AI should be treated as a function, not a magical genie. Treating AI like a search box for one-off queries is fine, but it's inefficient for repetitive tasks. To build a useful AI assistant, start with a single, narrow, repetitive, boring, and verifiable task. Think of prompts as function signatures, with clear inputs and structured outputs. To ensure reliable summaries, force the AI to quote supporting sentences from the source material. For data cleanup, avoid in-place changes and instead have the AI output a mapping for review. Crucially, verify AI outputs by creating a small "golden set" of correct examples. Implement cheap checks like code assertions and spot-checking to catch errors early. Only introduce agents for tasks that genuinely involve multiple steps or tool calls. Finally, iterate by incorporating successful AI outputs back into the prompt to create a predictable and reliable assistant.