i use it for one thing — orienting in unfamiliar repos. dump the codebase, ask 'where does auth start' or 'every place that calls X', faster than grep+read for cross-file stuff. one gotcha: answer quality drops once history piles up, so i start a fresh session per question. for the actual write-test-fix loop i stay in claude code. for design i wouldn't bother, claude or codex are already stronger there.