Posts
All the articles I've posted.
-
Push, Not Pull: On Anthropic's Week of Communication
The user surges of recent years were never pull. They were push. Both sides are just competing over who screws up less, and this week Anthropic acted the whole thing out again.
- Updated:
The Week of Qwen 3.8 Open Weights: From Model Card to A3B Getting Deleted
Qwen 3.8 27B open weights land, I run it, the 35B A3B shows up registered on Modelscope, and then it gets deleted. All within one week.
- Updated:
2026 模型脾氣觀察：Gemini、Claude、Codex 的個性對比
用了一整年下來，三家的旗艦模型各自有很明顯的「脾氣」——Gemini 3 像戲精博士、Claude 4.7 像油條前輩、GPT-5.5 反而是這波最務實的同事。把累積的觀察整理成一篇對比，順便講一個從「貧嘴密度」反推版本號的玩法。
- Updated:
AI 工具使用哲學——不選陣營，選思維
拍了快三十部教學影片後的感觸：工具會變，思維不變。不用否定別人的工具選擇，真正重要的是指揮的邏輯——拆分任務、規劃驗證、分配職責。
-
把佛法交給 agent，讓他用佛法約束自己
一個叫 buddhist-method 的 SKILL，用六條佛法概念約束 agent 的行為——查證、找根因、重新確認狀態、丟掉草稿、捨斷成因、不因壓力改立場。
- Updated:
跟 Claude 5 對話的 7 個實用技巧：不發散、不話癆、不偷懶、不雞婆
一些最近跟朋友分享的，跟 Claude 5 代系列模型對話的實用小技巧：自製 explain skill 抓外星話、一次回完別讓對話發散、語音輸入配置、first-principles 剪枝、用 hook 治偷懶、CLAUDE.md 極簡化，以及治雞婆的幻影反駁。
- Updated:
額度是怎麼被燒掉的：Claude Code 快取 bug、Opus 4.7 的 2x 消耗、subagent 遞迴繁殖
四起額度事故的完整紀錄：兩個官方快取 bug、JSONL 扒出來的 80 倍 API 請求、Opus 4.7 實測 2x 消耗，以及我自己一個 session 半小時燒掉 90% 的 subagent 遞迴。
- Updated:
網頁版 Claude Code 環境設定的兩個世界
網頁版 Claude Code 有兩種環境——雲端 VM 跟 Remote-control。這篇整理我實測下來的設定訣竅、會踩到的小 bug，以及為什麼有些設定不能無腦丟到雲端。
- Updated:
一年下來，我對「該用終端機還是桌面版」的答案變了三次
從三月的 Cowork 燒 token、四月的「答案還是終端機」、五月的「初學者直接從桌面版開始就好」，到八月的「CLI 是親生的」。同一個問題，我在五篇文章裡給過三種不同的答案，這裡照時間順序排出來。
- Updated:
數位遊牧是可遇不可求，不要把它當成追求目標
從 2016 年就開始世界各地旅遊加遠端工作的我可以負責任地說：數位遊牧是可遇、但不可求，不要把它當成 life style 的追求目標。
-
幫企業導入 AI，教案設計最常犯的三個錯
從實際的企業培訓案例中，整理出三個教案設計的核心錯誤：統一任務不等於統一方法論、開放式問卷無效、以及缺少「自主應用」這一步。
- Updated:
Fable 5 回歸一週：把最貴的模型變成 Skill Distillation 引擎
Fable 5 限時回歸一週，我沒拿去做日常小事，而是拿去做高槓桿判斷、蒸餾成便宜模型可照做的 skills。
-
免費只給 15 分鐘：知識變現的四條判準
教齡 15 年累積下來的四條判準：免費諮詢只給 15 分鐘、先當社群分享者再談變現、別看不起你覺得 google 就有的東西、以及怎麼分辨同業裡的詐騙。
- Updated:
Gemini 3.7 Flash——這半年內第一個能拿去做 agent 的 Gemini
Gemini 3.7 Flash 發布後我的體感：終於是這半年內可以用在 agent 工作上的 Gemini。比 Sonnet 5 略強但沒接近 Sol，價格對半折，額度根本用不完。另附一節：同一家的 API 和網頁版根本是兩回事。
