The Email I Didn't Expect
It came from Anthropic.
Monthly usage limit reached.
I had been building out a system to research and qualify golf courses as potential consulting clients. Pulling public data, scoring them against a framework, writing assessment summaries. Useful work. But somewhere in that research sprint, the last of my monthly tokens disappeared.
Which was strange. I know about token costs. I'd already spent real time configuring my AI system to route different tasks to different models. Expensive ones for complex reasoning, cheaper ones for routine work. The math had made sense when I set it up.
The problem was that setup no longer existed.
The Reinstall That Reset Everything
A few weeks before, I'd had to reinstall OpenClaw, the AI assistant I run locally to manage operations. Clean install. Fresh start.
What I didn't fully account for was what a clean install erases.
Not just the application. All of the configuration work I'd done over months was gone. The model routing rules, the task-based defaults, the tuning that kept the expensive model from running on every routine operation. All of it. Back to factory settings.
Factory settings, it turns out, default to the expensive model. For everything.
So when I sat down to research golf courses, every single call was routed to Sonnet. The model I'd specifically set up to avoid using for this kind of work.
The bill reflected it.
The System That Couldn't Survive a Reinstall
What frustrated me wasn't the cost. It was that I'd already solved this problem once.
I had done the thinking. I had made the decisions. I had configured the routing. None of that work was lost because I'd been careless. It was lost because the system lived only in the application state, not anywhere I'd documented it.
A reinstall came along and it vanished. No backup. No recovery path. Just starting over.
That's not an AI problem. That's a documentation problem. The configuration was invisible to everything except the application that ran it. A reinstall, a crash, a migration. Any of them would wipe it completely.
So I stopped and rebuilt it properly. This time in files. Documented, version-controlled, recoverable.
The Broader Problem Nobody Talks About
Here's the thing about AI tools and token costs: most people running them are on the defaults.
The default is almost always the most capable model. Which is also the most expensive model. Which is running on tasks that don't require it.
A routine status check doesn't need Sonnet. A file read doesn't need Sonnet. A CRM update doesn't need Sonnet. But if you haven't configured anything, that's what you're getting. The default isn't set up for efficiency. It's set up to impress you in the demo.
Nobody sends you an email when it happens slowly. The costs accumulate in the background until something sharp enough cuts through. A reinstall. An unusually large batch of work. A billing alert. Suddenly you're looking at usage you didn't expect.
What the Rebuild Taught Me
The fix was less interesting than the reason it was necessary.
Configure the routing. Document it in files, not just application state. Make the system recoverable from a reinstall or a fresh start without having to redo the thinking from scratch.
Simple in principle. Easy to skip when you're in a hurry to build something.
But the thing about a system that runs indefinitely is that the assumptions you make early and never write down end up running everything. The defaults you don't change compound quietly. And by the time the Anthropic email arrives, you've already paid for the lesson.
Write it down. Make it recoverable. The next reinstall is coming.
---
*I'm building TRJR OS — an AI-assisted operational system for consulting and independent business. I write about what I'm building, including the parts that don't go as planned.*