I feel like I'm still not understanding something. How does making the output from the LLM lower level help?

Concrete example: Next/Turborepo. These tools make your life easier if you drink some kool aid. Rather than have the agent scaffold the app you have the agent use a tool that scaffolds. Agents write specs to manage tools, and those tools scaffold the code, then the agents just sprinkle in business logic that is too bespoke for codegen.