和朋友争论 Mio 该走疗愈树洞还是恋爱模拟,争到最后发现两条路都不对。市场数据说这个品类的用户和角色平均只处 5 到 7 天,差评第一名是失忆。我觉得毛病出在关系是单方面的。这一篇讲 Mio 的新方向,双向奔赴,附一份完整的产品主线 deck。

AI 陪伴留不住人，毛病出在关系是单方面的

给盘盘猫做了一次从灵魂到呈现的完整体检，结论挺离谱：最值钱的三样东西——记忆、人格、统一体验——代码全在仓库里，没有一样真的跑在用户面前。这一篇讲免费大模型怎么把「生辰进、报告出」打成白菜价，参天占住的工具位旁边那个没人站的关系位，还有为什么最后选了灵魂先行这条路。附一份完整的产品主线 deck。

盘盘猫的记忆和人格建好了，一直没上线

上一篇让几个 claude 账号自己换着用，人还是拴在电脑前。这一篇把 agent 全挪到一台常开的机器上，手机随时连上去看进度、派新活，笔记本合盖它也接着跑。三个开源件各管一段——herdr 管 session、hcom 管通讯、Moshi 管手机——外加账号自动轮换兜底，凑成我目前跑得最顺的移动端 vibe coding 方案。这篇讲怎么搭，也讲踩过的坑。

我把 coding agent 挪到常开机器上，手机随时指挥

Last time I taught a few Claude accounts to rotate themselves — but I was still chained to my laptop. This time I cut the last cord: the agents live on an always-on box, and my phone connects in to watch and steer them. Close the lid, leave the house, lie in bed — they keep running. Three open-source pieces each own one layer — herdr the sessions, hcom the messaging, Moshi the phone — plus account auto-rotation as the safety net. The mobile vibe-coding setup that currently runs smoothest for me: how to build it, and the traps I hit.

I Moved My Coding Agents Into My Phone — and Close the Lid

天天用 claude code 的人都撞过五小时用量墙，切号还要重新登录，正在跑的 session 也会丢。 我搞了两个工具：clauth fork 在活跃 session 底下做无感热切换和无人值守自动轮换， 菜单栏的 ccsbar 把每个账号的余量、预报和健康状况摆在明面上。这篇讲这套东西是怎么搭起来的。

我让我的几个 claude 账号自己换着用了

I run several Claude accounts and kept hitting the 5-hour wall mid-session. Switching meant logging out. So I built a daemon that hot-swaps the active login under a running session — and the whole thing turned on one non-obvious fact about how macOS stores an auth token.

I Taught My Claude Accounts to Rotate Themselves

模型越强，你说不清楚想要什么的代价越大。从地图与疆域讲到四种空白，拆一套我自己验证过的和模型打交道的方法，最后讲怎么让每个坑只交一次学费。

模型越强，越得先把要什么说清楚

AI 把讲授的边际成本打到零，教育史上第一次，瓶颈从'谁来教'搬到了'人为什么想学'。顺着一个 AI 社群里关于'孩子为什么不肯学'的讨论，我把动机科学、学习科学和 alignment 的证据串了一遍：需求不是找出来的，是点着的；'以后用得上'输给的是一条折现曲线；动机满格的学生加一个直接给答案的 AI，考试成绩反而比不用的还低 17%。教育要解的是双重对齐——先让人想学，再保住那份不可外包的认知劳动。

教育的瓶颈从谁来教搬到了想不想学

AI has driven the marginal cost of instruction to zero, and for the first time in the history of education the bottleneck has moved from 'who will teach' to 'why would anyone want to learn.' Starting from a community thread about why kids refuse to study, I traced the evidence across motivation science, learning science, and AI alignment: demand isn't found, it's ignited; 'you'll need it someday' loses to a discount curve; and a fully motivated student with an answer-giving AI scores 17% worse on exams than one with no AI at all. Education in the AI era is a double alignment problem — first get people to want to learn, then protect the cognitive labor that can't be outsourced.

The Best Teacher Is Now Free. Learning Didn't Explode.

