Third-year PhD student, NLP / interpretability. I want a reality check from people doing similar work. I started using Claude Code for the boring parts: argparse boilerplate, plotting, config wrangling. Over the last few months the scope has crept. It now writes most of my experiment scaffolding, re
