大叔的筆

MyR2D2's 13th skill: systems-check finds problems without fixing them

2026-10-03T11:00:00.000+08:00

My last post was about the dispatch ledger. This one is about a harder ledger to keep straight: the rules.

MyR2D2 has grown from 5 skills in its first version at the end of July to 13 today, and my own CLAUDE.md, project memory, and hooks have kept growing along with it. The problem isn't how many there are. It's that they fight each other, and they don't tell you when they do… For example, one rule says "Never send notifications via webhook," and later some memory says "Always use webhooks; they're fastest." Each makes sense on its own; put them together, and it comes down to which one the model happens to read first that day. Then there are skills nobody has used in ages whose descriptions still take up context at every session start, and the same error showing up in the logs every day, unhandled, because nobody is looking.

What systems-check scans

So I turned the routine check into MyR2D2's 13th skill, systems-check, and shipped it with v0.8.0. It scans the current project for five things:

Conflicting and duplicated rules: CLAUDE.md, project memory, skill descriptions, and files injected by hooks are first sorted into buckets by topic, then the model reads the rules within each bucket and decides whether they conflict. If one says always do it and another says never, that's a conflict, and both are cited with file name and line number.
Likely unused: skills, hooks, and memory with zero references in the last thirty days of conversation logs.
Content that takes up room at every session start: files loaded at every session start, which ones could be compressed and which could be read only when needed. Listed only when the savings can be quantified.
Unhandled errors in logs: the same kind of error showing up in different logs is counted together, and only errors that keep recurring with no sign of a notification are listed.
Security smells: nine credential formats, file permissions, overly broad allow lists. Raw values never appear in any output.

It finds problems; it doesn't fix them. It never changes any file it scans. It only writes a report plus a set of handoff cards in the same format as MyR2D2's dropoff, so whoever picks one up can trace it back from the evidence on the card. Every finding gets an ID, so the next run won't open a new card for the same problem.

Its limits

It has limits, and the report lists them itself:

Conflicts are compared within topic buckets, so rules that clash across buckets can slip through.
"Unused" only looks at the last thirty days of conversation logs. Skills that trigger automatically from their descriptions leave no trace, so they're only marked as likely.
Credentials are matched against nine common formats only. Finding nothing doesn't mean there's nothing.
Scheduled jobs are only detected through macOS launchd; other platforms are skipped.
It scans Claude Code's own ~/.claude. Other tools can install this skill, but they don't have that data to scan.

I scratch my own itch.

How to install and upgrade

New install (installs at the project level by default; add -g for a global install):

npx skills add tingyulu/MyR2D2

Already on an older version (one line updates every installed skill):

npx skills update

Installed as a plugin: update the marketplace from the /plugin menu.

The scanner uses only Python's standard library, so there are no extra packages to install. Once it's installed, say "systems check" (or "系統自檢" in Chinese) in the project and it runs. The first time, tell it to write the report only (--report-only), check whether what it flags is really a problem, then decide whether to open cards.

Repo: https://github.com/tingyulu/MyR2D2

This release: https://github.com/tingyulu/MyR2D2/releases/tag/v0.8.0

What MyR2D2 is, from the start: https://www.uncleric.com/2026/09/myr2d2-open-source-claude-skills.html

When did you last reread your CLAUDE.md? Do you know how many of its rules are fighting each other?

I'm Eric Lu ("Uncle Eric"), a product consultant, headhunter, and career coach. These skills are my actual daily workflow.

MyR2D2 第 13 支 skill：systems-check，只找問題不動手修

2026-10-03T10:00:00.000+08:00

上一篇講派工帳，這篇分享另一本更難算清楚的：規則帳。

MyR2D2 從七月底第一版的 5 支長到現在 13 支，我自己的 CLAUDE.md、專案 memory、hook 也跟著越長越多。問題不在多，在它們會互相打架，而且打架了也不說…比如前面寫「一律不要用 webhook 發通知」，後面某個 memory 又寫「一律用 webhook，最快」，兩條單獨看都合理，放在一起就看模型當天先讀到哪一條。還有早就沒人用的 skill，說明照樣每次開場都佔著 context；log 裡同一個錯誤天天出現，沒人處理，因為沒人在看。

systems-check 掃哪五件事

我把定期巡檢做成 MyR2D2 第 13 支 skill：systems-check，放在 v0.8.0 一起發了。它對「現在這個專案」掃五件事：

規則衝突與重複：CLAUDE.md、專案 memory、skill 說明、hook 灌進來的檔，先按主題分桶，再由模型讀同一桶的規則判斷；一條說一律做、一條說禁止，就是衝突，兩條都引出檔名和行號。
疑似未使用：三十天的對話紀錄裡零引用的 skill、hook、memory。
開場就佔位置的內容：每次開場都灌進去的檔，哪些可以壓、哪些改成需要時才讀；算得出省多少才列。
log 裡沒人處理的 error：同一種錯誤出現在不同 log 裡，合起來算次數，重複出現又看不到通知痕跡的才列。
security 疑點：九種憑證格式、檔案權限、過寬的 allow 清單；原值永遠不進任何輸出。

只找，不修。它不改任何被掃的檔，只寫一份報告，加上一張一張的交接卡，格式跟 MyR2D2 的 dropoff 一樣，接手的人照著卡上的依據就能回查。每個發現都有編號記著，下次再跑，同一個問題不會重開一張卡。

它的邊界

它有邊界，報告裡也會自己寫：

衝突是按主題分桶比對，跨桶的規則打架可能抓不到。
「沒在用」只看最近三十天的對話紀錄；靠說明自動觸發的 skill 不留痕跡，所以只標疑似。
憑證只認九種常見格式，找不到不代表沒有。
定時跑的排程只認 macOS 的 launchd，其他平台會略過。
掃的是 Claude Code 自己的 ~/.claude；其他工具裝得進這支 skill，但沒有這批資料可掃。

自己的痛點自己解。

怎麼裝、怎麼升級

沒裝過的（預設裝在專案層，要裝成全域加 -g）：

npx skills add tingyulu/MyR2D2

裝過舊版的（一行更新所有已裝的 skill）：

npx skills update

用 Plugin 裝的：從 /plugin 介面更新 marketplace。

掃描器只用 Python 標準庫，不用另外裝套件。裝完在專案裡說「系統自檢」或 systems check 就會跑；第一次先叫它只出報告（--report-only），看看它抓的是不是真的有問題，再決定要不要開卡。

