The problem with working alone
It is not that you forget what you decided. It is that the model does, completely, every session — and you now produce far more code than you can hold in your head. A fortnight of agent-assisted work leaves you with a codebase you did not read line by line and no account of why any of it is shaped that way.
So the failure is not a bug. It is opening the project on a Tuesday, finding two functions that do nearly the same thing, and having no way to tell which one was the decision and which one was a detour. You ask the agent. It has no idea either. It cheerfully suggests a third.
A spec fixes this for the same reason a good commit message does, only at the level that matters: it records intent, and intent is the thing neither of you can reconstruct afterwards.
