every AI coding assistant on the market and developed a workflow that consistently delivers results.

Here’s my tried-and-tested method for solo developers looking to leverage AI to make SaaS products.

Last week, X exploded when a “vibe coder” announced his SaaS was under attack.

His business, built entirely with AI assistance and “zero hand-written code,” was experiencing bypassed subscriptions, maxed-out API keys, and database corruption.

His follow-up admission made this notable: “as you know, I’m not technical so this is taking me longer than usual to figure out.”

As someone deeply immersed in the AI code generation space, I’ve been watching this unfold with a mix of sympathy and frustration. Let me be clear — I’m not against AI-assisted development. My own tool aims to improve code generation quality. But there’s a growing and dangerous fantasy that technical knowledge is optional in the new AI-powered world.

After observing many similar (though less public) security disasters, I’ve come to a controversial conclusion: vibe coding isn’t just inefficient — it’s potentially catastrophic.

In 12 years of my dev career, I’ve spent countless hours battling browser bugs.

Recently, I discovered an MCP that’s cut my debugging time in half.

MCP as a term is being overused too much, but just understand them as APIs that AI agents can use.

I found an MCP to let AI see and interact with your browser, called BrowserTools. Once you integrate it with Cursor, you can ask it see what’s going on in your browser and the console.

Using AI for coding isn’t perfect, but it definitely makes me faster.

I’ve been coding with Cursor AI since it was launched now while building my SaaS, and I’ve got some thoughts.

The internet seems split between “AI coding is a miracle” and “AI coding is garbage.” Honestly, it’s somewhere in between.

Some days Cursor helps me complete tasks in record times. Other days I waste hours fighting its suggestions.

After learning from my mistakes, I wanted to share what actually works for me as a solo developer.

I’m not sure if I’m ready for what’s coming next with AI. Things are moving too fast.

The recent releases of Claude 3.7 and GPT-4.5 were… surprising. I read articles where AI researchers admitted these systems solved problems faster than they could.

We’re racing toward a world where humans might not be the smartest ones anymore. And we’re doing it willingly, even eagerly.

A year ago, I laughed off AGI fears. Then, I watched these new models solve in seconds what took me hours. I suddenly felt… outdated.