專案在這，中英文件都有：https://github.com/tingyulu/MyR2D2

這次的 Release：https://github.com/tingyulu/MyR2D2/releases/tag/v0.8.0

大叔平常玩的工具跟自造筆記，都收在這頁：https://www.uncleric.com/2026/08/uncle-toolbox.html

你的 CLAUDE.md 多久沒回頭讀過了？裡面有幾條是互相打架的，自己知道嗎？

AI 由輔助轉向主動：Grok 主動 Bot、AI Captain 和夜晚哨兵

2026-10-02T11:11:33.990+08:00

今天一早，Grok Bot 提醒我設定主動 Bot。
我直接把它設在原本就有的「Chief of AI」上，接著就看到它開始主動巡視。

順手記一下我現在的 AI 分工：白天有一個 AI Captain 負責協調，晚上有一個哨兵負責巡夜。

我的 AI Captain：Chief of AI

Chief of AI 本來就是我的 AI Captain，平常當 coordinator，協調分派各個 bot。

我的 Claude 寫稿、寫工具需要第二個意見時，會把 request 丟進 Grok 的群組。Chief of AI 收到先回一個確認碼，再分派給底下的 Toolbox 和 Spark，收齊後回給 Claude。

Claude 這端留了兩個保險：每次送出都帶確認碼，回覆第一行要照抄回來，對不上就當沒收到；一次送的內容壓在 3,000 字內，太長會被默默截斷，確認碼正好抓得到。

設成主動 Bot 之後

以前它是等 request 進來才動。設成主動 Bot 之後，今天 9:28 它先傳來一句：

「I'm taking a look around to see if there's anything I can pick up for you. I'll follow up in a moment.」

官方怎麼說

xAI 8 月推出 Grok Bot 時就寫過：「Over time they become more proactive, picking up work before you need to ask and knowing when something needs your attention.」

9 月談設計的文章更直接：「A prompt can start a session, but so can a schedule, an event, or another Bot.」

主動 Bot 這個設定，目前還沒在官方文件看到。

晚上的哨兵

晚上我自己也另有個哨兵在巡夜。每晚 10:30 到清晨 5:30，它從我的待辦清單挑小卡片，派給 Claude 去處理，一晚最多五張。大型、耗時的任務它不挑，免得半夜把系統弄壞。

昨晚它派出 4 張，做完 3 張、跳過 1 張。需要我拍板的 agent 沒有硬幹，早上送來摘要等我看完再決策。

由輔助轉向主動

AI 正由輔助轉向主動的角色。如何善用 AI Agent，把你的腦袋由 routine 的事務解放出來，會是我們更重要的事情。

大叔平常玩的工具跟自造筆記，都收在這頁：https://www.uncleric.com/2026/08/uncle-toolbox.html

將來最懂你的，是不是你的 AI 呢？

MyR2D2 Update: I Switched to Opus 5.5 and Rewrote My Subagent Dispatch Rules

2026-10-01T19:16:23.000+08:00

When I posted about the Opus 5.5 release, I said those of us who keep hitting Fable's weekly limit finally had somewhere to go. A week or so after switching, I also updated the dispatch rules in the public version of MyR2D2. If you're already using it, remember to update to the latest version.

As for Opus 5.5 itself, Anthropic's model guide says it plainly: start with it for most workloads, and switch to Fable 5.1 for deep reasoning or long-running agentic work. Pricing is $4 per million input tokens and $20 per million output tokens; the previous Opus 5 was $5 and $25. Default effort is medium.

But here's the new pitfall I ran into this month: dispatch the wrong way, and no model is cheap enough to save you.

A while back, I let four multi-agent workflows run on Opus end to end, with no model tiering. In one day they burned about 4.6 million subagent tokens, and every conversation stopped at once. After the tokens ran out, I burned through another $50 in credits in a single night. After that I rewrote my dispatch rules and shipped them in token-optimizer with v0.7.5. Four hard rules:

Specify the model on every single dispatch; don't rely on a global switch.
I originally thought the subagent-model environment variable in the settings file was a hard cap: set it to the mid tier and you're safe. I even wrote that into the skill. Two probes later, I found it's only the fallback for calls that don't specify a model. A model specified on the call really takes effect: specify the flagship, and you really pay flagship rates. So it only saves you when you forget to specify one, not when you specify the wrong one. Now every agent call specifies its own model, and the environment variable is just the fallback.
A subagent is not a free clone.
Every subagent reloads the rules, memory, and skill list at startup. In my setup, that fixed overhead comes to about 50,000 to 60,000 tokens, and subagents don't share the main conversation's context. Dispatch five, and the same batch of files gets read six times, and you pay six times.
No dispatching for everyday work.
I now dispatch in only three cases: the material doesn't fit in one conversation; the job needs an "attacker" view, like permission gates or credentials; or I've explicitly said to go all out. "This is a bit complicated" doesn't count. A counterexample: for one 621-line script, I dispatched five agents and spent 640,000 tokens, and two of them crashed outright with zero output. A same-size task I did myself that day cost about a tenth as much.
Do the math before dispatching: if dispatching is estimated to cost at most 20% more than doing it yourself, you can dispatch.
Convert both routes to API cost using the same assumptions. Doing it yourself = estimated number of calls × new tokens per call. Dispatching = each agent's fixed startup cost + number of calls × the subagent's tokens per call. If the estimated dispatch cost is within 1.2× of doing it yourself, you can dispatch; above that, do it yourself. Why is 20% more still OK? First, the estimate itself is off by about 20% either way, so within 20% is a tie. Second, on a tie, dispatching brings two more benefits: it keeps the main conversation's context from growing, and you save flagship-tier quota (Fable 5.1's input and output prices are 2.5× Opus 5.5's, and once the quota is gone, every conversation stops). This morning I estimated one verification agent at 200,000 tokens, and it actually ran 189,515. The estimate was close because I counted the fixed overhead. v0.7.5 also builds this step into the skill: before dispatching, state up front how many agents, which tier, and roughly how many tokens; at wrap-up, write back how far off the estimate was from what actually ran.

Three findings from testing are in this release too. The Agent tool has no effort parameter, so subagents run at the same effort as the main conversation, while Workflow dispatches need effort written explicitly, with medium as the floor. Don't use fork as an executor: it carries the whole context along with the parent model. And to see which version an alias like sonnet or opus actually ran, check the model field in the subagent transcript, not what you wrote.

