Give the first step a finish line
A broad request leaves many decisions unstated: what counts as a problem, which files matter, and whether modeling is appropriate. A smaller question creates a result you can check before committing to more work.
Try “Find missing values and repeated rows in this dataset, and explain which columns need attention. Do not fit a model yet.” This names the input, task, output, and boundary. You can judge whether the answer did what you asked.
Make the context explicit
Tell the assistant what one row represents and which column, if any, is the prediction target. If observations share a batch, laboratory, source paper, or experimental family, explain that relationship. Those details influence what a fair evaluation looks like.
Attach only the relevant materials. In a continuing project, check the selected file chips before sending a follow-up. Context is useful when it applies to the question at hand.
Use the plan as a conversation
Turn on plan review for an unfamiliar or expensive investigation. Look for an understandable connection between the goal, the proposed tools, and the expected evidence. Ask for a narrower plan if that connection is unclear.
After the first step, inspect the saved results. Then ask the next question the evidence supports. Iteration is valuable when each step reduces a real uncertainty.