Prompting Is Programming

A liberal-arts professor’s meandering conversation with Claude illustrates why many people get little more than drafts and summaries from AI, while software engineers report dramatic gains. Familiar explanations - models’ exposure to code, testable outputs, and investment in coding tools - are part of the story but do not explain why engineers also succeed with nontechnical tasks such as trip planning or grant writing. The central argument is that effective prompting is a form of programming: engineers have practiced specifying goals and constraints precisely, even when working in ordinary language.

Programming habits help users define the starting information and desired result, break a task into checkable steps, anticipate edge cases, and decide how to verify the output. Unlike people, who often ask clarifying questions, language models tend to fill gaps with plausible guesses, so vague instructions can produce convincing but unsuitable answers. Engineers are more likely to distrust such results, treat errors as clues about their instructions, and revise them deliberately rather than simply rerunning the same prompt. Domain expertise still matters for judging whether an answer is right, but careful specification is needed to get a useful answer in the first place. People who gain substantial value outside software often already use computational methods, such as complex spreadsheets or data scripts. Natural-language AI makes programming faster for those who think computationally; it does not automatically give everyone those habits.