Today I wrote down the goal that has governed every build since, and it fits in one sentence: every feature must run live and autonomously, on real data, with no restarts from a chat window and no human babysitting it.
That sounds obvious. It is not how most AI-assisted development actually works. The default failure mode looks like this: you sit in a session with an AI agent, it builds something impressive, it works while you watch, and the moment the session ends the work starts decaying. State lived in the conversation. The fix was applied to a tree that got reset. The cron job was registered in a way that dies with the login session.
I had weeks of that behind me. Features that existed only as long as someone was watching them.
The standing directives
The master goal came with sub-rules that are still in force today:
- Destroy silent failures. Every failure must surface loudly in a dashboard a human actually looks at. A system that fails quietly is lying to you by omission.
- Durable primitives over local duct tape. Anything that matters runs under real process supervision, writes to a real database, and survives a reboot without a human noticing.
- Live proof before "done." Nothing gets marked complete because the code looks right. It gets marked complete when it demonstrably ran, on real data, unattended.
Why this mattered more than any feature
The uncomfortable insight underneath this: the system "kept needing fixing" not because the individual pieces were bad, but because they were built on fragile foundations and work that evaporated on the next reset. Making a rotten foundation faster is not progress. It's just faster rot.
Writing the rule down didn't fix anything by itself. What it did was give me a test I could apply to every piece of work from then on: if I close this laptop for three days, is this still running when I come back? If the answer was no, the work wasn't done, no matter how good the demo looked.
Three days after writing this, I applied that test to the entire system at once. It did not go well, and the result was the biggest decision of the whole project.