Claude Code, three months in: what changed

The wins

• Multi-file refactors that used to take a morning now take an hour, with fewer regressions.
• CLAUDE.md files replaced a lot of tribal knowledge — new devs onboard against them instead of asking us.
• The plan-mode default reduced 'confidently wrong' edits by roughly half compared with Cursor's agent.

The friction

Token cost climbs fast on large monorepos. We had to be deliberate about which folders it could read, and we still burn more per developer than Copilot ever did.

It occasionally over-plans small tasks — for a one-line fix, dropping into direct edit is still faster.

Verdict

Still our daily driver. Nothing else touches it on long-horizon work, but pair it with a cheaper autocomplete for typing-speed edits.