课代表立正这期访谈，嘉宾是屠龙（杨滢）——清华生物、匹兹堡脑科学博士、给美军解码大脑，回国后却在一个"会计开法拉利、网红带黑社会大哥上门"的江湖里连开四家公司。标题问"为什么她干啥啥赚钱"，但她自己交代过卖首饰一件没卖出去。真正值钱的不是某种点石成金的天赋，是一套能从科研搬到美妆、图书、课程的操作系统。我顺着完整字幕又抠了一层：她到底带了什么下场。

屠龙博士那套方法，搬到哪个行业赢面都高

Reactions to an interview with Tulong (Yang Ying) — Tsinghua biology, a Pittsburgh PhD decoding the brain for the US military, who came home and built four companies in a world of "accountants driving Ferraris and influencers showing up with gangster bosses." The title asks why she makes money at everything; she's the one who admits the jewelry line sold zero. What's actually valuable isn't a Midas touch — it's an operating system she can carry from neuroscience into skincare, books, and courses.

"She Makes Money at Everything" Is Packaging. Her Real Edge Is Bringing the Model to the Battlefield

这是《前三次技术革命，红利最后都落到普通人头上》背后的完整框架。正文是给人读的一篇文章，把判断讲清楚就够了；这篇是给框架本身的——把'技术怎么洗牌财富'拆成第一性原理、一个财富捕获公式、六个传导机制、AI 的四层结构、一套九问研究法和几条可验证假说，方便你拿去套到下一个技术上，自己跑一遍。

财富洗牌这套框架，拿去自己跑一遍（研究附录）

The full framework behind Railroads, Electricity, the Internet Each Built a Middle Class. AI Might Not. The essay is the readable version; this is the machine — how technology reshuffles wealth, broken into a first principle, a wealth-capture formula, six transmission mechanisms, AI's four layers, a nine-question research method, and a few falsifiable hypotheses — so you can run it on the next technology yourself.

The Tech–Network–Distribution Model (Appendix to The Wealth Reshuffle)

听《十分吸引》串台《听懂涨声》聊 AI 财富再分配，最有意思的是他们把铁路、电力、互联网三轮洗牌的机制层层拆解。每一次生产力基础设施的革命都是一次财富大洗牌：技术重写经济网络，旧瓶颈失效，新瓶颈拿到定价权。前三轮都是先集中、后扩散，靠的是新生产方式需要海量的人；AI 是重资产、离散型的，可能只走完集中那一半。这是我对这轮最不乐观的判断。

AI 的红利可能不会再落到普通人头上

I listened to an episode on AI and wealth redistribution — the finance podcast 十分吸引 crossed with 听懂涨声 — and the best part was the history: how railroads, electricity, and the internet each reshuffled wealth. I dug a layer under their account and made it harder: the reshuffle runs one logic every time — technology rewires the economic network, the old bottleneck fails, a new one gets priced, and wealth follows the bargaining power. Railroads rewired the goods network, electricity the energy network, the internet the information network, AI the task network. The first three eventually grew a new middle class, because those production modes needed armies of ordinary people. AI is heavy-asset and dispersive, so it may run only the concentration half and skip broaden-the-middle.

Railroads, Electricity, the Internet Each Built a Middle Class. AI Might Not.

Agent 干错事，大多数人的反射是往 CLAUDE.md 加一句话。它错在两个方向：指令文件是 context 不是强制，规则在压力下根本不绑得住；而文件越长，越污染那个让 agent 可靠的 context 本身。承重的规则该进 hook、lint、CI，散文应该越写越短。

一份 CLAUDE.md 管不住 agent

The reflex when an agent misbehaves is to add a sentence to CLAUDE.md. It's wrong in two directions: instruction files are context, not enforcement, so the rule won't bind under pressure — and the longer the file gets, the more it pollutes the context that makes the agent reliable. Load-bearing rules belong in hooks, lint, and CI. The prose should get shorter.

Your AGENTS.md Is Not a Fence

Agent 能完美遵守一条它看得见的契约，也能毫无知觉地破坏一条它看不见的。能熬过 agent 时代的边界，是那些写进仓库、有类型、有测试、能被 import 的；会被弄坏的，是那些只活在团队记忆里的裸字符串约定。把每条边界都变成契约，是手里杠杆最高的一招。

agent 看不见没写下来的约定

An agent is excellent at honoring a contract it can see and terrible at preserving one it can't. The boundaries that survive the agentic era are the ones written into the repo as typed, tested, importable things — and the ones that break are the stringly-typed promises that lived only in a team's shared memory.

