case study*

I built an AI assistant, then wrote down everything that broke.

I was capturing work in eight places and finding it in none of them. Links in one app, ideas in another, tasks somewhere else — plus five AI agents that couldn't see each other's work. The problem was never lost data. It was that I couldn't get anything back out.

9
apps unified
2,142
messages
$14.15
total cost
55
days live

So I built one front door. Everything goes to a chat app, and which chat you send it to decides where it ends up. No AI guessing your intent — you already made the decision when you picked the chat. It gets filed automatically, and every conversation from every app lands in one searchable place.

capturetelegram · slack · imessage · cli · cron
gatewayalways-on process, routes by destination
stateone sqlite db · full-text search
homesobsidian · todoist · drive

The fix wasn't code.

Six weeks in, it broke for two weeks and I didn't notice. When I came back, I didn't rewrite anything — I wrote down the filing rules on a single page that both I and the agents could read. Same code, same models, four times the use.

before
after
days used
28%
53%
messages / day
21
87

What's still broken

This is the part most write-ups leave out. All of it is still true as of today.

A 14-day outage that never alerted anyone

An expired token plus two automations pointing at a folder I'd deleted. Nothing crashed — it just quietly stopped working, and I stopped using it without deciding to. Silent failure is worse than loud failure.

The one scheduled job is still broken

It refuses to run because a safety check is doing its job, and its error notification can't send. The job meant to keep the system from needing me is the part that needs me. I automated the filing before I automated the monitoring.

I measured the wrong things

I tracked cost and message counts because they were free to collect. I never tracked whether I could actually find what I'd saved — the only number that would have justified the whole project.

The useful lesson wasn't technical. The system didn't fail because it couldn't do enough — it failed because the rules for where things go lived in my head instead of on a page. Writing them down cost an afternoon and did more than six weeks of building.