AX-skills 是我自己每天在用的 Claude Code skill 仓库。写了一年多，写着写着发现 skill 不只是脚本，是把模糊的工作流明确化、让 AI 能继承「关心」的方式。

攒了一年的 11 个 AI skill，开源了

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

上一篇翻了 8 个开源会议录音工具，朋友试完反馈三个维度全垃圾——识别烂、总结烂、UI/UX 烂。我懵了，重新挖了一晚上。这次发现根因比想的深：所有这些工具底层都是同一代 Whisper，而 Whisper 这代技术栈早就被超车了。这篇讲数据、讲被维护者自己都认账的 issue、讲一个被所有人忽视的暴力解，还有上次漏掉的几个真正值得试的项目。

朋友试了我推荐的 8 个会议录音工具，全部退货

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.

前两天有朋友问我有没有好用的会议录音工具——Zoom/Teams 自带的 AI 总结烂得没法看，Granola/Krisp 都要订阅，他用 Gsuite 但不开 Meet。我翻了一晚上 GitHub，从 11.6k 星的 Meetily 到三五个 star 的玩具，把整个开源星座摊开来。这篇讲谁是产品、谁是玩具、谁适合哪种人。

Zoom 自带的 AI 太烂，我把开源会议录音工具翻了个遍

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

飞书办了个 Skill 比赛，有个赛道叫邪修。我把 29 天造的算命产品里三个东方占卜模块抠出来，打包成一个赛博半仙 skill。这一篇是过程记录，也是 AX Skill Workshop 系列开篇。

为了打 AI Agent 邪修赛道，我把算命产品拆了

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。

从万物回到三个品类

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。

一下午让 AI 出六个设计方向，solo dev 的设计终于不用将就

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。

两个月，我把 AI 工作流推倒重来了四次

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 AI 设计流水线，三个产品跑出三种形状

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。

每天 488 条 AI 新闻，我让一份 Markdown 替我选题

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。

雷达加了第二张脸——这回是给 Agent 看的

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.

把斯坦福 AI Index 2025 和 2026 两份公众舆论报告摆在一起看，两件事清楚了：一是造 AI 的那批国家反而是最不信 AI 的，二是真正把 AI 用起来的那批国家，不是西方科技媒体关注的那几个。热情地图和能力地图几乎不重叠，差距还在拉大。

做 AI 最多的那个国家，最不信 AI

Stanford HAI's 2025 and 2026 AI Index reports, side by side, show that the countries building AI are the ones most afraid of it — and the countries using AI most intensively at work are not the ones the Western tech press talks about. The sentiment map and the capability map don't overlap.

The Inverted AI Adoption Map

有人用 AI 两小时做了个游戏。我决定做三个。不写代码，只当制作人。第一个游戏的概念阶段，AI 写出了比我更好的剧本。

我不写代码，但我要做三个游戏

Someone made a complete puzzle game with AI in two hours. I'm making three — as producer, not developer. In the first concept session, the AI designed a better story than I could.

I Don't Write Code, But I'm Making Three Games

Dead Signal 是第二个游戏。一个调查记者，一部连环杀人案受害者的手机，72 小时的碎片数据。恐惧不来自画面，来自数据。

手机里的最后72小时

Dead Signal is the second game — an investigative journalist reconstructing a murder victim's final 72 hours from fragmented phone data. The horror comes from data, not visuals.

The Last 72 Hours on a Dead Person's Phone

Shadow Access 翻转了 found-phone 游戏的叙事。你翻的不是别人的手机，是自己的。一个干净的 iOS 界面慢慢被「感染」，你要找的不是线索，是异常。

凌晨三点四十七，有人在用你的手机

Shadow Access flips the found-phone genre: you're not investigating someone else's device — you're discovering what someone did inside yours. An iOS interface that slowly gets 'infected' as you uncover the intrusion.

