StartupCities







I'm a Neanderthal who still wants to read the code I merge. But agent PRs are too big! They drive me crazy.
With "accept all" and "auto mode" a single task touches dozens of files and edits hundreds or thousands of lines. It'll only get faster and more autonomous from here.
I've found a way to fix this: a hook that repeatedly nags Claude to commit once it has finished a logical unit of work.
(I call the pattern "AI behavioral nudge.")
Any time it changes a file, the hook reads the diff and says "Hey Claude, you've edited X lines in Y files, you sure it's not time for a commit?"
It's simple but effective. After some edits, Claude will proactively propose a slice back to me: "I think we can ship {some change} as one unit."
When I approve, a skill calls Graphite to land a focused, stacked PR. No more brain-breaking 10,000 line PRs.
It's a bit embarrassing to have to build tools to enforce this kind of engineering discipline, but such is software in 2026.
Repo with skill + hook: claude-stacked-prs


















