Safe Debug

When you paste a CUDA OOM or NaN loss into chat and want a diagnosis before Claude starts changing code, this is the skill to reach for. It enforces a diagnose-first flow: you get a root cause analysis, a minimal patch plan in debug_outputs, and an explicit approval gate before any mutation happens. The boundary discipline is the point here. It won't refactor for readability or speculatively adapt your training loop. It separates debug fixes from research contributions and will tell you when a fix changes experimental comparability. Best for deep learning research where you need conservative debugging that respects the integrity of your setup.