A policy analyst pastes a dataset summary into Claude and types: "What does this tell us?" Claude returns three paragraphs describing trends, noting that income correlates with outcome, and suggests that "the results indicate a significant relationship." The analyst copies the paragraph into a report. The problem: the variable labeled "income" in the dataset is categorical, not continuous. The "significant relationship" Claude described was never tested , Claude inferred it from the pattern of numbers in the table. The analysis sounds authoritative. None of it was verified.

Analysis support prompts earn their value at the planning stage , before the test is run, when a misspecified question can still be fixed. The next chapter applies the same specification discipline to a different kind of cognitive labor: learning. When the goal is understanding, not just output, the prompt must be designed so the human does the thinking. The stakes of skipping that design are different from analysis errors, but the mechanism of the failure is the same.

Analysis support prompts earn their value at the planning stage , before the test is run, when a misspecified question can still be fixed. The next chapter applies the same specification discipline to a different kind of cognitive labor: learning. When the goal is understanding, not just output, the prompt must be designed so the human does the thinking. The stakes of skipping that design are different from analysis errors, but the mechanism of the failure is the same.