The other set of hands.
Time for full disclosure, though regular readers may have already guessed: AI helps build this migration, and AI helps write this log. This entry is exactly how — and why I’m telling you on my own terms instead of waiting for a detector to tell you first.
Start with the honest inventory, because it’s bigger than you’d think. On this migration, I draft most code with Claude in a working conversation — design argued out first, code delivered as files I extract, commit, and deploy myself. A separate agentic instance works directly against the repository, read-only by default, doing reconnaissance: verifying field names, reading the platform’s source, confirming what’s actually deployed before anything gets written. My colleague on the sync side builds with a different AI assistant. Another colleague had a third one build an entire module — that story was entry 8. Three different AI systems have contributed code to this project, under two humans. That’s not an experiment anymore; it’s just how this shop works.
What makes that workable is the thing this series has been writing about all along: the gate. Nothing merges without review, whoever or whatever wrote it. The scar-tissue rules from entry 10 apply to machine output with extra force, because AI fails in a particular way — confidently, completely, and with excellent posture. It will overwrite your prices without knowing you have a pricing engine. It will assert that an API field exists because it plausibly should. The house rule I hold every assistant to is the same one I hold myself to: never guess. Say what you don’t know, verify against the real system, and bring findings before code. The assistants that work well here are the ones made to work that way; the failures this log has documented were, almost every time, a guess that nobody caught before it shipped.
Now the part you came for: this log is made the same way. Each entry starts with AI mining the actual working threads — the same conversations the code came from — proposing what’s worth telling, and flagging what the record can’t support. I decide what runs, correct what’s wrong, and cut what overstates. The accuracy discipline you’ve hopefully felt in these entries — “designed, not shipped,” “built, awaiting the carriers,” claims softened when the evidence is thin — is the review gate working on prose. More than once a draft has claimed something the record couldn’t back, and the draft lost. The words are collaborative; the judgment, the work being described, and the responsibility for every claim are mine.
Why disclose now? Partly consistency: a log that published “Reviewing the robot’s homework” would be ridiculous hiding its own co-author. Partly the calendar: AI-generated text now carries invisible watermarks — the major model providers began embedding them this year under new transparency rules — which means anyone with a detector may eventually be able to tell AI touched a given text. I’d rather you hear it from me, with the process attached, than from a tool with no context. And partly conviction: I think this is simply what competent software work looks like now, and a build log that hid it would be documenting a fiction.
So here’s the standing disclosure, now also a line on the log’s front page: these entries are written with AI assistance, from the real working record, verified and edited by me, published under my name — and everything described in them was reviewed by the same gate before it shipped. If that changes how you weigh the log, that’s your right. My own view is the one the whole series argues: don’t trust authors. Trust gates, receipts, and people willing to show you both.