像 Claude Code 这样的 AI 编码工具使执行变得廉价且快速，但需要清晰的前期规划以避免构建错误的东西。 AI coding tools like Claude Code make execution cheap and fast, but require clear upfront planning to avoid building the wrong thing.
AI 放大了半成型想法中的模糊性，使敏捷方法效果降低；规范驱动开发能带来更好的结果。 Ambiguity in half-formed ideas is amplified by AI, making agile methods less effective; spec-driven development yields better results.
用自然语言编写的行为契约，定义可接受的结果和回归测试，能改善 AI 输出并增强控制。 A behavior contract in plain language, defining acceptable results and regression tests, improves AI output and control.
AI 在给定工作容器时表现最佳，应作为原型工具而非协作者。 AI works best when given a container to work inside, not as a collaborator but as a prototyping tool.
设计的真正工作从执行转向决定什么值得制造，因为制造变得毫不费力。 The real work of design shifts from execution to deciding what is worth making, as making becomes effortless.
局限 · Limitations
作者的经验仅限于使用 Claude Code 的单个项目；其他工具或领域的结果可能不同。 The author's experience is limited to a single project with Claude Code; results may vary with other tools or domains.
该方法需要大量的前期思考和规范，可能不适合所有开发工作流程。 The approach requires significant upfront thinking and specification, which may not suit all development workflows.
AI 仍然无法在没有明确定义的情况下保持上下文或推断隐形边界，限制了其自主性。 AI still cannot hold context or infer invisible boundaries without explicit definition, limiting its autonomy.
作者指出与 AI 缺乏真正的协作，缺少人类团队合作的协同效应。 The author notes a lack of genuine collaboration with AI, missing the synergy of human teamwork.
论文章节 · Sections（共 8）
概述Overview
执行廉价，思考无价Execution Is Cheap. Thinking Isn’t
这是前期的大设计吗？感觉确实如此。Is this Big Design Up Front? It sure feels like it.
Claude 在思考过程中未能支持我Claude failed at supporting me in the process of thinking.
这一切都关乎契约吗？Is this all about contracts?
如果不能与 Claude 共同思考，我需要在 Claude 之前思考If I can’t think with Claude, I need to think before Claude
廉价且快速的对与错Cheap and fast right, cheap and fast wrong