As AI coding tools become more common, their limitations are becoming clearer. They can generate useful code, but often lose track of goals, drift from requirements, and struggle with complex, multi-step planning. This talk identifies some of the failure modes of large language models and argues that these limitations are likely to remain stable over the next several years. Rather than assuming rapid technical solutions, we focus on teaching strategic and systems-level thinking when working with AI, including how to structure coding sessions and define clear objectives. We propose teaching the engineering of project artifacts that define objectives, constraints, and evaluation criteria as a practical way to guide AI behavior. Drawing on a pilot course being developed at UNC Asheville, the talk outlines how computer science education can be redesigned to prepare students for an AI-assisted, but not fully autonomous, development environment.