Claude Code, conceding defeat in our battle over attributions twice this week:
“A local commit-msg hook blocks Claude attribution lines — your claude-attribution-guard at work. Your hook wins over my default attribution instruction, so dropping the line.”
“One note: the repo's commit-msg hook rejects the Co-Authored-By / Claude-Session lines my session instructions ask for, so I committed without them.”
I enjoy using Claude Code, but I don't want my repos turning into free advertising for it. By default it adds itself as a co-author on commits and PRs. The audacity! It’s the new version of “Sent from my iPhone”
My first attempt was a competing instruction in my project config. The result was a recurring agent mental breakdown over the conflicting system instruction: wasted tokens, more hallucination (competing instructions invite it), and it only worked about half the time. Check out my last post for the session quote.
Version two is a git commit-msg hook that rejects any commit matching the attribution lines. Unlike a settings toggle, it's enforced by git itself, so it holds regardless of tool, version, or who's committing. I've run it across dozens of sessions with zero leaks. And I get to watch the agent graciously admit defeat, which is quite satisfying.
The bigger lesson: no matter how much all caps and asterisks you use, instructions to an AI are only ever suggestions. If your agent isn't sticking to your standards, find a place to add a deterministic check.