Scaling large language models (LLMs) for broader codebases is a challenge without a clear solution. Investing in guidance and oversight is essential for improving how LLMs contribute to software engineering. Guidance involves providing context and environment, while oversight focuses on the skills needed to evaluate and validate the choices made by LLMs. One of the main goals is to enable "one-shotting," where an LLM produces a high-quality implementation in one attempt, avoiding the inefficiencies of rework. To enhance LLM performance, creating a prompt library is recommended. This library should compile documentation, best practices, and a clear map of the codebase to help engineers use LLMs effectively. Iteration is key; every time an LLM makes a mistake, the prompt library should be refined to clarify what could have been better communicated. The article emphasizes that a clean and well-structured codebase is crucial, as the quality of input directly affects the output from LLMs. If engineers struggle to understand the code, LLMs will likely falter as well. Oversight is equally important. Engineers must possess the ability to assess the decisions made by LLMs, such as whether choosing Redis over Postgres for data storage was appropriate. Building design skills within teams is vital for creating a robust architecture that facilitates future development. This can be achieved through studying existing code, replicating successful designs, and gaining practical experience. The article suggests that while some aspects of oversight can be automated—like implementing safety checks to prevent errors—human expertise remains crucial for aligning technical solutions with product goals. Overall, the interplay between effective guidance and oversight enhances the ability of LLMs to generate useful code. By investing in both areas, organizations can improve their engineering processes and ensure that LLMs serve as valuable tools rather than replacements for skilled engineers.