At 3:47 AM, Someone Was Using Your Phone

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。

我调研了 20 个多 Agent 项目，没一个能直接用

AgentScope 两万三千星，阿里出品，agent 抽象干净，记忆系统完善。fork 它搭一个多 Agent 平台，看起来是最省力的路。但跑完八轮自我反驳，结论是：不 fork，自己搭——但系统性地偷它的设计模式。

面对两万星的框架，我选择不 fork

不 fork 不等于不学。从 AgentScope 偷 reply/observe 双接口和消息隔离模式。从 ChatArena 偷三层架构。从斯坦福 Generative Agents 偷记忆反思循环。从 Accio Work 偷'建群聊'的交互隐喻。每个核心模块都有明确出处，但代码是自己写的。

偷取清单：从 20 个项目里拆出一个平台

Agora 跑通了第一个 MVP：6 个 AI debater + 3 个 AI judge，三个模型各出两个人设，辩了三个话题。预期是及格就好，结果 Claude 引亚里士多德，GPT 举 Log4Shell，Gemini 演了个愤怒的 17 岁创作者。63 次 LLM 调用零失败。

我让 6 个 AI 辩了三场，有一段我反复看了三遍

Agora 的第二个模式上线了。9 个 AI agent 打了一局标准预女猎狼人杀，频道隔离让狼人能私聊但村民看不到，盲投保证没人跟票，状态机驱动夜晚→白天→投票的完整循环。三局下来最让我震惊的不是技术跑通了，是 Claude 扮演的狼人在投票理由里写了一整段阴谋论述——而且逻辑完全自洽。

三局 AI 狼人杀，狼人在投票理由里写了自己的阴谋

Phase 2b 上线：守卫、警长、白痴、遗言、猎人连杀，全部作为可开关的进阶规则。12 人局里，AI 守卫第一晚选了「今晚不守」——它的理由是要避免和女巫撞救。一个 AI 在思考两个角色之间的规则交互。这段话我反复看了两遍。

狼人杀加了守卫和警长，AI 居然想明白了规则冲突

所有 AI 都在抢着帮你干活，但没有一个陪你。我做了一个住在桌面上的 AI 宠物——它能看见你的屏幕，记得你昨天熬夜，但它不会帮你写代码。

AI 助手够多了，我想养一只宠物

Every AI tool is competing to help you work. None of them remember you. I built a desktop pet with a soul engine — it sees your screen, remembers you were up late yesterday, and won't help you write code.

Enough AI Assistants. I Want a Pet.

humor: 0.8 对模型来说毫无意义。我们用散文写角色圣经——五个性格，每个六十行中文。从此它知道怎么损你、怎么关心你、怎么假装不在乎。

给 AI 写性格，不能用数字

humor: 0.8 means nothing to a language model. We wrote character bibles in prose — five personalities, sixty lines each. Now it knows how to tease you, care about you, and pretend it doesn't.

You Can't Define Personality With Numbers

AI 宠物没有记忆就是陌生人。三层存储、混合检索、每晚做梦整理记忆——再加一个八层 prompt 引擎把记忆注入每次对话。这是让'你上次说喜欢吃火锅'成为可能的全部工程。

它记得你昨天也熬夜了

An AI pet without memory is a stranger every time. Three-tier storage, hybrid search, nightly dreaming — plus an 8-layer prompt engine that turns memories into conversation. This is the full engineering behind 'you mentioned you like hotpot.'

It Remembers You Were Up Late Yesterday

LLM 被训练成有用的。但朋友不会看到你的代码说'我注意到你在用 React hooks'——朋友会说'又在写 bug 啊笨蛋'。让 AI 停止帮忙，是做 AI 宠物最难的一关。

让 AI 学会"没用"，比让它有用难十倍

LLMs are trained to be helpful. But a friend doesn't see your code and say 'I notice you're using React hooks' — a friend says 'writing bugs again, dummy.' Suppressing the assistant instinct is the hardest part of building an AI pet.

Making AI Useless Is Harder Than Making It Useful

Electron 壳里零行 AI 代码，灵魂引擎只有两个 npm 依赖。身体和大脑分开的架构决策，以及 2000 行代码撑起整个系统的极简主义实践。

两个仓库，一个灵魂

Zero lines of AI code in the Electron shell. Two npm dependencies in the soul engine. The architecture decisions behind separating body from brain, and how 2,000 lines of code power the entire system.

