Let me know if this sounds like you: you ask Claude Code (or “Tor” in my case) to make a change on one branch, you don’t like it, then you ask it to back out everything and try again. It’s slow, it’s frustrating, and you end up losing good work along the way. In this video I show you the workflow I use instead: Git worktrees. The core idea is simple—create multiple independent copies of the same repo on your file system, point a separate Claude Code instance at each one, and let your agents run in parallel. I walk through a practical demo where I generate a landing page for a fitness tracker app (“Claude Fit”), then spin up three worktrees: one where I install a front-end design skill, one where I feed a reference image and have Claude generate a full design system into a docs/design folder, and one where I switch models entirely and use Gemini 3.1 Pro in Anti Gravity IDE. Then I compare the results side by side (including the original vanilla design), pick the winner, and merge that worktree back into main—while cleaning up the rest. The takeaway: worktrees make parallel agent workflows actually usable, especially for comparisons across approaches and models, without constantly switching branches.