Vibe Coding, Explained — And Why It's Not Just Vibes

Karpathy coined the term in February 2025 as a half-joke and the internet, predictably, turned it into both a movement and a culture war. So let's settle it.

What it actually means

Vibe coding is when you describe what you want in plain English (or Spanish, or any human language), an AI writes the code, and you ship it without reviewing every line. You judge the result by whether the thing feels right — does it work, does it look ok, did the demo land — instead of auditing the implementation.

That's it. The novelty isn't that AI writes code. It's that you trust the output enough to not even read it on the way out.

Where it works brilliantly

• Prototypes & demos. A weekend hack to show your cofounder that the idea is alive. Ship in 6 hours, throw it away on Monday.
• Internal tooling. A script your ops team needs today. If it breaks, your ops team sends a Slack and you fix it.
• Throwaway one-shots. A landing page for an event in 2 weeks. A scraper for a one-off report. A playground to test an API.

For all of those, vibe coding is genuinely magical. The opportunity cost of notusing it is now embarrassing.

Where it ships a Chernobyl in disguise

Production software that handles money, identity, regulation, scale, or anyone's private data is notvibe coding territory. Not because AI can't do it — it can, sometimes shockingly well — but because the failure modes you don't see will eat you alive in month three.

Specifically:

• Subtle security holes. The auth bypass that only triggers when two specific roles intersect. The AI didn't test for it. You didn't test for it. Your customer's pen tester will.
• Schema drift you didn't notice. Today the LLM generated a column named user_id. Tomorrow another agent calls it userId. In six months you have three half-broken integrations.
• Performance cliffs. The query is fine at 1,000 rows. It's a 30-second hang at 100,000. You don't know which until your biggest customer signs.

The honest middle path

The right framing isn't "vibes vs. real engineering." It's AI velocity with senior judgment. The AI types. A human reviews the stuff that matters: architecture, security boundaries, data model. The combination ships in weeks instead of months and still passes the audit.

We do this for a living. Boring confession: most of the time we agree with the AI's first draft. We just don't skip the read.

Should you vibe code your startup?

If your startup is the demo, sure. Vibe everything until something works.

If your startup's value is not breaking once paying customers show up, you want vibes andtaste. That's not a contradiction — it's the actual job.

The vibe is real. The code still needs to compile in production at 3 AM.