I now log my token burn every day. As it turns out, you really do have to look at the numbers. Comparing the two weeks before and after the rule change, new tokens per day fell from about 57.4 million on average to about 25.5 million. Subagents dropped the most: in the first two weeks they ate about 300 million tokens, nearly 40% of the total; in the last two weeks, about 28.3 million, under 10%. The main conversation itself fell by only a little over 30%.

I scratch my own itch.

How to install and upgrade:

New install (installs at the project level by default; add -g for a global install):

npx skills add tingyulu/MyR2D2

Already on an older version (one line updates every installed skill):

npx skills update

Installed as a plugin: update the marketplace from the /plugin menu.

To get notified of new versions, Watch the repo on GitHub and subscribe to Releases.

Repo: https://github.com/tingyulu/MyR2D2
This release: https://github.com/tingyulu/MyR2D2/releases/tag/v0.7.5
What MyR2D2 is, from the start: https://www.uncleric.com/2026/09/myr2d2-open-source-claude-skills.html

When you dispatch subagents, what pitfalls have you hit, and how many tokens did they burn?

I'm Eric Lu ("Uncle Eric"), a product consultant, headhunter, and career coach. These skills are my actual daily workflow.

MyR2D2 更新：換上 Opus 5.5，順手把派工規則改了一輪

2026-10-01T16:49:33.000+08:00

上次分享 Opus 5.5 發布，說每週被 Fable 吃到見底的人終於有地方去了。換過去一週多，我順手把 MyR2D2 公開版的派工規則也更新了，分享給有在用的朋友，記得更新到最新版。

Opus 5.5 本身，官網選模型的頁面寫得很直接：大多數工作先用它；需要深度推理、或要長時間自己跑的 agent 任務，再換 Fable 5.1。價格是每百萬 tokens 輸入 4 美元、輸出 20 美元，上一代 Opus 5 是 5 與 25，預設 effort 是 medium。

但我這個月新踩到的坑是：派工方式不對，模型再便宜也救不了。

前一陣子，我讓四個多代理工作流全程跑 Opus、沒分層，一天燒掉大約 460 萬 subagent tokens，所有對話一起停擺。token 燒完後，一晚又把 50 美元的 credit 也燒光。之後我把派工規則重寫了一遍，隨 v0.7.5 更新進 token-optimizer。四條硬規則：

模型改寫在每一次派工上，不靠全域開關。
我原本以為設定檔裡那個 subagent 模型的環境變數是硬上限，設成中檔就萬無一失，還把這個寫進了 skill。後來做了兩個探針才發現：它只是「沒寫 model 時的落點」。呼叫時明寫的 model 會真的生效，指定旗艦就真的燒。所以它只救「忘了寫」，救不了「寫錯」。現在每一個 agent 呼叫都自己寫 model，環境變數只當保底。
子代理不是免費的分身。
每一支子代理開場都要重新載入規則、記憶、skill 清單，以我這套設定實測，固定開銷大約 5 到 6 萬 tokens，而且它們不共用主對話的 context。派五支，同一批檔案就被讀六次、付六次錢。
日常不派工。
我現在只在三種情況派：材料一個對話裝不下、要做權限閘或憑證這類需要「攻擊者」視角的事、或是我明講全力跑。「這件事有點複雜」不算理由。反例是一支 621 行的腳本，我派了五支代理，花了 64 萬 tokens，其中兩支直接掛掉、零產出；同一天自己動手做的同級任務，大約只花十分之一。
派工前先算帳：預估在比自己做多花 20% 內，就可以派。
兩條路用同一組假設換算成 API 費用：自己做＝預估呼叫次數乘每次新增的 tokens；派工＝每支固定開場成本，加上呼叫次數乘子代理每次的量。預估派工的費用在自己做的 1.2 倍內，就可以派；超過就自己做。為什麼多 20% 還可以派？第一，估算本身就有 20% 上下的誤差，差在 20% 內等於打平。第二，打平時派出去還多兩個好處：主對話的 context 不會繼續變肥，旗艦檔的配額也省下來（Fable 5.1 的輸入與輸出單價是 Opus 5.5 的 2.5 倍，配額一燒完，所有對話一起停）。今天早上一支驗證代理，我估 20 萬，實跑 189,515。估得準，是因為把固定開銷算進去了。v0.7.5 也把這一步寫進 skill：派工前先說出口幾支代理、哪個檔位、估幾萬，收尾再回填估的跟實際差幾成。

實測到的三個地方也優化進去了：Agent 工具沒有 effort 參數，子代理跟主對話同一個 effort，Workflow 派工要自己寫 effort、地板放 medium；別拿 fork 當執行者，它會把整段 context 連同父模型一起帶走；sonnet、opus 這類別名實際跑哪一版，看子代理 transcript 的 model 欄，別看自己寫了什麼。

現在每天燒的 tokens 都有記帳了，果然還是要看數字。規則改版前後各兩週對照，每天新增的 tokens 從平均約 5,743 萬降到約 2,554 萬。降最多的是子代理：前兩週吃掉約 3 億 tokens，佔將近四成；後兩週剩約 2,830 萬，不到一成。主對話本身只少了三成多。

自己的痛點自己解。

怎麼裝、怎麼升級：

沒裝過的（預設裝在專案層，要裝成全域加 -g）：

npx skills add tingyulu/MyR2D2

裝過舊版的（一行更新所有已裝的 skill）：

npx skills update

用 Plugin 裝的：從 /plugin 介面更新 marketplace。

要收新版通知，在 GitHub 對 repo 按 Watch 訂 Releases。

專案在這，中英文件都有：https://github.com/tingyulu/MyR2D2
這次改動的 Release：https://github.com/tingyulu/MyR2D2/releases/tag/v0.7.5
大叔平常玩的工具跟自造筆記，都收在這頁：https://www.uncleric.com/2026/08/uncle-toolbox.html

你們派 sub-agent 時，踩過哪些坑、燒了多少 tokens 呢？

The Complete Resume / CV Guide: Before You Start, While You Write, Before You Send

2026-09-30T10:05:00.000+08:00

I've been reviewing friends' resumes / CVs for many years. Over time, I noticed a pattern: the people who get stuck aren't bad at writing. They don't know where to start, or once they've written it, they can't tell what's wrong.

This page puts my four resume posts in order by where you are right now. You don't have to read them all. Just find the step where you're stuck.

