My Claude Code Git Workflow: Small Commits, Clean History, Zero Fear

Before I built this workflow, my AI-assisted sessions produced exactly what you would fear: 800-line commits titled 'fix stuff.' Claude Code writes excellent commit messages, but only if you structure the session so each commit is one coherent change.

The unit of work is the task, and I state it explicitly: 'We are doing one thing: extract the email validation into a shared module. Nothing else.' Claude is happy to fix the three other bugs it notices, and every one of those well-meant detours is how you get a diff nobody can review. Related bugs go on a list for later, on paper, out of the session.

Between tasks I have Claude run the ritual: tests, then stage the diff, then 'write the commit message from the diff, conventional commits, under 72 characters.' Because the session context contains exactly one task, the message is accurate. Because the diff is small, the message is specific. This is the whole trick and it is entirely about session discipline.

Branch strategy matters more with an agent, not less. I work on feature branches and make Claude the reviewer before merge: 'review the full branch diff against main as a skeptical senior engineer.' It catches the cross-commit issues that single-diff reviews miss, like a helper added in commit two and abandoned in commit five.

The surprise benefit was archaeological. Six months in, my history reads like a well-run human project: every change explainable in one line, every revert surgical. When Claude Code sessions are structured as small committed tasks, the AI's tendency toward breadth gets channeled into documentation instead of becoming tech debt.