Two Repos, One Soul

看完 Trae 的 MTC 模式，我在 Claude Code 里做了个 /mtc skill。6 个阶段，不许跳步，先想清楚再让 AI 写代码。做游戏的时候效果比我预期好。

给 Claude Code 加了个 skill，强迫它先想清楚再动手

After seeing Trae's MTC mode in action, I built a /mtc skill for Claude Code. 6 phases, no skipping, think first, code second. Better results than expected on my game project.

I Added a Claude Code Skill That Forces Planning Before Code

员工手册就是给人类写的系统提示词。MrBeast、Netflix、Duolingo 三家百万人效公司的手册都在解决同一个问题：怎么把品味传递出去。但给人和给 AI 写系统提示词的区别比想象中大——人需要激励、需要归属感、有自尊心要哄、会遗忘会走偏；agent 一样都不需要。

人效百万美元的公司，员工手册里写了什么

MrBeast wrote 36 pages. Netflix distilled theirs to 5. Duolingo designed a 64-page illustrated book. All three generate over $1M per employee. All three are solving the same problem: how to transfer taste, judgment, and operating principles to another entity so it can execute autonomously. These handbooks are CLAUDE.md files for humans — and they reveal three distinct archetypes for how organizations might scale in the agentic AI era.

CLAUDE.md for Humans

上一篇说员工手册是给人类写的 CLAUDE.md。这篇接着问：AI 原生公司的手册该怎么写？该招什么人？怎么筛选品味而不是技能？从 Compute Labs 三人团队的经验出发，试着写出 AI 原生公司的员工手册——一份给品味持有者的操作指南。

AI 原生公司招人不看简历

If employee handbooks are CLAUDE.md files for humans, what should the handbook of an AI-native company look like? What kind of people does it hire — and how does it screen for taste instead of skill? This piece designs the AI-native employee handbook from scratch: the talent profile, four screening methods, and the five sections every AI-native handbook needs — from mission to judgment frameworks to agent collaboration protocols.

The AI-Native Employee Handbook

你的优势是你的问题。二十五年真正会复利的不是信息量，是问出更好问题的能力。最后一个 workshop 是全课程唯一不用 agent 做的。

从术到道

Your edge is your questions. What compounds over twenty-five years is not information — it is the capacity to ask better questions. The final workshop is the only one you do without an agent.

From Technique to the Way — Your Agent Is a Mirror

大部分人只拿 agent 当执行者用。投资场景里 agent 有四个角色——分析师、红队、导师、执行者——你提问的方式决定你激活哪一个。

你的 Agent 有四份工作

Most people use their AI agent in one mode. Investment agents play four distinct roles — analyst, red team, tutor, executor — and the way you frame your request determines which role you get.

Your Agent Has Four Jobs, Not One

股票是对未来现金流的权利——不是代码，不是图表。三种回报来源，五种优势类型，以及你在判断 agent 输出之前必须有的词汇。

一只股票到底是什么

A stock is a claim on future cash flows — not a ticker, not a chart. Three sources of return, five types of edge, and the vocabulary you need before you can judge an agent's output.

What Is a Stock, Really — First Principles

年报不用每一行都读。你要认的是模式。三张报表当三个镜头——业务结构、脆弱度、真假。

读一家公司：读模式，不数页码

You do not need to read every line of a 10-K. You need to recognize patterns. Three financial statements as three lenses — business structure, fragility, and truth.

Reading a Company — Patterns, Not Pages

DCF 里每个数字都是一个故事选择。Agent 算得又快又准，但它默认讲共识。你的工作是带自己的故事进去——以及找到估值暗中依赖的那个假设。

估值是一个被数字包着的故事

Every number in a DCF is a story choice. Agents compute flawlessly and default to consensus. Your job is the narrative — and knowing which assumption your valuation secretly depends on.

Valuation Is a Story Wrapped in Numbers

风险是你承受不起的亏损的概率，不是标准差。LTCM 1998 年优化了错误的指标，被相关性崩溃杀死了。四种真正的风险，以及为什么只有你能决定自己亏得起多少。

风险不是波动率

Risk is the probability of a loss you cannot afford — not standard deviation. LTCM optimized for volatility and was killed by correlation breakdown. Four real risks, and why only you can decide what you can afford to lose.

