Boris Cherny ships 20-30 PRs a day. Opus introduced maybe 2 bugs in an entire month. He would have introduced 20 writing by hand. But the number that matters is the one nobody is talking about: the kill rate. The Claude Code team prototyped "probably hundreds of versions" of agent teams before shipping. Built ~30 prototypes of condensed file view, then dogfooded for a month. Ran 50 to 100 iterations on a terminal spinner. 80% of those never shipped. That ratio is the entire point. When building drops to near-zero cost, the bottleneck moves from "can we build this" to "should we ship this." And the answer to "should we ship this" is no, 4 out of 5 times, even when the prototype works. This is where most teams get the AI workflow backwards. They use AI to build faster and ship everything. Boris uses AI to build faster and kill almost everything. The speed is in service of judgment, not output. The printing press analogy lands hard. Scribes became authors. The skill shifted from production to editorial selection. PMs are watching the same transition happen to their craft in real time. The 5 Lenses framework in the piece is the most practical thing I've seen on this topic. Problem-solution fit, interaction cost, edge case exposure, technical debt signal, business model alignment. Run those on a working prototype in 15 minutes and you catch what a 15-page PRD never would. The PMs who prototype one feature per month and evaluate it ruthlessly will outperform the PMs who ship 10 features per month with zero filtering. Volume of judgment reps compounds faster than volume of output.