Finding focus and engagement in the fog of AI

The speaker reflects on their recent shift in productivity, emphasizing a renewed focus on slow, deliberate work rather than rushing through tasks. They describe their process of detailed review and interaction with AI-generated code, highlighting the importance of understanding every line and decision. The discussion emphasizes working on one thing at a time, prompting responsibly, and maintaining deep engagement rather than multitasking or rushing to ship features.

The podcast delves into the speaker’s approach to working with AI in coding tasks. They describe a meticulous process of prompt engineering, step-by-step review, and iteration with language models (LLMs). The speaker critiques some of AI’s tendencies, such as scaffolding and reliance on hacks like monkey patches, and stresses the importance of internal consistency, API design, and fundamental system architecture over quick fixes. They believe deeply understanding and controlling the codebase leads to better, more harmonious results.

A major theme is the necessity of discipline and focus when working with AI. The speaker emphasizes that engagement and managing one’s own focus are crucial to success, rather than blaming AI or external tools. They advocate for working slowly, prompting responsibly, and avoiding distractions, which in turn fosters better results. They believe that a clear methodology—working one thing at a time and understanding every detail—enhances the development process and produces more reliable, harmonious code.

The speaker discusses their values around API consistency, internal and external quality, and the pain of scaffolding hacks. They prioritize building API systems that make sense and fit well together, resisting the temptation to introduce shortcuts or precedents that could lead to inconsistency. The focus is on patience and waiting for abstractions to reveal themselves naturally, in order to maintain a harmonious and predictable codebase.

The discussion touches on the variety of AI workflow terminology and the rapid evolution of concepts like loops, agents, and orchestration. The speaker compares this to early programming patterns and emphasizes that, for their style of work, the most effective use of AI is through direct, intimate interaction—working with a single session in a slow, deliberate manner. They reflect on how engagement and focus, rather than complex architectures or paradigms, are key to harnessing AI effectively in their workflow.