记录我如何把模糊的想法，变成可以使用、理解和继续迭代的东西。Notes on turning fuzzy ideas into things people can use, understand, and keep improving.
部分文章目前仅提供中文正文。Some article bodies are currently available only in Chinese.
在跑通 H3 之后，继续拆解官方提示词规范和八类内容 Skill，用简易案例说明适用场景、输入模式与减少无效生成的方法。After the first successful run, examine the official prompt rules and eight content skills through simple examples, input modes, and ways to reduce wasted generations.
面向新手说明如何检查硬件、选择本地或云端环境，并在 ComfyUI 中从官方 MiniMax H3 模板开始生成第一条带声音的视频。A beginner guide to checking hardware, choosing local or cloud execution, and generating a first video with audio from the official MiniMax H3 templates in ComfyUI.
AI 人物的真实感不只来自毛孔与瑕疵，更取决于人物、场景、镜头、动作和生活痕迹是否共同成立。Believable digital humans depend on more than pores and imperfections; the person, setting, camera, motion, and traces of lived experience must agree.
面向零基础梳理 FDE 的三条入行路径、能力清单、面试准备与真实工作体验，帮助判断这个岗位是否适合自己。A beginner roadmap to forward deployed engineering covering three entry paths, the core skill set, interview preparation, and the realities of the role.
FDE 是真正推动 AI 落地的前线工程师，还是换名后的驻场外包？通过岗位职责、授权范围与交付方式识别两者差异。Is a forward deployed engineer truly responsible for putting AI into use, or is the role merely renamed on-site outsourcing? Distinguish them through scope, authority, and delivery.
从成本、错误风险、客户体验、部署投入和团队培训五个常见顾虑出发，解释中小企业与社区运营该怎样务实引入 AI 客服。A practical response to five common concerns about AI customer service: hidden cost, errors, customer experience, deployment effort, and team training.
从 Loop Engineering 走向 Graph Engineering：把复杂任务拆成可并行、可汇合、可检查的多 Agent 图结构，并附可直接使用的分工模板。Move from loops to graphs by splitting complex work into parallel, mergeable, and reviewable agent tasks, with a reusable delegation template.
从头像、照片、品牌吉祥物或草图出发，设计一款可识别的个人 IP，再把它制作成陪伴自己编程的 Codex 宠物。Start with a portrait, photo, mascot, or sketch, shape it into a recognizable personal character, and turn it into a Codex pet that accompanies your coding sessions.
知识库让 Agent 理解过去，Canvas 让人与 Agent 共同面对复杂任务的现在：目标、资料、判断与待办可以在一张共享画布上并行组织。A knowledge base helps an agent understand the past; Canvas gives people and agents a shared surface for organizing goals, evidence, decisions, and tasks in the present.
Boris Cherny 说"我不再 prompt Claude 了。我让一堆循环跑着，由它们去提示 Claude、去琢磨该干什么。我的工作变成了写循环。"Boris Cherny said, “I no longer prompt Claude. I let a set of loops run, prompting Claude and figuring out what to do. My work became writing loops.”
外包一个几十秒的产品视频要上万块，但我没打开任何剪辑软件，只用代码就把它跑出来了。工作流的关键在于：先用 GPT-Image-2 把每个场景的视觉预演图出好，再把图交给 Codex 复刻成代码。A few dozen seconds of product video can cost tens of thousands, but I did not open an editing app. I built it with code: first create visual previsualizations for every scene with GPT-Image-2, then hand them to Codex to reproduce in code.
一人公司听起来很爽和自由。没有考勤，没有日报，没有老板在背后走来走去。但另一面是：也没有人替你盯着你自己。Running a one-person company sounds great and free: no time clock, daily reports, or boss pacing behind you. But there is another side: no one else is watching you either.
AI 明明就在收藏夹里，但根本没想到用。这才是很多人一直“学了很多，却始终用不上”的真正原因。 不是因为笨，也不是因为教程没用，而是因为你还没有养成一个最基础、但也最重要的习惯：遇到任何问题时，先想到一句，这件事，AI 能不能帮我做？AI is already sitting in your bookmarks, yet you never think to use it. That is why many people keep learning without ever applying what they learn. It is not because they are incapable or the tutorials are useless; they have not built the basic habit of asking, whenever a problem appears, whether AI can help.
装好 Claude Code 之后，下一步该怎么办？CC 本身是空白的：它知道怎么帮你，但它不知道你是谁、在做什么、你的工作有什么规律。从"装好了"到"真正进入生产力"，中间要做一件事：What comes after installing Claude Code? CC starts as a blank slate: it knows how to help, but not who you are, what you do, or the patterns in your work. Moving from “installed” to real productivity requires one thing in between:
2025 年，AI发展的速度超乎人的想象。如果说23，24年我们还在以一种娱乐态度看AI“它能不能写诗”，“能不能画明星吃面条的图片”，而发展到 2025 年，AI正式进入了“价值生成”的深水区。泡沫在左，机遇在右。AI moved faster in 2025 than most people imagined. In 2023 and 2024, we treated questions like whether AI could write poems or draw celebrities eating noodles as entertainment. In 2025, AI entered the deep waters of value creation. The bubble is on the left; opportunity is on the right.
我在幻想着，如果这套Agent跑通了，就意味着我可以在睡觉时，让AI帮我自动通过“搬运+写文”来赚取流量收益，这不妥妥的“睡后收入”吗。I imagine that if this agent workflow worked, I could let AI automatically repurpose and write content while I sleep to earn traffic revenue—the ultimate “sleep income.”
我们探讨了AI的基础概念、如何编写Prompt，还探讨了AI作为“模拟器”的本质。知道了这些固然不错，但这些难道只是为了跟AI聊天吗？AI对于我们来说就只是个聪明点的聊天助手吗？We explore AI’s basic concepts, how to write prompts, and the nature of AI as a simulator. These ideas are useful, but are they only for chatting with AI? Is AI merely a slightly smarter chat assistant?
AI根本不是一个人，也不是某种“实体”，请停止把它当成一个人类来聊天。AI is not a person or an entity. Stop treating it like a human when you chat with it.
一会儿是LLM，一会儿是AIGC，还没搞懂NLP是啥，那边又在喊AGI要来了。大家都在焦虑“被AI取代”，但对大多数人来说，第一步的门槛其实是先明确概念。One day it is LLM, the next it is AIGC, and you still have not figured out NLP before everyone starts shouting about AGI. While people worry about being replaced by AI, the first hurdle for most is simply getting the concepts straight.
