20 years in Enterprise IT. One month with Claude Code. Never going back.
I'm not a software engineer. I haven't solved a LeetCode problem in my life. For nearly two decades, I ran IT infrastructure at companies where "writing code" meant updating a spreadsheet macro.
Then AI happened.
Taiwan (1985-) -- Studied CS at National Central University. Spent more time running the campus network, co-building a MUD (still running after 20+ years), and managing the highest-traffic personal BBS board than writing code.
US (2008-) -- Master's in CIS. Then nearly 20 years of enterprise IT:
| Domain | Stack |
|---|---|
| Identity & Endpoints | Google Workspace, M365, Intune, Entra ID |
| ERP | Dynamics NAV |
| BI | Domo |
| CRM | Salesforce |
| Security | NIST CSF, ISO 27001 |
One year, I wrote maybe one line of code. Never touched GitHub.
Now -- Downloaded Claude Code. Built an AI agent framework in weeks. Shipped more in a month than most side projects ship in a year.
I design from an IT veteran's perspective -- not "how do I write the cleverest code," but "what actually works for people who aren't engineers."
Ai-Sister · live
A web app that orchestrates ChatGPT, Claude, Gemini, and Grok from a single input -- open to everyone, running on my own paid API keys. Parallel chat, structured debate, an 8-step coding loop, a multi-round roundtable, plus single-AI agent modes. Tiered plans with quota + subscription billing, per-tier model routing, SSE streaming. Hono + SQLite + React on Oracle Cloud ARM behind Caddy. Grew out of my multi-ai-chat Chrome extension.
A security hardening layer for AI agent hosts. A chattr +i policy baseline no agent can rewrite, a guardian subagent, and a deterministic cross-patrol heartbeat / dead-man's-switch -- bolted onto OpenClaw without touching its code. Apache-2.0.
Claude Code skills wrapper. Lightest weight, curated features, maximum performance. One generator command, you're running. 25 MB RAM, 672 KB on disk.
Unified long-term memory for AI agents. SQLite + Oracle + Vector Search.
A hands-on study of IDN homograph attacks -- the "looks legit, isn't" threat I spent years defending against.
This isn't code written by a 20-year software engineer. It's the best path found by a 20-year IT / MIS / Cybersecurity veteran who knows what enterprises actually need.
The best developer tools come from people who've spent decades as users of technology. I know what breaks at 3 AM, what IT teams actually adopt vs. what dies after the demo, and why "just works" beats "highly configurable" every time.
Building a multi-AI agent fleet -- Claude, GPT, Gemini, and local models working as a team with defined roles, real accountability, and honest performance scoring. Not a demo. Running in production, every day.
Taipei → Los Angeles
20 年企業 IT。一個月 Claude Code。回不去了。
我不是軟體工程師。這輩子沒解過一題 LeetCode。將近二十年,我在企業裡管 IT 基礎建設,「寫程式」對我來說就是改 Excel 巨集。
然後 AI 來了。
台灣(1985-) -- 中央大學資工系。比起寫程式,更多時間在管校園網路、共同開發 MUD(營運至今 20+ 年),以及經營全校最高人氣 BBS 個版。
美國(2008-) -- CIS 碩士。接著近 20 年企業 IT:
| 領域 | 技術棧 |
|---|---|
| 身分與端點管理 | Google Workspace, M365, Intune, Entra ID |
| ERP | Dynamics NAV |
| BI | Domo |
| CRM | Salesforce |
| 資安 | NIST CSF, ISO 27001 |
一整年,我大概寫了一行 code。從沒碰過 GitHub。
現在 -- 下載 Claude Code。幾週內打造一套 AI Agent 框架。一個月的產出超過多數 side project 一年的進度。
我用 IT 老兵的視角 設計工具 -- 不是「怎麼寫最聰明的 code」,而是「什麼東西非工程師也能用」。
Ai-Sister · 線上服務
一個 web app,一個輸入同時調度 ChatGPT、Claude、Gemini、Grok -- 開放給所有人用,跑在我自己付費的 API 上。平行對話、結構化辯論、8 步 coding loop、多輪圓桌論證,外加單一 AI 的 agent 模式。分級方案(額度 + 訂閱計費)、各級模型路由、SSE 串流。Hono + SQLite + React,部署在 Oracle Cloud ARM、前掛 Caddy。源自我的 multi-ai-chat Chrome 擴充。
AI agent 主機的安全強化層。任何 agent 都改不掉的 chattr +i 政策基線、守護 subagent,以及確定性的 cross-patrol heartbeat / dead-man's-switch -- 直接加掛在 OpenClaw 上,完全不動它的程式碼。Apache-2.0。
Claude Code skills wrapper。最輕量、精選功能、極致效能。一條指令,直接跑。25 MB RAM,672 KB 磁碟。
AI Agent 統一長期記憶。SQLite + Oracle + Vector Search。
IDN homograph 攻擊的實作研究 -- 那種「看起來合法、其實不是」的威脅,正是我防守多年的東西。
這不是一個 20 年軟體工程師寫的程式。這是一個 20 年 IT / MIS / 資安老兵找到的最佳路徑。
最好的開發者工具,來自於花了數十年當技術 使用者 的人。我知道什麼東西凌晨三點會壞、IT 團隊真正會採用什麼 vs. demo 完就丟的東西,以及為什麼「就是能用」永遠打贏「超級可設定」。
打造多 AI Agent 艦隊 -- Claude、GPT、Gemini、本地模型作為一個團隊運作,有明確角色、真實問責、誠實績效評分。不是 demo,每天都在 production 跑。
台北 → 洛杉磯