- Updated:
規則越加，Claude 越不聽話——派一隊 AI 重整設定，常態上下文省 36%
半夜拿到 Opus 4.8、額度剛好重置，我做的第一件事是減肥我的 harness。一個 COO 學員把 1M 上下文用到 88% 的案例，讓我認真面對一件事：規則加越多，模型反而越不聽話。
-
你要在一個業界打滾非常久，才知道真正的痛點是什麼
用 Fable 翻新 GMAT 教學主力產品，抓出隱藏 bug、改善演算法，順便講兩件做產品的感想。
- Updated:
iPad 工作流進階——斷網斷電的備案，與我為什麼放棄妙控鍵盤
iPad 跑 Claude Code 全攻略的續篇：談的不是安裝，是穩健性——曼谷家用網路會斷、雨季會停電，主機端怎麼撐住；以及我為什麼放棄妙控鍵盤、改用一張嘴加手勢。
-
只換外面那圈 loop
LongHorizon-Harness 架在 Claude Code、Codex 這些 agent 外面，不訓練模型也不取代你的 agent，只做迴圈，WeaveBench 完成率從 51.8 拉到 80.7。
-
不是拉力，是推力：從這一週 Anthropic 的溝通說起
近幾年用戶暴漲從來不是拉力，是推力。兩邊就是在比誰少犯錯，而這一週 Anthropic 的幾則發言，剛好把這件事演了一遍。
- Updated:
Qwen 3.8 開放權重的這一週：從模型卡到 A3B 被刪
從 Qwen 3.8 27B 模型卡放出、實測體感、到 35B A3B 在 Modelscope 註冊又被刪掉，一週之內的完整過程。
-
Your AI Tools Are Now the Attack Vector: npm and Python Supply-Chain Backdoors, a $1000 Stolen-Key Bill, and a Scan-Before-You-Install SOP
Attackers are planting instructions in Claude Code, Cursor, and Gemini CLI configs so your own assistant runs the exfiltration script. Two supply-chain waves, one $1000 stolen-key bill, and the five minutes to spend before installing anything.
-
Your Automated Video Pipeline Is Silently Dropping Work: Gotchas From Remotion, ffmpeg, the YouTube API, and HyperFrames
Four months of video automation gotchas, from Remotion through HyperFrames: truncated downloads that exit 0, thumbnail titles cut mid-word, transcript chunks returning zero lines, large uploads that hang. None of them raise an error.
-
你的 AI 工具已經是攻擊媒介：npm/Python 供應鏈後門、外洩金鑰盜刷 1000 美元，與裝 MCP 前的掃描 SOP
攻擊者已經在 Claude Code、Cursor、Gemini CLI 的設定檔裡植入指令，讓你的 AI 助手幫他們跑竊取腳本。這篇串起兩波供應鏈攻擊、一次一千美元的盜刷，以及安裝前該做的那五分鐘。
-
自動發片管道正在悄悄漏成品：Remotion、ffmpeg、YouTube API、HyperFrames 的踩坑總表
從 Remotion 到 HyperFrames，四個月的影片自動化踩坑合輯：exit 0 的殘檔、靜默截斷的縮圖標題、回 0 行的轉錄合併、卡死的大檔上傳。共同點是全都不報錯。
-
ABC Legal's Agent Fleet: Turning AI Experiments Into a Governed Production System
Six cards summarizing Anthropic's ABC Legal case study: 50+ production agents, agents managed as software in git, a steering committee with no software developers on it, and the principle that trust comes before automation.