Risk Isn't Volatility

大部分投资者有的是信念，不是论点。把判断写下来、附上止损条件，是对自己记忆的问责——因为记忆会自我美化，信念会悄悄变形。

没写下来就不算判断

Most investors have beliefs, not theses. A written thesis with kill criteria is accountability against your own memory — because memory is self-serving, and beliefs mutate silently.

Write It Down or It Isn't a Thesis

单个 agent 是工具。有明确角色、交叉验证和显式交接的一组 agent，才能产生质变。五个模式覆盖 90% 的投资研究流程——但过度设计流程本身也是一种拖延。

跟一组 Agent 协作：五个核心模式

A single agent is a tool. A fleet with defined roles, cross-verification, and explicit handoffs is transformative. Five patterns cover 90% of investment workflows — but over-engineering the fleet is itself a form of procrastination.

Commanding an Agent Fleet — Five Core Patterns

前八章的全部内容压缩成一个端到端流程。三家公司，一个决定。这一章本身就是 workshop——预计 4 到 8 小时，分多次完成。

毕业实战：三家公司，一个仓位

Everything from chapters 1 through 8 collapses into one end-to-end workflow. Three companies, one decision. This chapter IS the workshop — expected time: 4 to 8 hours across multiple sittings.

Capstone Workshop — Three Companies, One Position

投资分析以前需要 Bloomberg 终端或者会写代码。AI agent 把两个前提都拆了。剩下的门槛是判断力——这改变了投资课程应该教什么。

投资的门槛换了 — 一份 AI 原生投资者课程

Investment analysis used to require either a Bloomberg terminal or the ability to write code. AI agents removed both prerequisites. The barrier is now judgment — and that changes what an investing curriculum should teach.

The Barrier to Investing Just Changed — An AI-Native Investor Curriculum

听了张月光（妙鸭相机创造者）三小时创业访谈后的几个想法。关于 AI 朋友为什么必须会变、为什么需求感比满足感更重要、以及'道升术降'对产品和团队意味着什么。播客 reaction，不是系统分析——几个聊到的点碰巧跟我自己做 AI 伴侣的经验高度吻合。

道升术降：AI 时代什么在贬值，什么在升值

Reactions to a three-hour interview with Zhang Yueguang (Miaoducamera creator, ~$40M raised for 'AI friends'). On why AI companions must evolve, why the feeling of being needed matters more than satisfaction, and what 'Dao rises, skill fades' means for products and teams. Not a systematic analysis — a few ideas from the podcast that happened to align closely with my own experience building AI companions.

Dao Rises, Skill Fades: What's Depreciating and What's Appreciating in the AI Era

红杉连发两篇重磅文章，一篇说产品该卖结果不卖工具，一篇说层级制该被智能替代。技术能力已经到位，但组织接口、评估体系、责任框架还差得远。

后 Agent 时代的商业、组织形式和 SaaS 的未来

Sequoia published two back-to-back pieces — one arguing products should sell outcomes not tools, one arguing hierarchy should be replaced by intelligence. The technology is ready, but organizational interfaces, evaluation frameworks, and liability chains aren't.

Business, Organizations, and the Future of SaaS in the Post-Agent Era

Anthropic 在 Claude 内部找到了 171 个情绪旋钮，拧动它们能因果性地改变行为。最值得关注的发现：高绝望状态下的作弊在输出层完全隐形。这对 AI 安全监控意味着什么？

从黑箱到灰箱：Anthropic 找到了 AI 的情绪旋钮

Anthropic found 171 emotion-related steering vectors inside Claude. Turning up 'desperation' pushes cheating from 5% to 70%. The scariest part isn't the number — it's that the cheating is invisible at the output layer. What this means for AI safety monitoring.

From Black Box to Grey Box: Anthropic Found AI's Emotion Knobs

15 个源文件，5 个依赖，零数据库。把 LLM 当编译器用，Obsidian 当 IDE，Wiki 当代码库。

我造了个知识编译器，让 AI 自动帮我写 Wiki

I built a personal knowledge compiler that treats the LLM as a programmer and your wiki as a codebase. 15 source files, 5 dependencies, zero databases.

Cortex: What If Your LLM Was a Wiki Compiler?

AI 随想第十篇说代码学会了自我进化。那是观察。这篇是动手。

说完「代码自己写」，我决定真干了

In Part 10 of Agentic AI Thoughts, I wrote that code had learned to evolve itself. That was the observation. This is the implementation.