Contracts, Not Tribal Knowledge

Agent 最危险的产出不是烂代码，是"我搞定了"这几个字。纯核心、一条命令的验证、一个独立的判官、一份证据账本，全部加起来就为了一件事：把"做完"的裁决权从作者手里拿走，让它变成被证明出来的，而不是被感觉出来的。这是前三篇一直在铺垫的底座。

agent 自己说做完了，不算数

An agent's most dangerous output isn't bad code — it's the word "done." The whole point of a pure core, a one-command verify, a separate verifier, and an evidence ledger is to take the verdict out of the author's hands and make "done" a thing that's proven rather than felt. This is the substrate the rest of the series was building toward.

Let the Agent Prove It's Done

这几个月改动基本全是 agent 写的：人出判断和认知，执行层全是 AI。代码的读者换人之后，仓库就变成了 agent 要对着编程的接口——tribal knowledge 记录下来就能像软件一样直接分发，目录是导航 API，契约是约束 API，验证是校验 API。

repo 得当成 agent 的 API 来设计

The maintainer of your repository quietly changed from a human who remembers to an agent that re-derives everything from the files each session. Once you accept that, the repository stops being documentation and becomes an interface. Directory layout is the navigation API, contracts are the constraint API, and verification is the validation API.

Your Codebase Is an API for Agents Now

朋友给我一份用拉丁学名的 omakase 认鱼课，可板前点菜时，菜单写的是 Hamachi、是 Sea Bream。我想要一个标美式菜单常用名的版本，于是指挥一队 agent 搭了出来：九个并行调研 agent 上网核实菜单名、gpt-image-2 出约 150 张写实图、一段确定性代码拼成 PDF。73 种鱼、92 页，外带两个踩坑故事和一份免费下载。

在板前叫不出鱼名，我让一群 agent 做了本认鱼图鉴

I wanted a sushi-counter field guide that used the names you actually order by in a US restaurant — not Latin binomials. So I pointed a team of agents at it: parallel research workflows verifying English menu names with web search, gpt-image-2 for ~150 photoreal neta shots, and a deterministic HTML→PDF assembler. 73 fish, 92 pages, two debugging war stories, and a free download.

How I Built a 73-Neta Omakase Field Guide in an Afternoon

一个朋友半夜问我，要不要也做个工具去跟 X 上很火的 Serenity 买进卖出。我的第一反应是没意义：一个策略真能稳定赚钱，最理性的做法是闷声加仓，不会拿出来开直播。但聊到后面我改了主意。值得蒸馏的东西确实有，藏在他提问的顺序里，跟他的持仓没关系。我们把这套研究流程做成了一个开源工具。

公开出来的 alpha，等你跟基本就没了

A friend pinged me at midnight asking whether we should build a bot to follow Serenity, the supply-chain researcher who blew up on X, in and out of his trades. My first reaction was that it was pointless: a strategy that actually compounds gets quietly levered up, not broadcast. But by the end of the night I'd changed my mind. There is something worth distilling here — it just isn't in his holdings, it's in the order he asks his questions. So we turned that process into an open-source skill.

Serenity's Alpha Died the Day It Went Public

群里聊升职，起点是鸭哥那篇文章：用 AI 把效率拉爆、成了模范老黄牛，就能加薪升职吗？现实恰恰相反——你越快、越好用，越容易被组织重新定位成一个超高吞吐的执行节点。我把升职拆成一个乘法模型：能力 × 叙事 × sponsor 动机 × 组织坑位 × 政治/时机，任何一项接近 0，整体都垮。AI 主要增强的是能力和速度，叙事和定位反而容易被它带坏——你越快，别人越懒得理解你的判断，越习惯把你当工具用。

能干的人反而更难升上去

A group chat got into promotions, kicked off by a friend's essay: if you use AI to crank your output 10x and become the model workhorse, do you get the raise and the title? The reality is the opposite — the faster and more useful you are, the easier it is for the org to recast you as a high-throughput execution node. I break promotion into a multiplicative model: capability × narrative × sponsor motive × org slot × politics/timing, where any factor near zero collapses the whole case. AI mainly boosts capability and speed, but can quietly damage your narrative and your positioning. The danger isn't being too slow. It's being so fast nobody bothers to understand your judgment.

