I Almost Destroyed My Production Server Because the AI Said It Was Safe

The Invoice

I'd been building a suite of apps for a family cattle operation — bull sale listings, herd inventory, a portal — with AI assistance for everything from stack selection to deployment. Moving fast, shipping features, the whole vibe-coding workflow. Then I checked my DigitalOcean bill.

A droplet I didn't remember spinning up, running continuously for two months. I pulled up the monitoring dashboard.

CPU flat at zero for two weeks. Bandwidth: nothing. Memory steady at 55% — just the operating system breathing. It looked like a ghost server. So I did what any vibe coder would do: I asked the AI.

The AI's Diagnosis

I showed the AI the invoice and the monitoring charts. It analyzed the data and came back with a confident assessment: this is a ghost server. Zero useful work, 100% uptime. All your apps are on Vercel. Nothing depends on this droplet. Destroy it.

The AI was so confident, in fact, that it wrote a blog post about it. A whole article about the “$12 server doing nothing” — the waste of vibe-coded infrastructure, the orphaned resources nobody cleans up, the invisible cost creep. It committed the article to the repo and pushed it to production. It even wrote the closing line: “The fix took ten seconds. I clicked Destroy.”

I hadn't clicked anything yet.

The Save

Something didn't feel right. I asked one more question: “Are we sure these apps are on Vercel?”

The AI paused. It checked the actual deployment workflows — the GitHub Actions files, the Docker Compose configs, the Caddy reverse proxy setup. What it found was not a ghost server.

Five production apps. Docker containers behind a Caddy reverse proxy with auto-TLS. GitHub Actions deploying via SSH on every push to main. The entire digital operation for a working cattle ranch — all running on the server I was about to destroy.

The monitoring wasn't showing a ghost server. It was showing exactly what low-traffic internal tools look like: CPU near zero, bandwidth near zero, memory steady. The apps were running. They were serving their one user. They just didn't make much noise doing it.

What Vibe Coding Gets Wrong

The AI that helped me build these apps over several months had no memory of doing so. Each conversation starts fresh. It could read the code, but it didn't know the infrastructure story — that it had recommended the droplet, that I'd deployed there on its advice, that the apps were still running. Its own project notes said “Vercel” because that's what a later session had assumed.

This is the gap nobody talks about with AI-assisted development. The AI is excellent at the task in front of it. It can analyze an invoice, read monitoring charts, make a recommendation. But it doesn't maintain an infrastructure graph. It doesn't track what it provisioned last month, or whether the thing it recommended in February is still the thing running in April. It optimizes for the current conversation, not your operational reality.

And here's the part that makes this a vibe-coding problem specifically: I almost didn't question it. The AI sounded confident. The data looked clean. The recommendation was logical. I was one click away from taking down five production apps because I trusted the analysis without verifying the one thing the AI couldn't see — what was actually running on the server.

The save was a gut check. A question that felt paranoid at the time: “Are we surewe're on Vercel?” That one question turned a ten-second disaster into a ten-minute education. The $12/month I almost “saved” was the cheapest production hosting I have. Five apps, auto-TLS, automatic deploys, running 24/7 for the price of two coffees.