Nuanced应用的开发者近日分享了一项重要发现：在AI代码生成领域，传统的规范化计划模式正在失去价值1。该应用最初的设计理念是通过计划模式帮助开发者在AI快速生成代码时保持对系统的理解1。然而，经过实践验证，作者发现这一方案存在根本性缺陷1。
计划模式曾被寄予两个主要期待：为AI代理提供精确的指令，以及帮助人类理解构建的内容1。但随着AI模型能力的提升，显式指令的必要性大幅下降1。同时，用户普遍缺乏阅读AI生成的冗长文档的意愿，而将计划与构建严格分离反而破坏了开发工作流的连续性1。
作者以Codex的工作流为例说明了更有效的方式：理解、行动、检查、澄清、调整、再次行动的循环过程，其中计划隐含在整个循环之中，而非作为独立的规范性文档存在1。这一认识引出了核心结论：将计划作为离散的文档工件本身就是根本错误1。真实的软件开发应当是计划与构建交织进行，未来的改进方向应是设计流程来增进人类理解，而非生成形式化的计划文档1。
The developer behind Nuanced, an AI programming application, has concluded that the explicit planning model that was central to the product's design has become obsolete.1 Nuanced was built on the premise that a structured planning phase could address a critical problem in AI-assisted coding: as language models rapidly generate code, developers risk losing understanding of what the system is building.1
The planning mode served two traditional purposes—providing AI agents with sufficiently precise instructions and helping humans comprehend the generated output.1 However, the developer found that this approach failed for several reasons.1 As AI model capabilities have advanced, explicit instructions have become less necessary; users proved unwilling to read lengthy AI-generated planning documents; and the rigid separation between planning and implementation disrupted natural workflow continuity.1
Rather than treating planning as a standalone documented artifact, the developer argues that real software development interweaves planning and construction continuously.1 The Codex workflow exemplifies this pattern: understand → act → inspect → clarify → adjust → act again, with planning embedded within the iterative cycle rather than isolated as a preliminary phase.1 The fundamental lesson is that processes should be redesigned to improve human comprehension through other means, not by generating normative planning documents.1
评论
还没有评论，欢迎留下第一条。