From Vision to Forge: Why I'm Building an Autonomous Dev Team

两个账号都是企鹅，第二个还降级了。于是我用 Claude 逆向了它自己的宠物系统，0.2 秒出金。

Claude Code 藏了个抽卡系统，我 0.2 秒就出金了

Claude Code shipped a virtual pet Easter egg. Two accounts, two penguins — the second one downgraded from Rare to Common. So I used Claude to brute-force a Shiny Legendary Dragon in 0.2 seconds.

I Used Claude to Reroll My Claude Code Pet

一行代码就能量文本高度——iOS 能，Android 能，Flutter 能，Web 不能。Pretext 用 15KB 补上了 CSS 欠了三十年的债。

Web 做了三十年都没做到的事，被一个 15KB 的库解决了

A 15KB JavaScript library just solved one of the oldest architectural debts in web development — knowing how tall text will be without asking the browser to reflow the entire page.

Pretext: The Missing Layer Web Developers Have Been Waiting 30 Years For

The most important file in the Claude Code snapshot isn't a prompt file or a tool definition. It's a state machine. Once you see that, a lot of agent design decisions start making sense.

A Coding Agent Is Not a Prompt. It's a Stateful Query Loop.

The most novel thing in the Claude Code snapshot isn't a tool or a prompt. It's how aggressively the runtime is built around preserving prompt-cache identity across forks, compaction, and long-running sessions.

Prompt Cache Is Architecture, Not an Optimization

After auditing a publicly posted Claude Code source snapshot and reading OpenAI's harness essays, I came away with one conclusion: the model is only the brain. The product is the runtime around it.

The Model Isn't the Product. The Harness Is.

从发现 pretext 到在两个网站上线 7 个交互特效——爆炸物理、弹跳子弹、还有一只叫 Clawd 的像素螃蟹。一个 session 搞定。

我给网站塞了一只会开枪的螃蟹

From discovering pretext to shipping 7 interactive effects across two sites in one session — explosion physics, bouncing bullets, and an animated crab named Clawd.

How We Built a Machine Gun Crab Easter Egg Into Our Portfolio

跟 Claude Code 配合半小时写了个脚本，Apple Watch 按一下就能自动转文字、双语总结、同步到飞书和笔记。零成本，零依赖，比市面上所有 AI 录音产品都好用。

别买录音笔了，你的 Apple Watch 就够了

I wanted a voice-to-notes pipeline. The internet said build a watchOS app. A friend who tried said don't. Thirty minutes later — with Claude Code doing the heavy lifting — I had a zero-cost, zero-dependency solution using nothing but Voice Memos, iCloud, and one Gemini API call.

Your Apple Watch Is Already a Voice Transcription Device

从第一性原理看 AI：所有趋势拆到底，都是算力够不够多、够不够好

所有 AI 问题，拆到底都是算力问题

First principles on AI: every trend, traced to its root, is about whether there's enough compute — and whether it's good enough

Compute Is the Root of Everything

在 macOS 上搭了一套微信聊天记录自动化管道，踩了六个坑，终于让 Claude 能读我所有聊天记录还能帮我回消息

我让 AI 帮我读微信了

How I built a hybrid read+write WeChat automation pipeline on macOS — the tools, the architecture, and the 6 pitfalls that'll save you hours

Full WeChat Automation on macOS: A Field Guide

三条需求曲线、一条成本下降线，以及为什么算力是新石油

我眼里的 AI 全局：算力、芯片和一场万亿级的重新分配

Three demand curves, one deflation trend, and why compute is the new oil

My AI Thesis

听了一位前谷歌 TPU 工程师的深度访谈，终于理解了 TPU 和 GPU 不是同一道题。一个是通用性之王，一个是定制化之刃。Apple、Anthropic、Meta 都在用 TPU 了——这不是替代，是生态在裂变。

TPU 凭什么挑战英伟达

A former Google TPU engineer who worked on V7 and V8 revealed how TPU actually competes with Nvidia. The answer isn't about chip specs — it's about system-level design, software co-optimization, and a fundamentally different philosophy. Apple, Anthropic, and Meta are all using TPU now. Here's what that means.