The same question asked two ways can produce a generic answer or a precise one. These habits reliably produce the second kind.
Ground in your data
Name the module, the period, and the population: "in Revenue OS, deals over $5k created since June" beats "our deals". Cortex resolves vague questions by guessing scope; precise questions remove the guess.
One question per message, then iterate
Cortex handles follow-ups well because conversations persist. Ask, read, then narrow: "now only the enterprise segment", "exclude renewals". Iterating is cheaper and faster than composing one giant question.
Answer the clarifying questions
When Cortex asks "which pipeline?" or "calendar or fiscal quarter?", that is the system refusing to guess. Answering takes seconds and costs nothing extra.
Choose the right depth
- ›Quick answer (about 2 credits): lookups, summaries, "what changed".
- ›Deep analysis (about 10 credits): causes, comparisons, forecasts, recommendations. Worth it when you would otherwise spend an analyst-hour.
- ›Simulate (about 8 credits): "what happens if we raise prices 10%" - explicit scenario modeling.
Give feedback
Correct wrong assumptions in the thread; corrections persist as memory. If an answer cites thin data, that is a prompt to import more or connect a source - see Import your data.