When did you realise you were depending on claude too much?

I build most of my app with Claude, so the question I ask myself is a slightly different one: if the chat vanished tomorrow, would I still know why the code is the way it is? My answer was to make it write things down. Every non obvious finding, a vendor quirk, a toolchain trap, a decision and its reason, goes into a topic file in the repo at the moment it is learned, not into a commit message or the chat. There is one file per area, and a short rules file on top. That turns the dependence into something I can read and check. The part I keep for myself is anything hard to walk back: destroying data, or anything that reaches real users needs my word first.