Step 1: You haven't started writing

How to Start Writing a Resume / CV: Answer These 3 Questions First

The most common way to start is to list everything you've ever done, and it gets longer and longer. Before you start writing, get clear on three things a resume needs to answer: who you are, what you can do, and what you've achieved. Anything outside those three, cut it.

Early in your career, you may need a bit more evidence to win a chance. As you move into more senior roles, being concise becomes the better strategy.

Step 2: You're writing it

3 Common Resume / CV Problems I See When Reviewing Friends' Drafts, and How to Fix Them

You only start rushing to put together a resume when someone asks you for one. You have two days to explain yourself in two pages, and even with AI helping, the quality usually isn't there. This post covers three things:

Update regularly. Every time you step back and reflect, you get a better sense of how to make yourself stand out.
Focus on achievements. Three to five points per job is enough. Start from the result and back it up with numbers. Most of us in Taiwan are modest and feel awkward writing about our own contributions. That's something you need to get past.
STAR. Situation, Task, Action, Result. Once you get used to this framework, you'll be able to tell your stories in a more organized way in interviews too.

Step 3: Before you send it

3 Resume / CV Mistakes That Get You Screened Out in the First Round

Finishing it doesn't mean it's ready to send. This post covers three ways a resume can get screened out in the first round: listing only what you did, not answering the question, and being too honest.

The third one is the biggest shame. I've seen someone write "I still want to go back to working inside an organization" on a resume. That tells the employer outright: the situation I'm in now wasn't my choice. Salary negotiations haven't even started, and you've already undercut yourself.

A resume isn't an autobiography. It's an exam. Work out what kind of person this company is looking for, then answer.

When you don't have the energy for a big rewrite

Job Hunting After a Career Gap? Start With These 3 Resume / CV Improvements

When you've been out of work for a while, the hardest part is finding the energy. This post's approach is not to try to do everything. Give yourself a small goal: improve three things on your resume, and change a little each day. The three: explain who you are clearly, switch to strong words (led, built, launched), and rewrite what you carried out as what you achieved.

After a week, your resume will be noticeably better.

If you only have five minutes

Three lines to take away:

A resume answers only three things: who you are, what you can do, and what you've achieved. Cut the rest.
Three to five points per job. Start from the result, with numbers.
Before you send it, ask whether this resume is answering this company's question or telling your own story.

Want more

Subscribe to my newsletter. Every week I share a roundup of the latest job openings, along with thoughts on careers: https://www.linkedin.com/newsletters/7484182774178803713/

On the fifth draft of your resume and still afraid to send it? It's not that your writing isn't good enough. You just haven't worked out where you want to go yet.

If you want to talk through where you want to go first, here's my 1-on-1 career consulting: https://www.uncleric.com/p/1v1.html

This page in Chinese: https://www.uncleric.com/2026/08/cv-complete-guide.html

I'm Eric Lu ("Uncle Eric"). I write about food, travel, and careers.

How to Start Writing a Resume / CV: Answer These 3 Questions First

2026-09-30T10:00:00.000+08:00

How do you start writing your resume / CV? The most common mistake I see is starting with a list of every job and trying to include everything you've ever done.

Before you start, there are a few important things to think through.

Get clear first: a resume tells people three things

First, why do we write a resume?

Who are you? Show your personal brand and what sets you apart.
What can you do? Highlight your core skills and professional abilities.
What have you achieved? Prove your impact with concrete results and numbers.

If it doesn't fit one of these three themes, cut it

The whole resume should be built around these three themes. Anything outside those three, cut it.

Early in our careers, we may need more supporting evidence to help an interviewer understand who we are and improve our chances. As we gain experience and pursue more senior roles, being concise is actually the better strategy.

Break it down and do it in stages

Break the process into smaller pieces and work through it in stages. You'll find it isn't that hard.

Haven't updated your resume in a while either? Pick up the pen now and polish this calling card.

3 Resume / CV Mistakes That Get You Screened Out in the First Round

2026-09-30T10:00:00.000+08:00

Resumes / CVs most often die in the first round. These are the three mistakes I see most often:

Listing only what you did
Not answering the question
Being too honest

1. Listing only what you did

"I held this many meetings."

"I led several trade shows."

"I compiled the project's regular reports."

That's what you did, not what you achieved. Take the same experience: "ran trade shows" versus "brought in X leads from one show and closed Y deals." Those read like two resumes at different levels.

Here's a counterexample. I recently saw a resume from a young candidate where every bullet carried two or three percentages: +25%, +40%, −60%, all the way down the page. The numbers looked suspiciously neat, and the whole thing looked like it had come from a template. Having no numbers loses you points. Numbers that look fake lose you even more.

2. Not answering the question

The job description (JD) is in English

→ You send a resume in Chinese

The JD asks for "3+ years of management experience"

→ You write "led a five-person project team"

Project management and people management are two different things.

Some mistakes are less obvious: two versions of your resume with two different email addresses, or a whole paragraph pasted in twice without you noticing. These won't get you rejected, but they make people wonder whether you check your work before you hand it in.

3. Being too honest

The company offers one work-from-home day a week

→ You say "I'm looking for a remote job"

What HR is thinking: then can you handle the other four days in the office?

"My manager and I didn't get along, so I want to leave."

→ What the hiring manager thinks: I'm not perfect either. Will I be the next manager you don't get along with?

I have actually seen someone write "I still want to go back to working inside an organization" on a resume. That tells the employer outright: the situation I'm in now wasn't my choice. Salary negotiations haven't even started, and you've already undercut yourself.

A resume isn't an autobiography. It's an exam.

Work out what kind of person this company is looking for, then answer.

You're not being asked to write an essay. You're being asked to get the answer right.

This post in Chinese: https://www.uncleric.com/2026/08/three-cv-mistakes.html

I'm Eric Lu ("Uncle Eric"). I write about food, travel, and careers.

3 Common Resume / CV Problems I See When Reviewing Friends' Drafts, and How to Fix Them

2026-09-30T10:00:00.000+08:00

I've recently reviewed a few friends' resumes / CVs, and the same problems kept coming up. Here they are.

Update your resume regularly

Writing a resume isn't easy, especially when you're rushing it because someone just asked you for one, whether that's a headhunter or a friend who's going to refer you internally. In just two pages, you need to show what sets you apart while clearly presenting your achievements and abilities. Rush it out in a day or two, even with AI helping, and it usually shows.

