我的 AI 工作流程从提示演进到近乎自主的流程。
1 分•作者: hoangnnguyen•4 个月前
在过去的 6 个月里，我一直在构建 AI DevKit 的同时，不断改进我的 AI 编码工作流程。<p>最重要的转变并非在于更好的代码生成，而在于从可复用的命令和模板，转向能够携带上下文、触发正确行为并自动验证工作流程。<p>最近的一个功能使这种差异变得显而易见。我使用 Codex 为 "ai-devkit skill add" 构建了交互式技能选择。我只给了一句话的指令，工作流程就完成了需求、设计、规划、实现、验证、测试和代码审查的整个过程。<p>整个过程耗时不到一个小时，实际的功能流程大约 30 分钟。<p>我发现有趣的不只是 AI 编写了代码，而是工作流程留下了需求文档、设计文档、规划工件、从需求中派生的测试，以及针对规范的验证，而不仅仅是一个差异（diff）。<p>在实践中，有几件事感觉很重要：<p>- 记忆功能找回了我之前忘记存储的旧 CLI 规则
- 审查阶段可以向后循环，而不是盲目地向前推进
- 验证功能捕捉到了实现和设计之间的偏差
- 我仍然自己做出了产品决策，并修复了最后一个未通过的测试<p>我很好奇这里的其他人是如何思考这个问题的。<p>你们是主要在优化提示词，还是现在开始尝试优化模型周围的工作流程层？
查看原文
Over the last 6 months, I have been evolving my AI coding workflow while building AI DevKit.<p>The shift that mattered most was not better code generation. It was moving from reusable commands and templates into a workflow that can carry context, trigger the right behavior, and verify work automatically.<p>A recent feature made the difference obvious. I used Codex to build interactive skill selection for "ai-devkit skill add". I gave one sentence of instruction, and the workflow carried the task through requirements, design, planning, implementation, verification, tests, and code review.<p>The whole session took under an hour. The actual feature flow was around 30 minutes.<p>What I found interesting was not just that AI wrote code. It was that the workflow left behind requirements, design docs, planning artifacts, tests derived from requirements, and verification against the spec instead of just a diff.<p>A few things that felt important in practice:<p>- memory pulled back an old CLI rule I had forgotten I stored
- review phases could loop backward instead of blindly moving forward
- verification caught drift between implementation and design
- I still made the product decisions and fixed the last failing test myself<p>I am curious how others here are thinking about this.<p>Are you mostly optimizing prompts, or are you now trying to optimize the workflow layer around the model?