Coding agents are insanely smart for some tasks but lack taste and good judgement in others. They are mortally terrified of errors, often duplicate code, leave dead code behind, or fail to reuse existing working patterns. My initial approach to solving this was an ever-growing CLAUDE.md which eventually got impractically long, and many of the entries didn’t always apply universally and felt like a waste of precious context window. So I created the dev guide (docs/dev_guide/). Agents read a summary on session start and can go deeper into any specific entry when prompted to do so. In my original project the dev guide grew organically, and I plan to extend the same concept to my new projects. Here’s an example of what a dev_guide might include:
以下展示一个UniScientist进行的完整科研推理链条，详细推理内容可以在Blog链接中查阅：https://unipat.ai/blog/UniScientist，这一点在新收录的资料中也有详细论述
，更多细节参见新收录的资料
Sure enough, I found a broken crescent bond on the long bond wire crossing over the controller die, right where it should have connected to the package.
贵州航宇科技为整机厂商提供核心锻件，订单持续饱满；山东豪迈机械聚焦燃气轮机零部件加工，产能高位运转；烟台杰瑞石油服务集团突破技术瓶颈，成功挺进北美数据中心燃气轮机配套市场，实现国产零部件海外高端市场的突破……，这一点在新收录的资料中也有详细论述