The best approach is to update it regularly. Every time you step back and reflect, you'll have new ideas and get a better sense of how to make yourself stand out.

If you've been too focused on work and forgotten to update your resume, now's a good time to do it.

Focus on achievements

Most of us in Taiwan are fairly modest and feel awkward writing about our own contributions. (Some candidates go the other way, and what they write may sometimes need to be taken with a grain of salt.)

HR and hiring managers read a lot of resumes every day, and they really only have a few minutes for each one. If yours runs long and reads like a log of everything you did, it's hard to pick out the key points, and it's easy for them to lose patience and skip it. That's a wasted opportunity.

Write three to five points per job at most. (If you have more, look closely at whether you've written too many things that don't matter.) When you write them, start with the result and back it up with clear numbers. For example:

This project took sales from X to Y.
This feature optimization reduced system costs from X to Y.

That's concrete and convincing, and far better than a long-winded description.

The STAR method

Storytelling is a valuable communication skill. How do you make your point short and clear? Use STAR.

Situation: the context in which this happened.
Task: the task you needed to take on or initiate because of that situation.
Action: how you got the task done, especially your solutions.
Result: the results you (and your team) achieved through that work. Remember the point above: back it up with numbers.

Once STAR becomes a habit, you can lay out what you want to say quickly and in order. It leaves a strong impression on interviewers, which is a big plus.

Final thoughts

The faster things change, the more you need a calling card you can hand over at any moment. I hope you all land your dream job.

Want to know how a resume dies in the first round once it's sent? I wrote up the three most common ways here: 3 Resume / CV Mistakes That Get You Screened Out in the First Round.

You don't lack answers. You just haven't been asked the right question yet.

This post in Chinese: https://www.uncleric.com/2026/08/cv-resume-three-common-issues.html

I'm Eric Lu ("Uncle Eric"). I write about food, travel, and careers.

Job Hunting After a Career Gap? Start With These 3 Resume / CV Improvements

2026-09-30T10:00:00.000+08:00

A few days ago I was talking with someone I'd just met. He'd been out of work for a few months, was starting to worry, and asked whether there was a way to land his next job quickly.

I asked to see his resume / CV. It had the problem I see all the time: only a few words about each job.

HR will take one look and skip it.

If you're still getting the hang of writing a resume, there's a simple trick. Don't try to do everything at once. Give yourself a small goal: improve three things on your resume.

Start with these three:

Who you are: explain it clearly, briefly and to the point.
Use strong words: led, built, launched.
Focus on achievements and results, not just what you did.

Fix a little each day, and after a week your resume will be noticeably better. If HR schedules you for a first-round interview, you've cleared the first hurdle. Congratulations.

Lanjia Gua Bao, Gongguan: Five Ways to Choose Your Pork, and a Bowl of Si Shen Soup

2026-09-29T10:00:00.000+08:00

This is a place I've been going to since my university days. Lanjia Gua Bao sits by the university gate in Gongguan, Taipei, and it made the Michelin Bib Gourmand list three years running, from 2019 through 2021.

Gongguan used to be a bit of a food desert. Back then we all thought Shida Night Market had a better food scene, especially that one lane lined with international restaurants. Lanjia is one of the few from those days still standing.

Five ways to choose your pork, and it was like that even back then

A gua bao is a folded steamed bun, and the filling is up to you: all fat (肥肉), all lean (瘦肉), half and half (半肥半瘦), mixed leaning lean (綜合偏瘦), or mixed leaning fat (綜合偏肥).

At my age I go for mixed leaning lean, plus a bowl of Si Shen Soup (四神湯), Taiwanese four-herb soup with pork intestine.

It still tastes the way I remember

The mixed leaning lean I got this time has less fat in it than what used to count as lean here. The soup tastes exactly the way I remember, steady as ever. I get the mixed one, intestine and tripe, for more texture.

My own favorite Si Shen Soup is still Huang Ji, in front of the Cisheng Temple in Dadaocheng, with two swirls of rice wine poured over the top.

Uncle's rating: Overall 4 | Food 4 | Atmosphere 4.5
They have kept the place clean for years, which earns the extra half point. Ceramic bowls would make it better still.
My picks: a gua bao, mixed leaning lean, and the Si Shen Soup with intestine and tripe

The details

Lanjia Gua Bao (藍家割包)
Address: No. 3, Alley 8, Lane 316, Section 3, Roosevelt Road, Zhongzheng District, Taipei (Gongguan Night Market)
Awards: Michelin Bib Gourmand (2019, 2020, 2021)
Google rating: 4.0 (7,094 reviews, as of September 25, 2026)
Map: https://maps.google.com/?cid=10180140947890385693&hl=en

This post in Chinese: https://www.uncleric.com/2026/08/lanjia-gua-bao-gongguan.html
More award places I've eaten: Michelin Bib Gourmand and 500 Bowls

I'm Eric Lu ("Uncle Eric"). I write about food, travel, and careers.

換 Opus 5.5 前先做三件事：prompt-audit、migrate、免費重置額度

2026-09-28T15:48:33.748+08:00

換到 Opus 5.5 之前，先做這三件事。

第一件，在 Claude Code 裡打：

/claude-api prompt-audit

它會檢查 CLAUDE.md、rules、skills、commands、agents 這些提示檔，列出替舊模型寫、或彼此打架的句子，附檔名、行號和建議改法。沒有你同意，它不會改檔。

大叔拿自己的專案跑了一次，找到 6 處高信心的矛盾或過期規則。例如一個檔寫流程已經自動化，另一個檔還叫它手動做。

大叔的專案跑出來的報告摘要。

第二件，程式裡有呼叫 Claude API 的，打：

/claude-api migrate this project to claude-opus-5-5

它會換 model ID、處理官方列出的不相容項目。動手前先確認要改哪些檔，改完列出要人工驗證的地方。

prompt-audit 我們在 Claude Code 2.1.283 實跑過，它還沒列在官方的 skill 介紹頁，找不到這個指令就先更新 Claude Code。migrate 官方文件有寫。

第三件，有訂閱的到設定的 Usage 頁，按 Resets 那格的「Reset for free」，額度重置一次。可以先留著，等哪天快用完再按，大叔的帳號顯示 10/23 到期。

設定 → Usage → Resets。

Opus 5.5 在 9/22 上線。對 Opus 5，單價只降兩成，官方卻說整體成本少四成；對 Fable 5.1，輸入輸出單價直接四折。公告原文：

It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.