Too Fast to Promote

群里聊到一个很沉重的问题：AI + 具身智能把活都干了、生产力大涨之后，普通人会因此更自由吗？还是说红利全被大公司、大平台、大资本拿走，普通人只是从打工人变成被系统管理的人？我越想越觉得，未来不会自动走向乌托邦，而更可能是三条路线混在一起——AI 红利社会、平台封建主义、高自动化治理社会。决定 AI 时代是人被解放，还是生产力被解放了、人没有。

AI 和机器人之后，普通人会更自由吗

A group chat landed on a heavy question: once AI and embodied robots do most of the work and productivity explodes, do ordinary people actually get freer? Or do the gains flow to a few platforms, capital pools, and states, while everyone else goes from worker to managed case? The more I sat with it, the less I believed in an automatic utopia. The likely future is three paths braided together — AI dividend society, platform feudalism, and high-automation governance — and which one dominates decides whether AI frees people, or just frees productivity while people stay stuck.

Productivity Will Be Liberated. People Won't.

从爱迪生那根烧红的灯丝，到能指挥电子的开关——电子管、PN 结、晶体管，一部人类驯电简史。

人类是怎么管住电子的

From the stray current Edison found in a light bulb to a switch that steers electrons on command — vacuum tube, PN junction, transistor, the short history of taming electricity.

How Humans Learned to Boss Electrons Around

一个电路怎么记住一件事情？从两个非门互相提醒，到寄存器，再到决定游戏性能的缓存，底层其实是同一个朴素的问题。

CPU 是怎么算数的

A circuit that can remember a single 0 or 1, grown step by step into a register — and why cache, not clock speed, is what makes a gaming CPU fast.

How a CPU Actually Does Math

机械硬盘用磁存、SSD 用电子存、RAID 用多块盘换可靠——以及那个吓人的「RAID 5 重建失败率 99%」到底是真是假。

硬盘是怎么存数据的

Hard drives store with magnetism, SSDs trap electrons in an insulator, RAID trades extra disks for reliability — plus the truth behind that scary "RAID 5 rebuild fails 99% of the time."

Where Your Data Actually Lives — and Why RAID 5 Won't Save You

内存和显存本是同根生，却走上了两条相反的路——一个死磕延迟，一个死磕带宽，到 HBM 干脆把存储架到了 GPU 头顶。

内存和显存是怎么分家的

Main memory and video memory came from the same place and then split — one obsesses over latency, the other over bandwidth, and HBM ends up bolting memory onto the GPU's roof.

RAM and VRAM Both Store Data. Here's Why They're Nothing Alike.

光既是造芯片的刻刀（光刻），也是造屏幕的画笔（量子点、AR 光波导）——人类在硅片方寸之间的驯光记。

人类是怎么管住光的

Light is both the chisel that carves chips (lithography) and the paint that makes screens (quantum dots, AR waveguides) — how humans learned to tame it in the space of a silicon chip.

Taming Light Inside a Chip

转录不难，难的是「谁在说」。为了搞定说话人分离，我把 Gemini 切块、Senko、豆包 auc 挨个试了一遍，最后落在火山妙记上。5 段真实录音，每段两人对话它都精准判 2 人。顺带踩了火山两套鉴权、跨境上传被限速到 34KB/s 这些坑。

说话人分离，我试了四套方案才搞定

Transcription is easy. The hard half is knowing who spoke. Chasing accurate speaker diarization, I ran Gemini chunking, Senko, and Doubao's auc models into the ground before landing on Volcano's 妙记 (Lark Minutes ASR) — which nailed the speaker count on every two-person recording. Plus the gotchas: Volcano's two auth systems and a cross-border upload throttled to 34 KB/s.

