黄潇 Shaw Huang

关于我About

我是做远程 IO 模块技术服务的，三年多。售前选型、售中调试、售后诊断 都干，攒了 600 多份工单，阀岛通讯、现场总线这类问题处理得最多。 用 AI 比这早。2025 年 2 月 DeepSeek 刚上线我就开始用网页版了， 算最早一批用户。那会儿它服务器老是繁忙，问的人太多。

I'm a technical service engineer for remote I/O modules, three years and counting. Pre-sales selection, on-site commissioning, after-sales diagnostics: I've done the lot, 600+ work orders, a lot of them valve-terminal communication and fieldbus problems. I got into AI earlier than that. I started using DeepSeek's web version in February 2025, right after launch, as one of its first users. The servers were busy half the time back then, too many people asking.

一开始只是把它当个更高效的搜索：查资料、学知识、排查现场问题， 也问生活杂事。用着用着就深了：帮我写日报周报，把我口述的一段话 整理成周报、月报、年报，陪我整理现场报告和案例，连案例分享 PPT 的 模板都是跟它一起定的。频率也是慢慢上来的，从有事才用，到一周好几次， 再到现在基本每天都在用。2025 年我只用 DeepSeek，够用；今年才加了 ChatGPT 这些，一个模型的视角不太够。

但有件事我那时就想明白了：不敢让它碰真工单。它没有我的上下文： 不了解客户现场、客户背景，也没见过我们公司以前处理过的案例； 我喂给它的信息还隔着好几手，我要跟客户确认，客户有时候还要跟现场 确认，每过一手就丢一手。所以 2026 年 5 月我从「问它」转向「让它干活」、 开始折腾 Agent 之后，做的最重要的一件事就是把 600 多份历史工单 整理成知识库，先把上下文补上。后来顺手写了一批开源工具：自进化引擎、 Agent 审计协议、缓存代理、上下文优化器、安全门禁，都在 GitHub 上。 这些工具的共同点就一个：AI 干的每一步，我得能查它是怎么干的。

At first it was just a better search engine to me: looking things up, learning new topics, troubleshooting field issues, plus everyday stuff. It crept deeper into my work over time: drafting daily and weekly reports, turning my spoken notes into weekly, monthly, and yearly reports, organizing field reports and case studies with me, even setting the template for our case-sharing PPT together. My usage ramped the same way, from "when I needed it" to several times a week to pretty much every day. DeepSeek alone was enough through 2025; I added ChatGPT and others this year, because one model's perspective wasn't enough.

But one thing was already clear to me back then: I didn't dare let it touch a real work order. It didn't have my context. It didn't know the customer's site, the customer's background, or the cases my company had handled before. And what I fed it had passed through too many hands: I confirmed with the customer, the customer sometimes confirmed with the site, and something got lost at every hop. So when I moved from asking questions to letting it do the work in May 2026, the most important thing I built was a knowledge base out of 600+ old work orders, to fix the context problem first. Along the way I released a string of open-source tools: a self-evolving engine, an audit protocol for agents, a cache proxy, a context optimizer, a security gate. They have one thing in common: I want to be able to check how the AI did every step.

平时琢磨最多的是 prompt、context、harness 这三块工程问题。 现在的原则也简单：能用 Agent 自动化跑的事，我不打算自己动手。 有 AI Agent 方向的活儿或者想法，欢迎来聊。

Most of my thinking goes to prompt, context, and harness engineering. My rule now is simple: if an agent can run it automatically, I'm not doing it by hand. If you're building something with AI agents, I'm happy to talk.