An engineering notebook on AI, systems, and everyday experiments.

为了解决车库门忘关的问题，我用拇指大小的ESP32-CAM单片机与AI协作做了一个端侧神经网络传感器。本文记录了如何通过打通全自动串口研发闭环、解决数据极度不对称与QAT量化衰减，在10小时内跑通嵌入式AI的完整过程。

To solve the problem of accidentally leaving my garage door open overnight, I collaborated with AI to build an edge neural network sensor on a thumb-sized ESP32-CAM. This article details how we completed the end-to-end telemetry loop, tackled extreme data imbalance and QAT quantization loss, and got edge AI running on a microcontroller in just 10 hours.

月食时月面边缘有一条青绿色的窄带，科普说那是臭氧吸收。但为什么是窄带不是整圈？为什么全食最深时反而看不见它？我们从最土的白圆盘开始，一层层加物理，硬算出这条带，在一路翻车里发现了一个更大的问题：AI 太懂物理，反而会把你带进前人留下的近似里。

During a lunar eclipse, a narrow green-blue band appears at the moon's edge; popular science says it is ozone absorption. But why a narrow band instead of a full ring? Why does it vanish at deepest totality? We start from the simplest white disk, layer in physics, and compute the band — and through repeated failures discover a bigger problem: AI knows physics too well, and will happily lead you into approximations left behind by earlier researchers.

我用AI提效很成功，产出和rating都是org最高之一，但升职两次都失败了。后来发现一个讽刺的陷阱：正因为手快好用，老板把你当手而非脑，项目零散多变，反而讲不清一年的成果。最擅长用AI的人，反而最容易被AI替代。破解之道是主动设计奖赏系统，把省下的时间用来做判断而非交付更多。

AI made me a top performer. I was denied promotion twice. Speed makes bosses treat you as a hand, not a brain. The best AI users paradoxically become the most replaceable. The fix: design the incentive structure.

从最直白的"一星一像素"出发，八次翻车、六亿颗星，一步一步把银河从真实星表里逼出来。在这个过程中才发现，以前从来没认真想过头顶的星空为什么长这个样子。

Starting from the simplest "one star, one pixel" rendering, eight failures and 600 million stars later, the Milky Way slowly emerged from a real star catalog.

用好AI的第二步不是更会写 prompt，而是先外化、再复用。本文讲清 Skill 如何承载工作知识、好 Skill 的三要素，以及如何组织 Skill 文件夹让 Agent 自动找到。

Step two isn't better prompting. It's externalize first, reuse second. This post explains how Skills carry work knowledge, the three parts of a good Skill, and how to organize them so agents find the right one.