-
Once It Ships, It's Gone: Three Times an Agent Misfired on an External System
Calendar invites, replies, and the internal notes on a hold event are all one-way doors. I got each of them wrong once this month, so here are the shapes of the accidents and the checks that catch them.
-
Channel Value and Deal Value Belong in Separate Ledgers
Three rules that came out of two weeks of back-to-back BD calls: score a deal's channel value separately from its deal value, use The Pumpkin Plan to decide which slice to take, and write down a hard ceiling for free consulting.
-
My CLAUDE.md Boiled Down to Eight Lines of "Take Pride In…, Take Shame In…"
My CLAUDE.md cut down to eight paired maxims, plus a human-machine collaboration circular written deliberately in Party-document style, plus a sixteen-character guideline.
-
Codex Falls Off the Pedestal: People Caught the Quota Nerf
A month ago I was writing about the Codex quota honeymoon. Now people are measuring a 50% nerf on Plus and 77% on Pro 5x, and my "off the pedestal in two months" call turns out to have been optimistic.
-
Eleven HyperFrames Video Gotchas
Transition flicker, a worker-parallelism myth carried over from another framework, transcript format drift, slide sizing, and silently truncated thumbnails — eleven things that bit me on the HyperFrames video line in one week of August.
-
A Local Model Is Not a Way to Save Money
If you are buying hardware to run a local model because you want to save money, let me do the math with you first: NT$100k minimum, commercial models at bleeding prices, and privacy as the only advantage left.
-
A Chatty Model: Teaching Opus 5 to Converge
Opus 5 does not speak plainly, and it also overwrites. A cue card turns into a treatise. I use a first-principles skill to teach it to converge.
-
Five Things I Never Saw Coming Two Years Ago
Five changes in myself I never saw coming two years ago: the terminal, cancelling Office 365, voice input, dispatching AI agents, and a desktop that finally got cleaned up.
-
ABC Legal 的代理人艦隊：把 AI 實驗變成可治理的生產系統
整理 Anthropic 官方 ABC Legal 案例的六張圖卡：50+ 生產環境代理人、把 agent 當軟體放進 git、非工程師組成的 steering committee，以及「可信任，才自動化」的營運原則。
-
寄出去就收不回來：agent 動外部系統的三次誤操作
行事曆邀請、回信、佔位事件的內部備註，這三種動作都是寄出即無法收回。這個月我在這三處各犯一次，記下事故形狀與防呆判準。
-
通路價值和成案價值，要分開記帳
這兩週連著跑幾場 BD，逼出三條判準：一筆生意的價值要拆成通路和成案兩層分開算、用《南瓜計畫》決定接哪一塊、以及免費諮詢的上限寫死在哪裡。
-
我的 CLAUDE.md 精煉到只剩八句「以…為榮，以…為恥」
把 CLAUDE.md 一路砍到只剩八組對句，外加一篇刻意寫成公文體的人機協同宣導文，還有十六字方針。
-
Codex 走下神壇：額度偷縮被網友抓出來了
一個多月前我還在寫 Codex 的額度蜜月期，現在網友實測 Plus 被偷縮 50%、Pro 5x 被偷縮 77%，我那句「兩個月跌下神壇」還是太樂觀了。
-
HyperFrames 影片產出的十一個坑
轉場閃動、worker 並行迷思、字幕格式漂移、投影片版式、縮圖靜默截斷——八月這一週在 HyperFrames 這條影片線上踩到的十一個坑，逐個記現象、根因、解法。
-
本地模型不是拿來省錢的
如果你買硬體架本地模型的目的是省錢，先聽我幫你算這筆帳：硬體十萬起跳、商用模型卷到流血價，本地模型唯一不可取代的優勢只剩隱私。
-
話癆的模型：教 Opus 5 怎麼收斂
Opus 5 除了不講人話之外還會過度寫作，小抄寫成萬言書。我用 first-principles skill 教它收斂，順帶談怎麼遏制 overthinking。
-
兩年前從來沒想過會發生在自己身上的五件事
兩年前從來沒有想過、卻真的發生在自己身上的五個變化：終端機、退訂 Office 365、語音輸入、派 AI Agent，還有被整理乾淨的桌面。
-
Two Skills That Make AI Talk Like a Human: ASD-STE100 for Grammar, ISO 24495 for Structure
A skill I threw together three weeks ago got passed around by people overseas and picked up three PRs, so I shipped a second one I use every single day.
