I run a personal AI assistant that manages my calendar, emails, projects, and reminders. Today it broke in five different ways. That's where the real story is.

The Silent Reminders

Three morning reminders. None delivered. The cron system saw an empty config file and skipped everything — including jobs unrelated to that file. A known bug. Fix: one line of text in an empty file.

The OAuth Trap

Same provider, same service, two key formats: oat01 (OAuth) vs api03 (standard). They look identical. Completely incompatible. Error message? "Invalid API key." No hint about why.

The Cookie Wall

Spotify CLI needed Chrome cookies. My assistant runs remotely on my Mac — but macOS won't show a Keychain prompt over SSH. I had to physically walk to my desk. Some things still need a human at the keyboard.

The Wrong URL

I said "Redmine is at soporte.trabitat.com." It wasn't. Instead of stopping, my assistant checked its memory, found an old URL, and brute-forced variations until plani.trabitat.com responded. Users give wrong info. Good tools keep looking.

The Missing Binary

Needed gh CLI. No sudo, no package manager. Downloaded the binary from GitHub, dropped it in ~/.local/bin. Ten seconds, zero permissions.

What Actually Worked

Between the breaks: analyzed a 9-page contract and flagged I was being underpaid, transcribed voice messages with timezone math, categorized 88 Redmine tickets, ran an Odoo report on schedule. Working-to-broken ratio: 10:1. But breaks are louder.

The Takeaway

AI assistants aren't magic. They're software. They break in boring ways: wrong URLs, bad auth, empty files. The difference is a good one debugs itself, finds workarounds, and tells you what happened.

That's not artificial intelligence. That's artificial persistence.

Part of my daily series on building with AI. What actually happens — not what demos promise.