The Vibe Is Dead. Here's What I Actually Do Now.

I watched that Karpathy video in early 2025 and thought, "This is it. I'm done writing boring code."

For three months, I vibed everything. APIs, dashboards, entire features. I felt invincible.

By July 2026? Totally different mindset. The magic wore off. And honestly, I'm glad.

My New Rule

I stopped treating AI like a magical code factory. I started treating it like a very fast, very confident, occasionally wrong intern.

I don't ask for code anymore. I ask for a plan.

First, I make the AI write out exactly what it's going to build—in plain English. No code. Just logic. Then I read it and ask: "Does this actually make sense?"

Most of the time, the plan is wrong. Not because the AI is dumb—but because I didn't give enough context. So we go back and forth until it feels right.

Only then do I let it generate code.

It sounds slower. But it's way faster than debugging nonsense for five days.

The Real Trend in 2026

Everyone talks about "prompt engineering" like it's a superpower. It's not.

The real skill? Writing better requirements.

Most stakeholders speak in feelings: "Make it faster." The AI guesses what that means, and we blame it when things go wrong.

Now I force clarity before I even open my IDE. Exactly what does "faster" mean? Two seconds? Half a second?

Once I get that answer, the AI delivers. It's boring how well it works.

What's Coming Next

I think we'll see a new role emerge in the next year or so. I'm calling it the "Vibe Manager" (hoping someone comes up with a better name).

For solo developers and indie hackers, this is already happening naturally. You're the product person, the developer, and the QA all in one. But with AI in the mix, there's a new hat to wear.

The Vibe Manager doesn't write much code anymore. Instead, they do three things:

1. Translate vague ideas into clear instructions. Your client says "make it modern." You figure out what that actually means in terms of layout, colors, and interactions.
2. Manage the AI's context. You remember what the AI has built before, what mistakes it made, and what patterns work. You keep a "playbook" of prompts that actually deliver results.
3. Know when to step in. You can tell when the AI is spiraling into complexity and when it's time to stop prompting and start making human decisions.

I've been doing this solo for about six months now. I take on freelance projects, vibe the core functionality, and spend the rest of my time being the human filter. The code is the easy part. The judgment? That's still all me.

And honestly? Clients pay more for that judgment than they ever paid for my coding speed.

My One Piece of Advice

Stop trying to replace yourself. Start trying to supercharge yourself.

I spent early 2025 terrified AI would make me obsolete. Now I laugh at that version of me. I was so focused on code that I forgot about everything else—problem-solving, people skills, judgment.

AI is terrible at all of that. It's good at syntax and speed. But it doesn't understand your users, your boss, your company politics, or what "good enough" means in your context.

That's your job. That's always been your job. The code was just the messy middle part.

And honestly? I'm finally enjoying my work again. Not because I do less—but because I get to do what actually matters.

I'm just some guy figuring this out as I go. The tech changes fast. The human stuff stays the same. Focus on that.