-
讓 AI 對你說人話的兩個 skill：ASD-STE100 管文法，ISO 24495 管結構
三個禮拜前隨手做的 asd-ste100-skill 被老外瘋狂轉載還留了三個 PR，順手再開一個天天在用的 iso-24495-skill。
-
Information Diet: I Pulled My Own Chrome History and Audited Where My Attention Goes
People obsess over whether every bite of food is clean, then swallow piles of dirty info online without a second thought. So I pulled my own Chrome history db and had it analyze my browsing habits. A few sections came out uglier than I expected.
-
Information Diet：我抓自己的 Chrome 瀏覽紀錄，做了一次注意力盤點
現代人極度在意吃進口裡的每一塊食物乾不乾淨，卻毫不在乎上網吃到的一大堆 Dirty Info。所以我抓了自己的 Chrome 瀏覽紀錄 db，讓工具分析我的上網習慣，結果有幾段比我預期的難看。
-
AI Will Never Go to Jail for You — I Figured Out What I Want to Teach, Then Got Told I Had Half of It Wrong
Since generative AI arrived I have been asking what AI can never do on a human's behalf. My answer is accountability, because AI does not get sentenced. Then I posted the argument and found out what I had missed about the risk carried by people at the bottom.
-
AI 不會代替人去坐牢——我想清楚要教什麼，然後被網友提醒想錯了一半
生成式 AI 之後，我一直在想有什麼是 AI 永遠不能替人類做的。我的答案是「負責」，因為 AI 不會被判刑。但這個論述貼出去之後，我發現自己漏想了基層承擔的風險。
-
NT$18,000 a Month for an AI Course: Price and Value Came Apart a While Ago
An AI-course cash grab overheard at a convenience store, set against the two things I can actually show: a SKILL distilled from fifteen years of material, and a mock exam interface pulling past a thousand USD a month. Price tracks marketing, not delivery.
-
My Context Window Is More Fragile Than Today's Frontier Models: I Open-Sourced a Document Review SKILL
Working with an agent drains my attention, so I worked out a review loop: he writes the full proposal, I review it by voice, he revises, I go eat and exercise, then v2 shows up.
-
AI 課一個月一萬八：教學市場的價格跟價值早就脫鉤了
在超商用餐區聽到的 AI 課吸金實況，對照我自己手上兩件真的能驗證的東西：15 年教材蒸餾成的 SKILL，跟每月破千美金的模擬考訂閱。價格反映的是行銷強度，不是交付密度。
-
我的 Context Window 比當今大模型還脆弱：我開源了一個文檔審查 SKILL
跟 Agent 合作會注意力耗弱，於是我摸索出一套審查流程：請他出完整提案書，我錄音審查，他改，我去吃飯運動，等 v2。
-
Why I Built My Own Newsletter
Once I started using AI, the feeds I follow changed completely. Threads is full of third-, fourth-, fifth-hand reposts with extra seasoning added, so I pulled together 300-plus sources and had AI build my own weekly digest.
-
為什麼我自己做了一份電子報
開始用 AI 之後，我關注的社群媒體整個換掉。脆上很多是三四五六手搬運還要加料，所以我自己整理 300 多個源頭，讓 AI 做出我個人的電子週報。
-
The CLI Is the Firstborn
New features and fresh bug fixes land in the CLI first. Remote control waits forever. Same company, wildly different update cadence.