My Transcriber Heard Five People in a Two-Person Conversation

从一篇「AI 写代码该选什么语言」的讨论说开去。当 AI 既写代码又当审稿人，选语言的判据不是它对 AI 多熟、也不只是循环多快，而是一次绿灯替你排除了多少错——而最大的陷阱是，那盏灯要是 AI 自己点的，它什么都没排除。

没人审 AI 写的代码之后，语言该怎么选

存算一体、光互连、异质集成、新型存储介质。当 HBM 也堆不动的时候，计算架构会怎么演化。

存储墙修不动了怎么办

When a wall gets this expensive, the industry attacks it from every side — tier through NAND, pool over CXL, ditch HBM, compute inside memory. None of them breaks the wall in 2026. Which tells you exactly where this cycle's turn is not going to come from.

Beyond the Memory Wall: The Workarounds, the Challengers, and the Turn

硅谷101 这期 AI-First 播客的听后感。多数人把 AI 当工具，跟 24 小时较劲，提速到头也就 10 倍；少数人把它当系统，重构整条工作流，人往后退到只做两件事——给意图、审产出。但我想说的是，审产出这道关，正在被自己淹没。

把 AI 当工具用，瓶颈永远是发起的人

Reactions to a podcast on AI-first org architecture. Use AI as a tool and you're racing the clock — 10x at best. Use it as a system and you rebuild the whole workflow, retreating to just two jobs: give intent, review output. But the review gate is drowning, and that's the part nobody wants to say out loud.

Tool or System: Why Most "AI-First" Tops Out at 10x

一个被 99% 的人忽视的事实：训练不是瓶颈，推理才是。推理需要内存，海量内存。而这个需求正在创造一个持续十年以上的半导体超级周期。

HBM 这波涨价其实是推理带起来的

Memory is the most cyclical business in tech — 40 years of up-10x, down-90%. This run has three genuinely structural differences. But 'this time is different' is the most expensive sentence in finance, so the bear case gets equal weight.

The Memory Supercycle: Why This Time Might Actually Be Different

我用 TypeScript 写过几个 web 游戏，但总觉得天花板太低。想认真做一个、让 AI 写大部分代码，到底该选 Godot、Unity 还是虚幻？查了一圈才明白：AI 最擅长的引擎，不是训练数据最多的那个，而是它能整个读进眼里、自己跑起来、自己看着改的那个。

想让 AI 帮我做游戏，查了一圈最后选了 Godot

I built a few TypeScript web games but always felt the ceiling was too low. If I wanted to build a proper game and let AI write most of the code, should I pick Godot, Unity, or Unreal? After a lot of digging the answer surprised me: the engine AI is best at isn't the one with the most training data — it's the one AI can read entirely as text, run headless, watch, and fix on its own.

The Best Game Engine for AI Is the One It Can Read

一张 H100 超过一半的成本是 HBM。它凭什么这么贵？TSV、硅中介层、垂直堆叠——以及为什么只有 SK 海力士能赚钱。

HBM 凭什么这么贵

HBM isn't faster DRAM. It's DRAM rotated 90 degrees — stacked, drilled through with thousands of vertical wires, and bolted next to the GPU on a silicon interposer. That manufacturing nightmare is exactly why three memory makers just crossed a trillion dollars.

How HBM Works — the Memory That Decides Who Wins AI

The code for my games? Agents handle it, faster and more correctly than I would. The thing that stopped me cold was the story. I'm not a writer, so I did the only thing I know how to do — turned "how do you write a game story" into a research problem, tore the craft apart, and built a skill that walks you from a blurry idea to a scene you can actually stage. This is the process, and the genuinely useful craft I dug up. AX Skill Workshop, part 2.

AI Wrote My Game's Code. It Still Can't Write One Line of Dialogue That Isn't Cringe.

一具坐在金色王座上的尸体，已经死了一万年。把它送上王座的，是它最宠爱、最信任的那个儿子。这是战锤40K 最像悲剧的那段往事——也是这一万年黑暗真正的源头。

荷鲁斯之乱是怎么回事

你的 SSD 为什么断电还能存数据？因为 NAND 用物理方法把电子锁在绝缘层——这个简单的原理构成了全球每年数百亿美元的存储市场。

SSD 凭什么断电了数据还在

The same chip lives in your USB stick, your phone, and the petabyte SSDs feeding AI clusters. It got cheap by trading endurance for density — and that exact tradeoff is why AI is now dragging it out of cold storage and onto the hot path.

