【2026-08-16】Mole CLI started as a few hundred lines written by a pool in Sanya. The Mac app has only been out for three months. This is what I actually learned along the way, about the three kinds of junk AI leaves behind, about showing people what you are deleting before you delete it, about keeping it out of your way, and about the calls you have to make when it is just you.

【2026-06-07】In April, I built a small robot dog with an STM32, ASRPRO, ESP32-C3, servos, and 3D-printed parts. Once it could hear a command, walk a few steps, and call a cloud model, embodied AI stopped being a broad idea for me. It became serial ports, PWM, space, motion, control frequency, actuators, failed samples, and manufacturing constraints.

【2026-05-01】Friends told me AI was recommending my projects unprompted. I hadn't done anything special, so I spent an hour structuring my content for AI readability. What I did, what works, and what to skip.

【2026-04-26】A practical Claude Code guide for product, business, and operations folks. Not about prompt engineering, more about how to actually use it: how to write CLAUDE.md, how to build Skills, when to use Plan mode, and which tasks actually fit Claude Code.

【2026-04-06】Before AI, learning was slow but joyful. Now I treat it like a workflow: collect, filter, outline, draft, refine, publish. Here is how that process actually works.

【2026-04-03】Most people think model upgrades mean bigger parameters. But the real differences you feel day-to-day usually come from the back half of the training pipeline. This piece follows the full chain from pretraining through distillation and deployment, with a focus on how data engineering, system recipes, post-training, reward design, and agent training together shape the final product. The conclusion is that models getting stronger is almost always a joint effect of weights, training pipeline, and deployment decisions, not just scale.

【2026-03-30】The title is borrowed from a rewritten version of a song by Omnipotent Youth Society that I loved more than a decade ago. The scene in that song is obviously different, but the feeling of watching a familiar world slowly get replaced feels strangely similar.

【2026-03-21】After writing my deep dive on Claude Code's architecture, I realized my understanding of the underlying agent foundations was still lacking. Given our team's growing experience deploying agents in production, we desperately needed a systematic overview. I revisited the literature, open-source implementations, and my own code to comprehensively cover control flow, context engineering, tool design, memory, evaluation, and security.

【2026-03-12】This handbook distills six months of hands-on Claude Code usage across two accounts. It covers context management, Skills, Hooks, Subagents, prompt caching, and CLAUDE.md design, with a focus on making agent collaboration stable, governed, and verifiable.

【2026-03-07】Watching the frenzy around Tencent Tower installing OpenClaw today gave me a lot to think about. Many big tech companies are aggressively pushing non-technical frontline employees to install this AI tool, with some even offering 500 RMB door-to-door installation services. Everyone's desperately searching for use cases, demanding implementation, and trying to prove this thing is too important to miss. The whole process gives me a strong sense of cyber-tech folding.