嗨，我是 Dustin。
Hi, I'm Dustin.
教了 12 年 GMAT，現在人在曼谷，用 AI 做教育產品。這裡寫的是一個沒有工程背景的人，怎麼靠 Claude Code 和 Gemini 把東西做出來的過程，還有途中踩過的坑。
12 years teaching GMAT. Now based in Bangkok, building education products with AI. This blog is about how someone with zero engineering background actually ships things — and all the pitfalls along the way.
精選
-
What I only learned after taking on enterprise AI training—prep time, the client liaison, instructor flexibility, and how to pick an instructor
Four things I learned after running a handful of enterprise AI training sessions. Prep time is 3-5x the session itself. The client liaison decides everything. The instructor has to handle wildly different skill levels. And if you're looking to hire a trainer, the signals are like picking a pop star.
ai-consulting -
親自接 AI 企業培訓之後才知道的事——準備時間、甲方對接、講師應變、選講師訊號
接了幾場企業 AI 培訓之後整理出來的四件事：準備時間是上課時間的 3-5 倍以上、甲方對接者決定一切、講師要能應付各種程度跟提問、想找講師的話訊號跟追星很像。
ai-consulting -
A Legal Counsel Used Claude Code to Catch Patent Infringement — Days Down to Hours
A legal counsel at a mid-sized company took our Claude Code course, then compressed a multi-day patent infringement audit into a single day — and actually caught a real case that is now in legal proceedings. This story finally turns "Claude Code is not just for engineers" from a slogan into a fact.
claude-code -
法務用 Claude Code 抓專利侵權，從幾天變成一天
一位幾十人公司的法務上完 Claude Code 課之後，把原本要好幾天的專利侵權查核壓到一天，而且真的抓到送上法律程序。這篇想記下這個案例，因為它讓「Claude Code 不是只給工程師用」這句話從口號變成事實。
claude-code -
The Truth and BS of Vibe Coding: Observations from a 10-Year Teaching Veteran
We shouldn't uniformly trash Vibe Coding, but we do need to recognize the traps in course quality. A teacher with 10+ years in the game shares how to tell the real deal from the fakes.
vibe-coding -
Vibe Coding 的真與假：一個從業十年老師的觀察
不該統一貶低 Vibe Coding，但也要認清課程品質的陷阱。一個教學從業十年以上的人，分享辨識有料與沒料的方法。
vibe-coding
最近在寫
-
Haiku 5.5 能接 Sonnet 的活嗎？我拿自己的工作考它
不看官方跑分，從自己 45 天、761 次的 subagent 派工裡挖題目考 Haiku 5.5。查核類七題它跟 Sonnet、Opus 同分，成本約十分之一；長逐字稿抽取還是不夠完整。
-
14 天 159 次確認，七成是誤判
盤點 PreToolUse hook 14 天跳出的 159 次 ask，約七成是誤判。攔截要比對指令位置而不是整段字串，成效從 transcript 的 hook 事件統計，不重播。
-
金鑰漏在你以為安全的地方
一週內三次金鑰差點外洩，入口都不是貼錯：skill 打包、遮蔽指令失手、build 設定。最後附上我判斷要不要輪換的準則。
-
某市政府 AI 推動評議：我給的四個建議
受邀參與某市政府研考會的 AI 推動工作評議，分享我看到的成績，以及我給的四個建議：人負責、風險排序、配套、額度看效率。
-
開 AI 課就是在割韭菜？我是怎麼回頭 refine 教學流程的
最近看到有人說開 AI 課就是割韭菜。我教 GMAT 15 年才駕輕就熟，剛開始教 AI 也超時、也浪費時間，後來回頭把整套教學流程重新做過。
-
為什麼 GPT 6 Sol / Luna 用起來這麼蠢
一個轉錄諮詢影片的小任務，聰明模型先搭管線、笨模型步步親力親為。請 Opus 5.5 接手 Codex 的爛攤子後，分析出 6 Sol / Luna 的四個失敗模式，以及 Cost per Task 該多一個時間變數。
-
Can Haiku 5.5 Take Over Sonnet's Work? I Tested It on My Own Tasks
Instead of official benchmarks, I built a test from 761 sub-agent jobs I actually delegated over 45 days. On seven checking tasks Haiku 5.5 matched Sonnet and Opus at about a tenth of the cost; long transcript extraction is still not complete enough.
-
159 Confirmation Prompts in 14 Days, and About 70% Were False Alarms
I audited 14 days of PreToolUse hook ask prompts: 159 of them, roughly 70% false positives. A guard hook should match the command position, not the whole string, and you can measure the fix from hook events in the transcript without replaying anything.
-
Keys Leak Where You Feel Safe
Three near-leaks of API keys in one week, and none came from pasting one somewhere public: skill packaging, a masking command that failed, and a build config. Plus the rule I use to decide whether to rotate.
-
Reviewing a City Government's AI Rollout: My Four Suggestions
I was invited to evaluate a city government's AI rollout. Here is what I saw and the four suggestions I gave: people own accountability, order by risk, fix the surroundings, judge quota by efficiency.
-
Is Teaching an AI Course Just Harvesting Chives? How I Went Back and Refined My Teaching Flow
Someone said running an AI course is just harvesting chives. I needed 15 years to get fluent at teaching GMAT, and when I started teaching AI I ran over time and wasted hours too, so I went back and rebuilt the whole flow.
-
Why GPT 6 Sol / Luna Feel So Dumb
A small job, transcribing consulting videos: the smart model builds a pipeline, the dumb one does every step by hand. After Opus 5.5 took over Codex's mess, it found four failure modes in 6 Sol / Luna, and why Cost per Task now needs a time variable.