How NAND Flash Works — From Your SSD to the AI Data Center

Ten thousand years before the grimdark present, the Emperor of Mankind's favorite son turned a galaxy-spanning crusade into a civil war that has never truly ended. This is the tragedy at the heart of Warhammer 40,000 — and why its wound still bleeds.

The Horus Heresy: How a Father's Greatest Son Broke the Galaxy

一个电容、一个晶体管。过去五十年几乎所有计算的内存都建立在这对搭档上。

DRAM 是怎么存下一个比特的

DRAM is a leaky bucket you refill thousands of times a second. That physical fragility is why it's fast, why it's volatile, and why it just hit a scaling wall at the exact moment AI demand exploded — sending RAM prices up triple digits.

How DRAM Works — and Why It Got Scarce

从六千万年前神族互砍，到第 41 千年还在打的圣战，这条时间线一篇捋完，顺带讲讲为什么设定里连日期本身都不可靠。

40K 六千万年的时间线，从头捋一遍

From two dead god-races fighting over the galaxy 60 million years ago to a resurrected demigod waging holy war in the present — the full history of the grimdark future, in order, with the dates that lie to you on purpose.

60 Million Years of War: The Complete Timeline of Warhammer 40,000

处理器的速度是指数增长的，内存带宽是线性增长的。这个裂缝是过去二十年计算架构最根本的瓶颈——而你每天都在感受它。

电脑卡，多半是处理器在等内存

I finance the machines AI runs on, and the single priciest, scarcest part inside the box isn't the logic die — it's the memory stacked next to it. This is why: inside a modern accelerator, raw FLOPS stopped being the binding constraint. Memory bandwidth did.

The Memory Wall

战锤40K 的银河里没有好人，也等不来结尾那份和平条约。这篇讲清楚这场仗为什么从结构上就停不下来，以及为什么这恰恰是它的全部用意。

战锤40K的战争为什么永远打不完

In the 41st millennium there are no good guys, no rescue, and no peace treaty waiting at the end. Here is why the galaxy of Warhammer 40,000 is structurally incapable of stopping the war, and why that is the entire point.

Why the War Never Ends: The Core Conflicts of Warhammer 40,000

一个垂死的神权帝国，四位用情感凝成的噩梦邪神，外加一打坚信自己才是主角的外星种族。这篇把战锤40K 银河里所有出场的狠角色，一个个介绍给你。

战锤40K里都有哪几伙人

A galaxy at total war needs combatants worth fearing. Here is the cast — a dying god-empire, four nightmare powers wearing reality like a glove, and a dozen alien races that all believe they are the protagonists.

Who Fights in the 41st Millennium? A Field Guide to the Factions of Warhammer 40,000

一百万个世界，一位已经死了一万年的神。在第41个千年，你能说出的最乐观的话，就是人类还活着——而这恰恰是最糟的部分。

战锤40K是怎么回事

A galaxy of a million worlds, ruled by a god-emperor who has been dying for ten thousand years. Welcome to the 41st millennium, where the most optimistic thing you can say is that humanity has survived — and that may be the worst part.

The Grimdark Galaxy: An Introduction to Warhammer 40,000

2015 年，Games Workshop 干了一件几乎没人敢干的事：把自己养了三十年的奇幻世界真的玩死了——星球裂成两半，神死的死逃的逃。这是旧世界怎么塌的，也是废墟里长出了什么。

GW 为什么把旧世界写死了

In 2015, Games Workshop did something almost no franchise dares to do: it ended its 30-year-old fantasy world for real — cracked the planet in half and let it die. Here is how the Old World fell, and what rose from its ashes.

The Day a Company Murdered Its Own World: Warhammer's End Times

帮非职业创作者从零打造游戏剧本的 AI skill。世界观、角色、主线支线、场景脚本——全流程陪伴式创作。

帮不会写故事的人写出一个游戏剧本

战锤奇幻的旧世界，南北极是两个永远堵不上的伤口，每座城市的地板下都住着鼠人。它打的不是一场仗，是六场。这篇讲清楚它为什么注定输——以及为什么这反而是它最动人的地方。

战锤旧世界的仗为什么打不赢

The Old World was never going to win. From daemon-haunted poles to a rat empire under every floor, here is why Warhammer Fantasy is a story about civilizations fighting entropy with one hand tied behind their back.