-
The Machines That Need the CLI Most Belong to People Least Likely to Learn It
For a machine short on resources the CLI uses a fraction of what a GUI does, but the people with the least headroom are also the least likely to ever learn it. Plus a note on how low the bar for local models actually is.
-
A July Full of Resets: A Light AI User Reviews the Subsidy War
Every time a reset landed in July I used it to the last drop, and the ledger says 57.5x. Plus notes on the subsidy war, Deepseek's price hike, and my bet on the next reset.
-
Installing Gemini CLI With Beginners Cures Most Claude Code Ailments
Five steps I walk Claude Code / Codex beginners through when installing the Gemini Antigravity CLI: multimodal use cases, getting comfortable with a CLI, a quick win from voice transcription, and finishing on data governance and local mode.
-
CLI 是親生的，其他都是後媽養的
新功能和剛修好的 bug 都優先落在 CLI，remote control 要等猴年馬月。同一家公司的產品，更新頻率差很多。
-
最需要 CLI 的電腦，主人最不可能學會 CLI
對資源有限的電腦來說 CLI 佔用是 GUI 的零頭，但電腦資源最不夠的族群也剛好最不可能學會 CLI；順帶聊本機模型的硬體門檻其實沒那麼高。
-
充滿 reset 的七月：一個 AI 輕量用戶的補貼大戰復盤
七月一有 reset 就用好用滿，帳面發揮 57.5x 的價值；順便記下補貼大戰、Deepseek 漲價，和我對下一次重置的預測。
-
帶初學者裝 Gemini CLI，治 Claude Code 百病
帶 Claude Code / Codex 初學者裝 Gemini Antigravity CLI 的五個步驟：從多模態 use case、CLI 介面的心理建設、語音轉錄的 quick win，一路帶到資料治理與本機模式。
-
The Alignment Gap Is Closing. Next Comes Taste and Verification
The gap between AI and human intent is closing fast, so what separates good output moves toward taste and verification — and verification is where human responsibility stays.
-
Other People Fly to Korea for Cosmetic Surgery. I Had Codex Do Mine.
Someone on Twitter chained Codex, Hyperframes, IndexTTS2 and HeyGen into a one-person media pipeline. I pulled the repo, tried it, played with a local TTS model, then rebuilt a video days later on Claude Code.
-
Six New Context Engineering Rules for Claude 5, and the 1,473 Lines I Cut
After reading Anthropic's context engineering guidance for Claude 5, I turned the key points into nine cards, then cut 1,473 lines from a harness I had built up over four model generations.
-
Fewer Prompts Is Not Less Control
Claude 5 and GPT-5.6 official guidance is converging on the same thing — retiring old-style prompt stacking. But trimming is not letting go; control just moves. And the payoff of maintaining that layering is switching tools without losing a step.
-
My Local Model Lineup on a Mac mini, Plus Three Bad Habits I'm Owning Up To
Ten days of offline models on a 48GB Mac mini M4 Pro — the picks, the task split, a terrifying swap write rate, and three bad habits I admit to.
-
Porting Old Prompts to Opus 5: A Nine-Card Migration Guide
Nine cards I made after reading Anthropic's official Opus 5 prompting guide: response length, agent narration, task boundaries, subagent delegation, self-correction, and what happens when you turn thinking off.
-
Supposedly the Smartest Model, and Opus 5 Spent My Whole Day Spinning in Place
Wrong languages, sudden Simplified Chinese, then whole turns with no visible output — and a session log that matched an issue open since June 15.
-
Two Habits for Voice-Driving Coding Agents, and I Am Still Finding the Balance
Short instructions go through push-to-talk; walking through a whole course outline or system architecture goes through a full QuickTime recording that I hand to the AI to transcribe and execute. Plus the setup I use to burn the AI quota that came free with Google Drive.
