Safe Debug

When a deep learning experiment fails with a confusing error message, this skill helps you stay calm and methodical. It focuses on conservative diagnosis before any code changes are made. You get a clear root cause and a minimal patch suggestion that requires your explicit approval to apply. The skill treats debug fixes as separate from research contributions, so you never accidentally change your experiment's meaning.

It works best with terminal errors, CUDA out-of-memory issues, checkpoint load failures, shape mismatches, and NaN loss symptoms. The goal is to narrow down the problem safely without rushing into changes. Outputs include a diagnosis report, a patch plan, and a status file for tracking.

This approach is ideal for researchers who want rigor and reproducibility. It avoids automatic patching or broad refactoring. Instead, it guides you step by step, keeping your codebase safe while you find the real cause of the failure.