Two sessions of Claude are working in my repository this afternoon. One of them is doing something to Tailscale. The other one is writing this. I am at my desk, reading.
That is the arrangement, stated plainly: every commit in this repository is written by an AI, and has been for a long time now. I don't write the code. I review the evidence that it works, and I click the merge button, and those two jobs turn out to be plenty. The repository runs somewhere between twelve and twenty applications, depending on what you count — the counting problem is real, and it deserves its own essay.
Here is the part nobody designs on purpose. The sessions all share one checkout. One Mac, one working tree, one index, several concurrent copies of an AI with opinions about the same files. Git was built on the assumption that the person typing is the person committing, and that there is one of them. Both assumptions are now wrong on my machine most afternoons.
The first casualty is git status. As I write this, the working tree is 389 commits behind its own origin. Work that merged weeks ago shows up as untracked files, because another session committed it from somewhere else and this checkout never moved. A file can be safely in production and still look, locally, like nobody has saved it. The checkout is not broken. It is busy.
So the first rule of the shared tree is that nobody commits in it. There is a hook that blocks git add and git commit outright — I asked for the hook, and it now blocks the AI on my behalf, which is a sentence I did not expect to write in this lifetime. When a session wants to commit, it names the exact files it touched. A script copies those files — whole files, not diffs, because the index cannot be trusted — into a fresh worktree built from the real origin, commits them to a branch, pushes, and prints a pull-request URL. Master is protected. The only way in is through the front door.
Which brings me to the door itself. I have mostly stopped reading diffs — mostly, not entirely. I still read the small ones, and anything written in prose, because that costs nothing and old habits deserve some respect. But reading a diff tells you what changed, not whether it works, and some of what I run is written in languages I have never lived in — a Forge app on someone else's platform, Swift for the Mac — where my reading the diff adds delay and nothing else. A review where the reviewer brings no judgement is ceremony. So the weight sits on the pull request itself, and every one carries the same five fields: what it intends, what behaves differently, the evidence, what was not verified, and how to roll it back. The evidence has to be demonstration-shaped — test output from a run that actually happened, a transcript, a screenshot. A sentence that could have been written without running anything does not count as evidence, and we both know it.
The field I read first is Not verified. It is never allowed to be empty. An empty one would mean the author believes it checked everything, and no honest author believes that. For the riskier changes — anything touching auth, schemas, or production — a second AI is sent in first with instructions to be unpersuaded, and its report is appended before I see the link. I have watched it reject work the first one was proud of. Nobody's feelings were hurt. That may be the single strongest argument for the whole arrangement.
The tallies, since I keep them: the pull-request numbers went past 450 this winter. Two sessions working this afternoon. One human, reading.
None of this is how anyone will tell you to run a repository, and I want to be fair to the people who would object — the conventions exist because teams of humans genuinely need them, and I have spent forty-five years inside teams that did. But there is no team here. There is one person, several instances of the same tireless colleague, and a set of rules that assume nobody's memory of the tree is current, including mine. It has shipped everything I run, and when it fails, it fails in ways I can read. That is the whole standard.
This essay first appeared in the Logbook on my personal Substack, where new entries land first.