A graduate student studying for a qualifying exam types into Claude: "Explain the difference between classical and Bayesian inference." Claude delivers four paragraphs. The student reads them, feels like she understands, and closes the tab. Three days later, on the exam, the question appears. She cannot answer it. This is not Claude's failure. The explanation was accurate. The failure was in what the prompt asked for , an explanation delivered to a passive reader , rather than what learning requires, which is active construction. Ambrose et al.

Learning prompts protect cognitive labor by structuring what the learner must do before Claude responds. Handoff prompts , the subject of the next chapter , protect a different kind of labor: the judgment that defines what an autonomous agent is allowed to do on your behalf. Both share the same discipline: a vague prompt produces an output you cannot supervise. A specified prompt produces an output you can check, correct, and reuse.

Learning prompts protect cognitive labor by structuring what the learner must do before Claude responds. Handoff prompts , the subject of the next chapter , protect a different kind of labor: the judgment that defines what an autonomous agent is allowed to do on your behalf. Both share the same discipline: a vague prompt produces an output you cannot supervise. A specified prompt produces an output you can check, correct, and reuse.