1. Say the grain firstFor every SQL prompt, state the row grain before joins or aggregation: one row per user, order, session, ticket, or day.SQL
2. Name the decisionOpen every case with the business decision: launch, rollback, target a segment, change pricing, or investigate deeper.Product case
3. Separate metric and diagnosisDo not jump from "conversion dropped" to a cause. Split the read into metric definition, segment, timing, and likely mechanism.Metrics
4. Keep one sanity check readyFor SQL, mention nulls, duplicates, timezone windows, bot traffic, deleted rows, or many-to-many joins before trusting the result.Quality
5. Practice a flat experimentExplain what you would do when an A/B test is neutral overall but positive for one segment and negative for another.Experimentation
6. Build a 60-second project storyProblem, data source, metric, method, result, tradeoff, and recommendation. Stop before it becomes a resume walkthrough.Story
7. Translate analysis into actionEnd answers with an operating recommendation, not just a chart: who should do what, when, and what would change your mind.Recommendation
8. Prepare one failure storyUse a real miss where the analysis, stakeholder read, or rollout could have been better. Say what you changed afterward.Behavioral
9. Ask for the missing constraintWhen a prompt feels vague, ask about goal, baseline, timeframe, user segment, data availability, and acceptable tradeoffs.Ambiguity
10. Rehearse out loudWrite less. Speak the answer, time it, then rewrite only the opener and recommendation until they sound natural.Mock