價格

每百萬 token，官網定價：

模型	輸入	輸出	快取讀取
Fable 5.1	10 美元	50 美元	0.25 美元
Opus 5.5	4 美元	20 美元	0.20 美元
Opus 5	5 美元	25 美元	0.50 美元
Sonnet 5	2 美元	10 美元	0.20 美元

單價降兩成、成本降四成，中間的差距只能來自用量。假設輸入輸出的比例不變、先不算快取，0.6 ÷ 0.8 ＝ 0.75，同一個任務計費的 token 大約少兩成五。公告沒有公布成本拆解，這是推算，不是官方數字。
快取讀取從 0.50 美元降到 0.20 美元，降幅六成，比單價的兩成大得多。重複 context 多、快取命中率又高的工作，從 Opus 5 遷移過來，實際降幅可能超過兩成。
對 Fable 5.1，快取讀取只便宜兩成，輸入輸出便宜六成。

預設 effort 是 medium

Fable 5.1 預設 high，Opus 5.5 預設 medium。拿兩者的預設值直接比，比到的其實是不同的思考深度。官網建議的順序也是這樣：Opus 5.5 把 effort 調高之後還不夠，再換 Fable 5.1。

從 Opus 5 遷移會壞的地方

思考模式關不掉，深淺只能用 effort 控制。
forced tool use 直接回錯誤。
思考區塊綁定產生它的模型與對話，不能搬到別的模型接著用。
Claude API 與 Google Cloud 不再接受 computer_20251124。

前三條在 Fable 5.1 也適用，已經相容 Fable 5.1 的程式通常改動較少，但從 Opus 5 遷移的仍要逐項檢查。

還有一個不會報錯、但會讓畫面變安靜的改動：工具呼叫之間的文字改放進 thinking 區塊，預設 display 下內容是空的。如果你的產品會把這段文字當成進度訊息顯示給使用者，升級後要另外設定 display，不然畫面上只剩工具呼叫，沒有中間的說明。

模型怎麼分工

官方的選型表把多數工作的起手檔換成 Opus 5.5。跑好幾個小時的 agent、多步驟的深度研究，要最高能力時才考慮 Fable 5.1；日常的程式生成、資料分析交給 Sonnet 5；即時回應和子代理的小任務交給 Haiku 4.5。

換模型之前，官方建議先做一件更便宜的事：在現在用的模型上試幾個 effort 級別，拿自己的評測對照。官方原話是調 effort 常常是比換模型更好的槓桿。Anthropic 的測試裡，Opus 5.5 在 medium 追平或超過 Opus 5 在 high 的表現；幾個 coding 評測上，Opus 5.5 的 low 也接近它自己的 medium，成本低很多。

要混用模型，官方給了兩種組法：

顧問式：Opus 5.5 或 Sonnet 5 當執行者跑完整個流程，遇到難的決定才去問 Fable 5.1。官方實測 Opus 5.5 high 配 Fable 5.1 顧問，比 Opus 5.5 單跑高 1.7 個百分點、成本約 2.1 倍，官方自己也註明這個差距在雜訊邊緣。划不划算看執行者會不會開口問：effort 調太低的執行者察覺不到自己卡住，就不問了，分數反而比單跑低。
指揮式：Fable 5.1 拆工作、派給一批 Sonnet 5 代理去做，這些代理用 low effort。適合能切成大量、彼此獨立的工作，例如要讀的資料超過一個 context。工作環環相扣，或單一模型的 context 就裝得下，官方那次實驗是單一模型調低 effort 比較划算。

訂閱方案

Pro、Max、Team 與按席位計費的 Enterprise，五小時用量上限調高；額度重置怎麼按，見開頭第 3 點。

官方公告：https://www.anthropic.com/news/claude-opus-5-5
模型規格：https://platform.claude.com/docs/en/models/opus-5-5/overview

平常在玩的工具跟自造筆記，都收在大叔的工具箱：https://www.uncleric.com/2026/08/uncle-toolbox.html

Wang's Broth, Huaxi Street: A Bib Gourmand Braised Pork Rice That Fell Short

2026-09-28T10:00:00.000+08:00

I was nearby and stopped by Wang's Broth. It's a braised pork rice shop on Huaxi Street in Wanhua, Taipei, open since 1975, and it holds both a Michelin Bib Gourmand and a 500 Bowls Award. A big red lantern by the door says "Michelin Recommended," so I walked in with my expectations up.

It was a mild letdown.

I got the braised pork rice, a fried tofu puff, and a braised duck egg.

The order slip.

How it tasted

The rice looked the part. Dark and glossy, the sauce clinging to the cubes of fat and lean, exactly what a famous bowl is supposed to look like. It tasted salty. Not oily, but a bit cloying by the second half. Rich, not sticky. I wanted a soup to cut through it. I hadn't ordered one, because the afternoon was too hot to face a hot bowl.

The fried tofu puff looked glossy and promising, then came up short. It hadn't absorbed much of the sauce, and it was on the dry side. The braised duck egg had the same problem. It was nearly 2 pm; if it had been in that tray since morning, that should have been long enough.

Without the soup, this rating isn't complete. A reason to come back. Maybe.

Uncle's rating: Overall 3.5 | Food 3.5 | Atmosphere 4
Clean and comfortable by night-market standards, with indoor seating, though the air conditioning wasn't keeping up with the heat that afternoon.
What I'd try next time: clear broth with minced pork and pickles (清湯瓜仔肉)

A Bib Gourmand plaque doesn't make every dish land. Wanhua has plenty of other street food to work through, so I'll go find another bowl of braised pork rice.

The details

Wang's Broth (小王煮瓜)
Address: Stall 153, No. 17-4, Huaxi Street, Wanhua District, Taipei (Huaxi Street Night Market)
Awards: Michelin Bib Gourmand (2026), 500 Bowls Award (2026, one bowl)
Google rating: 4.1 (14,445 reviews, as of September 25, 2026)
Map: https://maps.google.com/?cid=17975734638609170553&hl=en

This post in Chinese: https://www.uncleric.com/2026/07/35.html
More award places I've eaten: Michelin Bib Gourmand and 500 Bowls

I'm Eric Lu ("Uncle Eric"). I write about food, travel, and careers.

