Reading the standard library

If a config file has more than a dozen options, half of them aren't being used. The other half are being used in ways the original author didn't predict.

Most retry logic I've removed turned out to be hiding a real bug. The retries made the failure invisible, and the bug compounded for months in the background.

Latency budgets work better than latency targets. Targets are aspirational. Budgets force you to delete features when you go over.

Most of the trouble I see in production comes from boundary cases someone promised would never happen. The code is fine. The plan is fine. The promise is what ages badly.

When a system surprises me, my first instinct used to be to fix the surprise. Now my first instinct is to figure out which mental model of mine was wrong, and update that.

I keep a small log of decisions in a text file. Not architectural — just small ones. "Picked X over Y because Z." Future me reads it more than I thought he would.