I kept pushing code Claude wrote that I couldn't explain, so I made Claude quiz me before every push
retained text
After a few months of Claude Code doing most of my typing, I noticed something uncomfortable: my PRs looked fine, but if a teammate asked me "why this approach?" I had to go read my own code first. So I built grip, a git pre-push hook. It reads your diff, asks five questions about it, and grades your answers. Below the pass mark, the push waits. The Claude part: * It uses your existing Claude Code install as the model, so no API key needed. * There's a /grip skill: Claude asks you the questions in chat, and it's told to never answer them for you. * There's a gate that stops Claude itself from running git push until you've passed. The newest mode draws your change as a whiteboard in the browser and pins the questions on it, inspired by Mitchell Hashimoto's "whiteboard defense" post. Free, open source, Python: [github.com/guilyx/grip](http://github.com/guilyx/grip) Try it offline first if you're curious: pipx install grip-hook, then grip quiz --provider fake in a test repo. Curious if anyone else has a ritual for this, or if you just trust the tests.
document fetched 2026-10-22 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1xbjn2x/i_kept_pushing_code_claude_wrote_that_i_couldnt