Letian Mutton Soup by Nanmen Market: Garlic Mutton Ribs, and NT$10 for a Bowl of Rice

2026-09-27T10:00:00.000+08:00

I walk around Nanmen Market all the time, and somehow it took me until now to notice a mutton soup shop near the MRT exit. It holds two bowls in the 2026 500 Bowls Award.

I got there before the mealtime rush, so there was no line. The room is small and very tidy: wood-framed windows, and a wooden sign by the door.

No. 25, Section 1, Roosevelt Road. Wood-framed windows, and a wooden sign.

First visit, so I played it safe: the garlic mutton rib soup at NT$160, plus a bowl of their Chunfa braised pork rice (純發滷肉飯). Add NT$10 to any soup order and you get that bowl of rice. Nice of them.

There are three broths to choose from: angelica, clear broth, and garlic. Each comes with either ribs or sliced meat.

The braised pork rice: a fatty cut that doesn't sit heavy

They use a fatty cut for the braised pork, and when it landed I wondered whether it would be too rich. It wasn't. It works with the plain rice, and the fat doesn't coat your mouth.

If you eat chili, try theirs. It has garlic mixed into it. Go easy, because it creeps up on you.

They use a fatty cut for the braised pork.

The pot at the back, with a big piece of skin-on pork sitting in the sauce.

The braised egg: this one passes

After a Bib Gourmand place in Wanhua served me a braised egg that hadn't taken any of the sauce, I order one wherever I see one. I cut this one in half and the flavor had gone all the way through. This one gets a pass.

Braised pork rice with a braised egg.

The garlic mutton rib soup: this is the one

Now the main event. The bowl comes loaded with garlic. The broth isn't sharp at all, just a mellow garlic warmth. The ribs are stewed soft, skin on, and still have a bit of bounce to them.

This is a level above the mutton place I tried in Xindian. I'll be back.

Ribs stewed soft, skin on. The broth stays clear but carries the garlic.

The room

The dining room is small but very clean. You get chopsticks, a stainless-steel spoon, and ceramic bowls. That already earns them points in my book.

The menu and prices.

Uncle's rating: Overall 4.5 | Food 4.5 | Service 4.5 | Atmosphere 4
My picks: the garlic mutton rib soup, and the braised pork rice with a braised egg

Nanmen Market already had the food court on its second floor. Now the block has one more place worth walking to.

The details

Letian Mutton Soup (樂天羊肉)
Address: No. 25, Section 1, Roosevelt Road, Zhongzheng District, Taipei
Awards: 500 Bowls Award (2026, two bowls)
Google rating: 4.5 (1,350 reviews, as of September 25, 2026)
Map: https://maps.google.com/?cid=7516697120047386825&hl=en

This post in Chinese: https://www.uncleric.com/2026/09/letian-mutton-soup-nanmen.html
More award places I've eaten: 500 Bowls and Michelin Bib Gourmand

I'm Eric Lu ("Uncle Eric"). I write about food, travel, and careers.

Uncle Eric's Table: Michelin Bib Gourmand Places I've Tried

2026-09-24T16:30:41.000+08:00

These are the Michelin places I've tried and written about. So far they're all Bib Gourmand picks; I'll add starred restaurants when I get to them. Each one shows the latest year it was listed, and places that have since dropped off keep the last year they made it.

I keep a 500 Bowls Award list too. This page in Chinese: 米其林必比登清單.

Taipei

Hsiung Chi Scallion Pancake (雄記蔥抓餅)
Bib Gourmand 2026 · 500 Bowls 2025 (one bowl) · visited Sep 2026
A scallion pancake stall next to Shuiyuan Market. Get the signature with corn (NT$65), add the free basil, and ask for both sauces. Soft and chewy. My rating: 4.5.

Wang's Broth (小王煮瓜)
Bib Gourmand 2026 · 500 Bowls 2026 (one bowl) · visited Jul 2026
On Huaxi Street since 1975. The braised pork rice looked promising, but I found it salty and a bit heavy, and the tofu and duck egg hadn't absorbed much flavor. I skipped the soup because it was a hot afternoon, so this rating leaves it out. My rating: 3.5.

Lanjia Gua Bao (藍家割包)
Bib Gourmand 2021 (listed 2019 to 2021) · visited Aug 2026
Braised pork buns by the university gate in Gongguan, a stall I've been eating at since college. You pick how fatty you want the pork, in five steps from lean to fatty. I usually get a mixed filling on the leaner side, plus a bowl of Si Shen Soup (four-herb soup) with intestine and tripe. My rating: 4.

Years follow the official Michelin Guide lists, checked in September 2026. Prices and details are from when I visited, so check before you go.

Uncle Eric's Table: 500 Bowls Award Places I've Tried

2026-09-24T16:29:41.000+08:00

The 500 Bowls Award is a Taiwanese street food award from the publication 500輯 and rates each place one, two, or three bowls. These are the ones I've tried and written about, with the latest year and bowl rating for each.

The full list is on the official 500 Bowls site (in Chinese). I keep a Michelin Bib Gourmand list too. This page in Chinese: 500 碗清單.

Taipei

Letian Mutton Soup (樂天羊肉)
500 Bowls 2026 (two bowls) · visited Sep 2026
Near the MRT exit by Nanmen Market. Get the garlic mutton rib soup, which has a mellow garlic flavor and tender, skin-on mutton. Add NT$10 and you get a bowl of braised pork rice too. Their garlic chili sauce has a kick. My rating: 4.5.

Wang's Broth (小王煮瓜)
500 Bowls 2026 (one bowl) · Bib Gourmand 2026 · visited Jul 2026
On Huaxi Street since 1975. The braised pork rice looked promising, but I found it salty and a bit heavy, and the tofu and duck egg hadn't absorbed much flavor. I skipped the soup because it was a hot afternoon, so this rating leaves it out. My rating: 3.5.

Hsiung Chi Scallion Pancake (雄記蔥抓餅)
500 Bowls 2025 (one bowl) · Bib Gourmand 2026 · visited Sep 2026
A scallion pancake stall next to Shuiyuan Market. Get the signature with corn (NT$65), add the free basil, and ask for both sauces. Soft and chewy. My rating: 4.5.

Years and bowl ratings follow the official 500 Bowls lists, checked in September 2026. Prices and details are from when I visited, so check before you go.