A World Under Siege: How the Old World Lost Its War

矮人把每一笔旧仇都记进一本永远翻不到底的书，鼠人多到能淹没整个地表却忙着自相残杀，精灵因为一个没坐上的王座分裂成两族打了几千年。带你认全这个垂死世界里的所有玩家——以及让他们到死都不肯联手的那些仇。

战锤旧世界里都有哪几伙人

Dwarfs who write their grudges in a book that never ends. Rats who deny their own existence. Elves split by a single denied throne. A tour of the powers fighting over a dying fantasy world — and the hatreds that keep them from ever uniting against the thing that will kill them all.

Order and Destruction: A Field Guide to the Factions of the Old World

一个从来没存在过的中世纪欧洲，八股看不见的魔法之风，世界尽头天上的一个洞。这是旧世界——Games Workshop 造了三十多年的奇幻设定，2015 年自己把它毁掉了。

战锤旧世界是怎么回事

A dark-medieval Europe that never was, eight invisible winds of magic, and a hole in the sky at the top of the world. This is the Old World — the fantasy setting Games Workshop built over thirty years, then deliberately blew up in 2015.

The World That Was: An Introduction to Warhammer Fantasy

把半仙占卜做成 skill。梅花易数、小六壬、六爻——用户描述事情，选卦法，出真卦象，半仙化解读。一个 skill，三种东方占卜体系。

我真的做了个赛博半仙

Six months of accumulated Claude Code configuration — 50 skills, 50 subagents, 40 commands, hooks, rules, and the sync pipeline that pushes all of it to a public GitHub mirror with a fail-closed two-tier audit. What's inside, why I built it, and the workflow philosophy behind it.

I Open-Sourced My Entire ~/.claude

AX-skills is the public mirror of my personal Claude Code skills directory — eleven workflows for writing, research, image generation, knowledge sync, and one cyber half-immortal divination skill. After a year of writing them, I think a skill isn't a stored prompt — it's how you encode care.

I Open-Sourced the 11 AI Skills I Use Every Week

Last week I surveyed 8 open-source meeting recorders. My friend tried every one. Verdict: ASR is bad, summaries are bad, UI/UX is bad. So I dug deeper. The real problem is structural — every project I recommended is built on Whisper, and Whisper has been outpaced by 2-4× on Chinese benchmarks for over a year. This is the deeper survey: data, the issues that maintainers themselves admit, the dumb-but-effective hack everyone overlooked, and four projects I missed the first time.

I Recommended Eight Open-Source Meeting Recorders. My Friend Returned All of Them.

A friend asked for a meeting recorder. Zoom/Teams native AI is unusable. Granola and Krisp work but cost ~$200/yr per seat. He's on Google Workspace but doesn't use Meet, so the bundled Gemini summarizer is dead weight. I spent a night on GitHub and found a whole constellation — from Meetily at 11.6k stars down to single-digit-star personal toys, all racing to commoditize the same three Lego pieces.

Zoom's Native AI Sucks. So I Surveyed Every Open-Source Meeting Recorder

Took the three Eastern divination engines from PanPanMao and packaged them into a single agent-portable skill — algorithm engine + knowledge base + persona prompt, three folders deep. The pattern transfers to any product with real domain knowledge.

Packaging a Product into a Skill: A Cyber-Banxian for an Agent Contest

Part 2 的「6 槽位通吃 7 品类」是一层漂亮的抽象。跑了两周，测出牛仔外套生成「宝石星图」的 bug。这一篇讲抽象失守在哪一层、为什么没用户时分阶段上线是假安全感、以及把服饰和美妆各自推倒重来的过程。

一套 prompt 通吃三个品类，是我想多了

Part 2 claimed six slots could cover seven categories. Two weeks later, a denim jacket generated a gem constellation. This post walks through where the abstraction failed, why skipping the 'safe' phased rollout was the right call, and how each of three categories got its own six-template set.