-
對齊的 gap 正在縮小，接下來拚的是品味與驗證
AI 對齊人類意圖的 gap 正在快速縮小，區分產出品質的要素會往品味與驗證轉移；而驗證背後是人類永遠不會被取代的責任。
-
別人去韓國做醫美，我叫 Codex 數位醫美
從推特上看到有人把 Codex、Hyperframes、IndexTTS2、HeyGen 串成一條自媒體流水線，拉下來實測、順手玩 b 站的本地 TTS，幾天後換 Claude Code 重做一支。
-
Claude 5 時代的情境工程六條新規則，與我砍掉的 1473 行 harness
讀完 Anthropic 對 Claude 5 的情境工程指南後，我把重點做成九張圖卡，然後照著把累積四個世代的 harness 削掉 1473 行。
-
更少的 prompt，不是更少的控制
Claude 5 與 GPT-5.6 的官方指南正在合流，都在淘汰舊式 prompt 堆疊。但精簡不等於放任，控制只是換了位置——而維護好這套分層的紅利，是換哪個工具都能立刻接手。
-
Mac mini 上的本地模型陣容，順便招認我的三大劣根性
48GB 的 Mac mini M4 Pro 試了十來天離線模型，選型結論、任務分工、swap 嚇死人的寫入量，還有我承認的三個劣根性。
-
舊 prompt 搬到 Opus 5：九張卡的遷移指南
讀完 Anthropic 官方的 Opus 5 提示指南後整理的九張卡：回答長度、代理敘述、任務邊界、子代理委派、自我修正，還有關掉 thinking 的副作用。
-
號稱最聰明的 Opus 5，一整天在我對話裡空轉
從回錯語言、突然寫簡體，到整輪沒有可見輸出，查 session log 之後對上一個 6/15 就開著沒修的 issue。
-
語音下指令的兩種習慣，我還在找平衡
用 Claude Code 或 Codex 時，短指令我用隨按即錄，要盤點整套思路時我改開 QuickTime 完整錄音，再丟給 AI 轉錄執行。附上我拿 Google 雲端硬碟送的額度做轉錄的設定。
-
I'm Stuck at Stage 2.5 of AI Adoption
After watching Boris Cherny break down the stages of AI adoption, it hit home: I'm stuck at stage 2.5, where limited attention plus low trust becomes a vicious cycle. This week I used late-night schedules, a morning dashboard, and herdr auto-spawning sessions to cut a 4-5 hour workflow down to 1-2 hours.
-
A Zero While Sitting on a Gold Mine
I read a piece on the perceived value behind churn-and-burn courses, turned it into a skill, ran it on my own site, and got back a verdict: a zero while sitting on a gold mine.
-
Some of My Subagents Were Already Grandpas
A CCX quota incident with no guardrails set: subagents bred recursively, one session burned 90% in half an hour, and here is the full forensics and fix.
-
Codex Built Its Own Evidence Package and Went to Argue With Google Support
A leaked Gemini backend key at PDT Learning got abused, no spending cap, and burned 1000 USD. I pointed Codex's browser automation at Google's live support to fight the charge, and it went so hard it built a 15-page evidence package and sent it over.
-
Two Traps Running Claude Code on Local Ollama: Truncated Context, and A3B Buckling Under Heavy Verification
Two things I logged: cco (Claude Code driven by local Ollama) had long been giving off-topic answers, and the root cause was not a weak model but a full harness whose system prompt had ballooned to 30-50k tokens and was being silently truncated; then I put A3B on the Mac mini for two days as a night worker.
-
Instead of Begging the Model Not to Lie, I Wrote a Hook That Stops It
The sequel to the Opus 4.8 confabulation post: I moved "don't make up numbers" from a plea in CLAUDE.md to a pending-guard hook that blocks git commit at PreToolUse.
-
Second Harness Diet: Global Skills From 58 Down to 40
A follow-up to the harness diet series: six months later, a bigger cleanup that cut my global skills from 58 down to 40, start to finish.
