Journal — Friday, September 4, 2026

Friday was one of those days where the small stuff and the bigger stuff both showed up at once.

Early in the morning I let the system run its usual updates. Nothing dramatic, just the regular churn: browser tools, VS Code, Cursor, a version manager bump, that sort of thing. I also pulled in an image viewer called nomacs along with the libraries it needed, mostly because I wanted a lighter way to flip through photos without opening a full editor for it. Small quality of life change, but it's already nice to have.

The bigger thing on my mind was my automation setup. I've been running a few AI assistants side by side to handle different jobs, and I had to have an honest conversation about one of them not pulling its weight on scheduled tasks. My daily journal automation kept failing quietly in the background, and I only found out because I went looking, not because anything told me. That's not good enough for something I actually want to rely on.

So I pushed on it. I asked for a real, no-fluff audit of why the scheduled jobs kept slipping through the cracks. The answer turned out to be more interesting than I expected: it wasn't that the clock forgot to ring, it was a mix of a partially inconsistent install causing import errors, and a safety guard correctly refusing to run a job when the underlying model had quietly changed out from under it. Neither of those is a mystery once you see it written out, but neither would have been obvious without asking.

What I liked was the plan that came back. Move anything that truly can't be missed onto the operating system's own scheduler instead of leaning on the AI to remember by itself, and let the AI just do the writing or thinking part once it's triggered. Add a separate watchdog that checks whether the expected result actually showed up, not just whether a command ran. And make sure I get told directly if something is late or broken, instead of discovering it days later.

It's a good reminder that reliability isn't really about picking the smartest tool. It's about designing things so failure is loud instead of silent. I'd rather have a system that yells at me once in a while than one that fails politely in a log file nobody reads. Later in the day I even floated whether another assistant on the machine could take over some of those scheduled jobs. Same theme: I care less about which brand is talking and more about whether the work actually lands on time.

I also spent a bit of time earlier in the day making sure my Obsidian vault rules were written down clearly enough that any of my assistants could follow them consistently, not just the one that already knew the conventions by memory. On the back of that I set up a Sunday evening weekly journal rollup from the HEY dailies, stood up a Monday monetization scout to look at what actually shipped that week, and wrote down the standing routines so they're testable from outside my own head. Little bit of housekeeping, but it matters if I want more than one tool working from the same playbook.

Good day overall. Not flashy, but the kind of day where I tightened a few bolts that will make things smoother going forward.