Ask HN: How Do You Maintain Understanding and Velocity When Using AI?

Summary

The Hacker News discussion comes from a developer who sees clear leverage from AI for logical code search, design brainstorming, and investigative work, where results can be verified quickly. They find AI less helpful for writing design documents and implementing code because they review each output for alignment with intent, sound abstraction, comprehensibility, and overlooked edge cases. This creates extended back-and-forth with the agent: during design, the agent may initially dismiss an important dependency and later acknowledge that it is load-bearing; during coding, the developer questions the generated logic until they understand and accept the approach. The process can uncover subtle problems and occasionally major ones, so the author considers the caution justified for systems where reliability and performance are central, rollouts are slow, and failures may have a large blast radius. After documenting their reasoning and reviewing nearly every generated line, they feel they have saved little time compared with writing the work themselves and are more exhausted. The post asks whether a different workflow or set of assumptions can make AI-assisted development smoother, or whether other developers face the same difficulty.