-
我卡在 AI 導入的 2.5 階段
看了 Boris Cherny 分享的 AI 導入階段框架後有感而發：我卡在 2.5 階段，注意力有限加信任不足變成惡性循環，這週靠深夜排程、晨間儀表板、herdr 自動開 session 把 4-5 小時的協作壓到 1-2 小時。
-
坐擁金礦的零分
讀到一篇談割韭菜課程感知價值的文章，我馬上做成 skill 拿去盤點自己的網站，換來一句「坐擁金礦的零分」。
-
有些 subagent 都當阿公了
一次 CCX 沒設好護欄的額度事故：subagent 遞迴繁殖，一個 session 半小時燒掉 90%，事後鑑識與修法全記錄。
-
Codex 自己做了一份證據包，跑去跟 Google 真人客服吵架
PDT Learning 一支 Gemini backend key 外洩被盜刷、沒設 spending cap 怒噴 1000 USD，我用 Codex 的瀏覽器操作去跟 Google 真人客服爭費用，它認真到自己生出一份 15 頁證據包發給對方。
-
本機 Ollama 跑 Claude Code 的兩個坑：context 被截斷、A3B 撐不住重驗證
記錄兩件事：cco（本機 Ollama 驅動 Claude Code）長期答非所問，根因不是模型不夠聰明，是完整 harness 的 system prompt 早就膨脹到 3-5 萬 token 被靜默截斷；順手把 Mac mini 換上 A3B 當夜間工人測了兩天。
-
與其拜託模型不要騙我，不如寫一個擋得住的 hook
接續 Opus 4.8 捏造工具輸出那篇：我把「別亂編數字」從 CLAUDE.md 的拜託，升級成一個掛在 PreToolUse 擋 git commit 的 pending-guard hook。
-
第二次 harness 減肥：全域 skills 58 砍到 40
接續 harness 減肥系列，半年後做了第二輪更大規模的整理，把全域 skills 從 58 個砍到 40 個的完整過程。
-
A Few Days Into Switching From Claude Code to Codex: The Quota Honeymoon
A running log of switching from Claude Code to Codex this week: quota I could not burn through fast enough, a $20 Sol Ultra run that beat a $100 Fable plan on a professor friend's paper, and a tool-chain swap to Hyperframe.
-
從 Claude Code 換到 Codex 的這幾天：額度蜜月期
這週從 Claude Code 換到 Codex 的體感記錄：額度多到花不完、Sol Ultra 20 鎂幫教授跑論文，比 Fable 100 鎂方案還划算，工具鏈也跟著換成 Hyperframe。
-
Picking a Brand Look as a Non-Designer: AI Made Prototypes Cheap, So Taste Became the Hard Part
I am not a designer, but I spent this week iterating on a visual system for AgentCrew Academy with Fable, GPT-image-2, and GPT-5.6-sol. Here are the four rejected directions and the final spec applied to the website, slides, and documents.
-
外行人選品牌視覺：AI 把提案做便宜了，難的變成你的品味
我不是設計師，這週用 Fable、GPT-image-2、GPT-5.6-sol 迭代出 AgentCrew Academy 的視覺系統，記下被否決的四個方向與最後套用到網站、投影片、文件的規格。
-
Four Things I Have Learned from Teaching AI
Recent corporate workshops reinforced four lessons for me: teach in person when possible, cut the slide count, show smart people the result first, and let students do the work.
-
I Finally Switched My Daily Driver from Claude Code to Codex
I try new tools easily, but I am slow to leave the ones already built into my routine. This time, quotas, pricing, and product direction all crossed the line together.
-
最近教 AI 課學到的四件事
最近幾次企業培訓讓我重新確認四件事：實體課更好掌握節奏、投影片要做減法、聰明人先看結果，以及學員真正需要的是放手實作。
-
我終於把主力從 Claude Code 換成 Codex 了
我很愛試新工具，卻很難離開已經用習慣的工具。這次真正讓我換主力的，不是一次 benchmark，而是額度、價格與可預見的產品方向一起越過了臨界點。