Question
An agent wrote a 40-line analytics SQL query to compute monthly active users with retention cohorts across three joined tables. It returns plausible-looking numbers and runs fast. The query uses LEFT JOINs, a window function, and a DATE_TRUNC you didn't scrutinize. The output will drive a board metric. How do you verify the numbers are right before they leave your hands — given you don't fully follow the query?
Treat the AI’s output as a draft to verify, not an answer to trust. Name the specific flaw and the input that triggers it, say how you’d catch it — tests, edge cases, reading critically — and how you’d re-prompt or decompose to get it right.
Vibe coding: describe the solution in plain language (or narrate it) and the coach grades your approach. Generating runnable code from your description is coming next.