Shipping 30% faster with AI in the loop
Where Claude Code and Cursor genuinely pay off on a team of six — and the review habits that keep quality from sliding.
The 30–40% figure is real but narrow. It's not that we think faster. It's that the work between having a decision and having it shipped — scaffolding, tests, migration files, the fourth CRUD resource that looks like the first three — collapsed.
Where it pays
Test generation is the clearest win: describe the edge cases in a sentence, get a table-driven suite, then spend your attention on whether the cases are the right ones. Refactors across many files are second. Reading unfamiliar code is third, and underrated — asking why a module exists is faster than reconstructing it.
Where it doesn't
Anything load-bearing on a decision the model can't see. Our idempotency scheme, the breaker placement, the read-flip order in the Postgres migration — those came from knowing which failures we could tolerate. A model that can't be paged has no opinion worth taking on that.
AI shortened the distance between decision and diff. It did not shorten the distance to the decision.
The review habits that kept it honest
Three rules, and they're the whole system. The author reads every generated line before it goes in a PR — no exceptions, and it's a firing-line question in review. Generated tests don't count as review of generated code. And any diff the author can't explain out loud gets closed, not fixed.