Question
You want an AI agent to generate a complex analytical SQL query (Snowflake) computing 28-day rolling retention cohorts from an events table, and the result feeds an executive dashboard. A loose request gave you a query that looked right but double-counted users who churned and returned, and silently dropped users in timezones past UTC midnight. Describe the context, constraints, and acceptance criteria you'd put in the prompt so it's correct, and the verification you'd insist on.
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.