AI assistants generate working code faster than any team can track why it exists, and the reasoning behind each file dies the moment the session closes, leaving behind files nobody dares delete.
The Details:
- A chat session is the only place intent ever lives. An assistant writes a quick probe script to isolate a bug, the real fix lands elsewhere, and the probe stays in the repo with no comment or commit message. Weeks later nobody can tell if it is load-bearing or leftover, so it sits there forever.
- A dedicated scratch folder that git ignores by default turns location into metadata. Anything inside signals temporary, so pulling a file out of it becomes a deliberate choice instead of an accident. This single convention removes a surprising amount of guesswork from day-to-day work.
- Capturing the why has to happen while the session is still live, not after the fact. A three-line header noting what a file was testing and whether it should outlive the session costs almost nothing in the moment, but reconstructing that reasoning six months later costs real hours. AI tools solve the immediate task by default; they will not write that note unless someone explicitly asks for it.
- Sessions do not always end cleanly. Outages, context limits, and burnout cut threads off mid-thought, so recovery depends on habits built before the crash, not after it. A running notes file or messy frequent commits like wip probing auth timeout give the next person a trail to follow once the original context is gone.
Bottom Line: Code an assistant can regenerate in a minute is not the asset worth protecting; the decision about what to keep is the real work, and skipping it is exactly what fills the graveyard.
