从 2026 年 7 月下旬开始，我在真实项目、个人工具和日常问题中密集使用 AI。最重要的结论不是“AI 能不能写代码”，而是：如何建立一套协作流程，把 AI 的生成速度转化为可靠结果。 [!abstract] 核心结论 AI 擅长探索、生成和迭代，却不会自动理解真实目标、系统边界与完成标准。人的工作正在从“亲手写每一行代码”，转向“拆解问题、提供上下文、控制范围、验证结果和沉淀知识”。

1. 小步推进，而不是一次生成大量代码 一次生成大量代码看似高效，却容易引入不必要的抽象、隐藏逻辑和失控范围。代码越多，人越难审查，AI 也越难在后续修改中保持一致。原型可以容忍这种方式，生产代码则承担不起隐藏的复杂度。 更可靠的做法是： 明确本 …

In my current company, we recently started an exciting new project called WorkBrain. Its goal is to use customer activity data to generate new insights and help people work more efficiently. The entire project has been built with AI assistance—in fact, I have not written a single line of code myself. Instead, I provide …

After my proof of concept (POC) on creating graphs from activity signals and then reasoning over the graph data, we've seen a significant improvement in LLM query quality. Soon, an internal project will kick off, which I'm calling "Brain X." In addition to building on my POC, the new project also draws inspiration …

In recent projects, I’ve often needed to work across multiple repositories, as well as on supporting documents that are not stored in Git. This makes it important to introduce a “workbench” concept—a setup that brings together all relevant repositories and supporting artifacts. Once the workbench is in place, it …

I’ve used LLMs in development, daily operations, and product offerings within the company. They are powerful—able to interpret even vague requirements and still take meaningful action. However, the same prompt can produce different outputs each time. Sometimes this leads to pleasant surprises, but it can also introduce …

This blog outlines the step-by-step process for getting started with a Langchain demo. LLM Creation Go to the Azure portal and create an Azure OpenAI service. Within that service, navigate to the Foundry portal and deploy a cost-effective gpt-4o-minimodel (use version: 2024-07-18). Simple Chat Create chat.py 1from …

This is a MCP Getting Started, to set up the simplest mcp server and go through the simple scenario. 1mkdir my-mcp-server 2cd my-mcp-server 3 4python -m venv venv 5source venv/bin/activate 6 7pip install fastmcp 8pip install pytz Create server.py 1from fastmcp import FastMCP 2 3mcp = FastMCP("我的第一个MCP服务器") 4 5@mcp.tool …

I have been very interested in 3D modeling and visualization techniques for a long time. In 2017, I started a GitHub project to build a simple modeling tool using raw WebGL directly. I decided against using frameworks like Three.js, mainly for educational purposes. The project progressed slowly. After struggling for …

Recently, a new trend has emerged in utilizing agent skills, as seen at: https://agentskills.io/home. It appears to be quite straightforward to assign specific tasks to an agent, provided that your requirements are clearly outlined in the skill.md file. I decided to give it a try. As an avid book lover, I occasionally …

I recently used vibe coding to complete two prototypes for my side projects. The experience was fantastic! General Experience I primarily used Visual Studio Code’s Agent mode with the Claude Sonnet 4.5 model. I would describe my requirements to the Agent, and it would write the code for me. I was impressed by how much …