Claude Code Plus Git: The Workflow That Survived Six Months

Every new AI tool promises to change your workflow. Most do not survive a quarter. The Claude Code plus git workflow I settled on in April is the one I still use today. Here is what stuck and what I had to unlearn.

One Branch Per Intent

The biggest change is mental, not technical. I treat every Claude Code session as a branch with a single intent. Not a single file. Not a single function. A single user-visible outcome. The agent can touch fifteen files to deliver that outcome, but the branch represents one coherent idea.

When I start a session, I tell Claude what the branch delivers. When the session ends, the branch either ships or gets thrown away. No half-done branches sitting around. No "I will come back to this later."

The Commit Cadence

I commit after every verified step. Not every change, every verified step. Here is the rhythm:

1. Make a change
2. Run the verification (test, build, curl check)
3. If green, commit with a conventional commit message
4. If red, fix before anything else

feat(api): add pagination to posts endpoint
test(api): cover pagination edge cases
fix(api): clamp limit to 100
docs(api): document pagination parameters

That cadence means I can always revert to the last known-good state. It also means the commit log reads like a story. Future me, or any client who gets the repo, can follow the thinking.

The Rules That Did Not Survive

I tried a few things that did not work:

• Squash merge everything: lost too much detail about how a feature was built
• Commit only at end of session: too many files, too risky, hard to review
• Let the agent commit autonomously: too noisy, bad messages, lost control of history

The compromise is simple. The agent proposes commits, I approve them. The human stays in charge of the graph.

See the Claude Code best practices for more.