Walking Back the Abstraction

Solo dev 做设计，大多数时候 N=1——第一直觉就是唯一方案，因为做第二个版本要花一天。Claude Design 把这事拆了：一下午六个完全不同的视觉语言摆在面前。留下了什么、丢了什么、方向锁死之后 big-task 的 ui profile 接手怎么走——这篇讲这些。

我让 claude 一下午出了六个设计方向

For solo dev and small teams, design exploration is usually N=1 — you go with your first instinct because producing a second variant takes a day. Claude Design collapses that: six distinct aesthetic directions in an afternoon. What I kept from the pipeline, what I dropped, and how `big-task`'s UI profile takes over once a direction is locked.

Six Directions in an Afternoon: Using Claude Design to Kill First-Instinct Bias

先是裸跑 Claude Code。再是 GSD 多 agent 仪式。再是 superpowers 的 skill 库。最后在上面包了一个自己的 orchestrator。每一版崩盘都来自一个具体 session。最后我才想明白：solo dev 的 harness 跟团队的不是一个物种，而且真正的答案不是选一个默认引擎——是按工作形状路由。

harness 得按活的形状挑

Vanilla Claude Code. Then GSD's multi-agent ceremony. Then superpowers' skill system. Then a custom orchestrator on top. Each rewrite came from a specific session that broke. What I learned: solo dev harnesses and team harnesses are different species, and the right answer isn't a default engine — it's routing by work shape.

I Rewrote My AI Coding Harness Four Times in Two Months

同一条流水线——Claude Design 出方向、HANDOFF.md 锁死、`big-task` 落地——但每个产品都把它掰成不一样的形状。ÉLAN 是空地，需要宽探索。识川 app 已经上线，要按类目一个个迁移。盘盘猫是 monorepo，v2 跑三个并行品牌表达而不是一个方向。变的是什么，不变的是什么。

同一套流水线，三个产品最重的一步完全不同

Same pipeline — Claude Design explores, HANDOFF.md locks, `big-task` ships — but it bends around each product. ÉLAN is greenfield and needs wide exploration before converging. Shichuan has a live app and needs incremental category-by-category migration. Panpanmao is a monorepo with a v2 redesign that runs three parallel brand expressions instead of one. What changes and what doesn't.

One Pipeline, Three Shapes: Running Claude Design + big-task Across ÉLAN, Shichuan, and Panpanmao

AX 的 AI 雷达每天从 50 多个信源抽约 500 条内容，挑 15 条给我。选题不靠 prompt，靠一份能版本化、能 diff 的 Markdown 编辑策略——一个 Agent 从反馈里迭代它。讲这把雷达的架构取舍和关键设计决策。

我把读新闻的口味写成了一个 md 文件

AX Radar filters ~500 stories a day from 50+ sources down to about 15. Editorial judgment isn't a prompt — it's a versionable, diff-able Markdown file that a Claude Agent iterates against user feedback. A breakdown of the architecture and the design calls.

488 AI Stories a Day: Why My Editorial Policy Is a Markdown File, Not a Prompt

上一篇讲的是我用 AX 雷达的一手体验。这一篇讲我给它加了 HTTP API、MCP server 和一份 Claude Skill——让 Claude 自己能查、能搜、能保存。两张脸共用一个后端；sha256 不是 bcrypt；pgvector 早就在了；Skill 才是放 domain 知识的地方。

claude 现在能自己查我的雷达了

Part 1 covered AX Radar from my side as the operator. Part 2 covers what I shipped this week: an HTTP API, an MCP server, and a Claude Code skill that let Claude read, search, and mutate the radar directly. Two surfaces, one backend. Why sha256 beats bcrypt here. Why semantic search rode the existing pgvector index. Why the skill is where domain knowledge lives.

AX Radar Needed a Second Face. One for Agents.

想跑团，凑不齐人。想玩剧本杀，差两个。想来一局狼人杀，只有四个人在线。如果 AI agent 能坐进来补位——不是那种复读机式的 NPC，是真的有人设、会推理、能演戏的 agent——会怎样？

凑不齐人的时候，AI 能不能顶上

想让几个 AI 坐一起辩论或来一局狼人杀？GitHub 上有二十几个相关项目。框架那头星数不少但没有'房间'概念，游戏那头全是一次性 hack。中间缺一个平台层——而搭好这个平台层之后你会发现，做好多 Agent 游戏和做好多 Agent 协作，用的是同一套东西。

Latest

Series

Building

Essays

Explainers & Lore

Workflow

Game Field Manuals

Subscribe