We batched software into releases for one reason: shipping was expensive. Sprints, versions, launch days: all of it was a coping mechanism for a world where every deploy hurt.
That world is gone. Now batching is the risk.
A release is fifty decisions in one crate. When one is wrong, you roll back all fifty and spend a day finding it. So at Ledger Collective we stopped batching. Every pull request is one thought — one decision — carrying its own proof: what changed, why, and the checks that show it holds. When something breaks, you know exactly which decision it was. You take back just that one.
Verify, don't skim. At AI tempo nobody can read every line, and pretending is the real risk. A rubber-stamp review is worse than none, because everyone downstream believes a human checked. So ask a better question than "did someone read this?" Ask "what would prove this is right?" In our world the purest answer is a reconciliation: the math tools do the math and the numbers foot to PASS, WARN, or FAIL, within a penny. Nobody's opinion is invited.
Records that can't lie. Documentation is honest the day you write it, and drifting the day after. We didn't fix that with discipline. We deleted the second copy. Our tool directory is generated from the code; the automation we review in the pull request is the same file production runs. The record is true as of the last commit.
Our security reviewer doesn't re-read the code. They check that the gates are intact and the evidence trail is unbroken. That's the whole trade: we don't work more. We ship smaller, with proof.
Review scales with headcount. Verification scales with discipline.
— Ashley
First shared on stage at the Capital One AI Speaker Series in Plano, Texas. Thank you, Howard Dierking, for hosting. One of four essays from that talk.
Keep reading Prose Is the Programming Language Now. → If you can write a memo to a new hire, you can write software. Most teams still can't write the memo.