The AI PM's Operating Loop

A working cadence: Monday — read last week's failure traces, not tickets: actual transcripts and output samples. Midweek — eval suite health: pass rate, new failure classes, eval additions from production. Release review — every candidate change gets an eval diff, a cost delta, and a rollback plan. Friday — capability watch: what did the frontier models just make feasible, and does it change the roadmap?

The loop's spine: production failures → eval cases → fixed behavior → shipped change.