Sean Lu
Sean Lu

Sean Lu

元智資管大二。用 $0 成本的工具和自動化,把自己養活。

Sophomore in Information Management at Yuan Ze University, Taiwan. I build tools and automation that cost nothing to run, so I can support myself.

關於我About

2029 年畢業,住在桃園、學校附近。持有 Anthropic Claude 101 認證(2026 年 7 月)。校內參加閱人社和資管系排球隊。

I graduate in 2029 and live in Taoyuan, near campus. I hold Anthropic's Claude 101 certification (July 2026). On campus I'm in the 閱人社 club and the IM department volleyball team.

正在做的事What I'm working on

目標很直接:在家裡還支持我的這幾年,先做到經濟上自給自足。

The goal is plain: become financially self-sufficient while my family is still supporting me.

技術背景Technical background

我做東西的原則How I build

自己不用的工具,不算做完。我只信自己感受過的需求。我曾以為可以先做出來再找需求,後來發現太多東西只有我自己會用。現在只做我真的缺、缺了會去解決的東西。

A tool I don't use myself isn't finished.I only trust demand I've felt. I used to think you could ship first and find demand later; too much of what I built turned out to be useful only to me. Now I build what I actually lack and would otherwise fix by hand.

作品Projects

考場筆記 · Personal IELTS Desk Cloudflare Pages · Gemini · Web Speech

繁中 IELTS 自學系統。單字庫、Leitner 五盒複習、AI 模擬題、錯題本、寫作評分、每日規劃,全部跑在瀏覽器。

A Traditional-Chinese IELTS study desk. Vocabulary bank, Leitner review, AI mock questions, mistake notebook, band-descriptor marking and a daily planner, all running in the browser.

問題
市面 IELTS 工具要付費、要帳號、資料不在自己手上。
Problem
Existing IELTS tools cost money, require an account, and keep your data.
解法
全部跑在瀏覽器。API key 存 localStorage,資料可匯出 JSON,雲端同步是選配。
Approach
Everything runs in the browser. The API key lives in localStorage, data exports as JSON, cloud sync is optional.
取捨
Gemini 直接從瀏覽器呼叫,沒有自己的後端。伺服器成本是零,代價是使用者得自己申請 key。複習演算法用 Leitner 不用 SM-2,因為使用者看得懂自己在第幾盒。
Trade-offs
Gemini is called straight from the browser, with no backend of my own. Server cost is zero; the price is that users bring their own key. The review algorithm is Leitner rather than SM-2, because you can see which box you're in.
學到
動態抓模型清單,比把模型名稱寫死耐用。另外寫了一份給 Claude Code 的交接包,後續迭代成本大幅下降。
Learned
Dynamic model discovery outlasts hardcoded model names. A handoff package for Claude Code made later iterations far cheaper.
vid-for-agents Shell · ffmpeg · whisper.cpp

一行指令把任何影片變成取樣畫格 + 本地逐字稿,讓 AI coding agent 能「看」影片。

One command turns any video into sampled frames plus a local transcript, so AI coding agents can read it.

問題
Agent 只能讀文字。影片教學、螢幕錄影對它們是黑盒。
Problem
Agents read text. Video tutorials and screen recordings are a black box to them.
解法
把影片拆成 agent 讀得懂的兩種東西:畫格和文字。
Approach
Reduce the video to the two things an agent can read: frames and text.
取捨
語音辨識跑本地 whisper.cpp。慢,但不用 key、不上傳、$0。
Trade-offs
Speech-to-text runs locally on whisper.cpp. Slower, but no key, no upload, $0.
CountAgent Claude Code · iOS Shortcuts

跑在 Claude Code 上的個人記帳 agent。Siri 往檔案追加一行純文字,agent 之後負責解析、分類、去重。

A personal bookkeeping agent on Claude Code. Siri appends one line of plain text; the agent later parses, categorises and dedupes.

問題
記帳 app 的摩擦在輸入。開 app、選類別、填金額,三步就放棄。
Problem
The friction in expense apps is input. Open the app, pick a category, type the amount. Three steps and you quit.
解法
輸入端只剩一句話,理解交給 agent 批次處理。
Approach
Input shrinks to one sentence. Understanding is handed to the agent in batches.
取捨
行為寫在 CLAUDE.md 裡,是自然語言不是程式碼。改規則不用動 code。代價是行為不完全確定,所以保留人工審核那一關。
Trade-offs
Behaviour lives in a CLAUDE.md, written in plain language rather than code. Rules change without a code change. The price is non-determinism, so a human review step stays.

這個月This month

2026 年 9 月

September 2026

工作之外Outside work

哲學和知識論。在乎遊戲和電影的敘事品質。理想的一天裡有我的貓。

Philosophy and epistemology. Narrative quality in games and film. My cat is part of the ideal day.

聯絡Contact

合作、專案,或只是想聊聊。直接寫信。

Collaborations, projects, or just a conversation. Email works.