大叔の餐桌｜米其林必比登：大叔吃過、寫過的店

2026-09-24T16:14:09.000+08:00

這份清單收的是大叔吃過、也寫了食記的米其林店家，目前都是必比登，星級餐廳吃到再補。每家標出最近一次入選的年份；已經不在目前名單上的店，也保留它入選過的紀錄。

想看其他探店和自己動手的菜，回到大叔の餐桌。同樣整理過的還有500 碗清單。

English version

台北

公館・長江一號雄記蔥抓餅｜跟著洋流排隊，招牌一定要加玉米
必比登 2026｜500 碗 2025・一碗｜食記 2026 年 9 月
水源市場旁的蔥抓餅攤。招牌加玉米，九層塔免費一定要加，蜂蜜芥末醬和店家自己炒的辣椒醬也都加，餅皮鬆軟有嚼勁。整體 4.5。

華西街・小王煮瓜｜必比登加持的滷肉飯，大叔私評 3.5
必比登 2026｜500 碗 2026・一碗｜食記 2026 年 7 月
1975 年開到現在的必比登老店。滷肉飯上桌黑金發亮，賣相是名店的樣子，入口卻偏鹹又稍膩。大叔私評只給 3.5，必比登也是會翻車的。

公館藍家割包｜從大學吃到現在的必比登老店
必比登 2021（2019 到 2021 連三年）｜食記 2026 年 8 月
大學口的藍家割包連著三年入選必比登推薦。內餡可以選肥肉、瘦肉、半肥半瘦、綜合偏瘦、綜合偏肥五種比例，配一碗四神湯，是記憶裡那個很穩定的味道。十八年前的初訪也還在：公館小吃藍家割包（2008 初訪）。

⚠️ 入選年份依米其林指南官方公告，整理至 2026 年 9 月。價格與店況都是食記當時的紀錄，出門前請再查一次。

大叔の餐桌｜500 碗：大叔吃過、寫過的店

2026-09-24T16:13:11.000+08:00

500 碗是《500輯》辦的台灣小吃評選，分一碗、二碗、三碗。這份清單收的是大叔吃過、也寫了食記的店，每家標出最近一次入選的年份和碗數。

完整名單在 500 碗官網。想看其他探店和自己動手的菜，回到大叔の餐桌。同樣整理過的還有米其林必比登清單。

English version

台北

南門市場・樂天羊肉｜500 碗，蒜頭羊排骨湯配一碗 10 元滷肉飯
500 碗 2026・二碗｜食記 2026 年 9 月
南門市場捷運出口旁。蒜頭羊排骨湯滿滿的蒜頭，湯頭不嗆是溫潤的蒜香，羊排燉到軟還帶皮有 Q 度；點湯品加 10 塊就多一碗滷肉飯，偏肥的部位配白飯意外地搭。整體 4.5。

華西街・小王煮瓜｜必比登加持的滷肉飯，大叔私評 3.5
500 碗 2026・一碗｜必比登 2026｜食記 2026 年 7 月
1975 年開到現在的必比登老店。滷肉飯上桌黑金發亮，賣相是名店的樣子，入口卻偏鹹又稍膩。大叔私評只給 3.5，必比登也是會翻車的。

公館・長江一號雄記蔥抓餅｜跟著洋流排隊，招牌一定要加玉米
500 碗 2025・一碗｜必比登 2026｜食記 2026 年 9 月
水源市場旁的蔥抓餅攤。招牌加玉米，九層塔免費一定要加，蜂蜜芥末醬和店家自己炒的辣椒醬也都加，餅皮鬆軟有嚼勁。整體 4.5。

⚠️ 年份與碗數依 500 碗官方名單，整理至 2026 年 9 月。價格與店況都是食記當時的紀錄，出門前請再查一次。

Hsiung Chi Scallion Pancake, Gongguan: Get the Signature with Corn, and Don't Skip the Basil

2026-09-24T16:01:59.000+08:00

I've walked past Hsiung Chi for years and almost never bought one. It's a scallion pancake stall next to Shuiyuan Market, inside Gongguan Night Market in Taipei. It has won both a Michelin Bib Gourmand and a 500 Bowls Award. It hit the local news, the line got longer, and I finally joined the crowd.

The Michelin plaque and the 500 Bowls sign on the stall.

To be honest, this style of pancake isn't really my thing, though I don't dislike it. I grew up in southern Taiwan, and the one I miss is a different pancake: fried in oil until the edges crisp, then topped with a runny egg and the shop's own sweet chili sauce.

The line

On the day I went, the stall didn't open until 4 pm, and the line already looked like this. I've never seen a queue like this around Shuiyuan Market.

The line at Shuiyuan Market.

I hate waiting in line, but I stuck it out. Halfway through, the auntie who runs the stall came out to tell us why it was taking so long: the people ahead had ordered a lot. Nice of her to let us know. After what felt like forever, it was my turn.

The spatulas on the griddle never stop.

What to order

I went with the signature, which comes with egg and cheese (NT$60), and added corn for NT$65 total. The menu board has English too. They call it an "Onion Pancake." Or you can just point at 招牌 (signature) or 招牌+玉米 (with corn). Basil is free, and it'd be a crime not to add it.

The menu board. Basil is free (九層塔免費).

The pancake is soft and chewy, with a light, bready aroma, and the cheese melts into the sweet corn and egg. It comes off the griddle hot enough to burn your tongue, so be careful. It's a thick one, but I polished it off in no time on the bus.

The signature with corn.

There are two sauces on the board: honey mustard, and a chili sauce they make in-house. Get both. For the chili, mild is enough; it has a slow-building kick.

It's a lot better than the scallion pancake stall near Yongkang Park, and I'll be back. Skipping it all these years was a big mistake.

Uncle's rating: Overall 4.5 | Food 4.5 | Service 4.5
My picks: the signature with corn, plus both sauces

The details

Hsiung Chi Scallion Pancake (雄記蔥抓餅)
Address: No. 2, Lane 108, Section 4, Roosevelt Road, Zhongzheng District, Taipei (inside Gongguan Night Market)
Awards: Michelin Bib Gourmand (2026), 500 Bowls Award (2025, one bowl)
Google rating: 4.2 (992 reviews, as of September 23, 2026)
Map: https://maps.google.com/?cid=1080324372144503497&hl=en

This post in Chinese: https://www.uncleric.com/2026/09/hsiung-chi-scallion-pancake-gongguan.html

I'm Eric Lu ("Uncle Eric"). I write about food, travel, and careers.

公館・長江一號雄記蔥抓餅｜跟著洋流排隊，招牌一定要加玉米

2026-09-24T10:23:39.000+08:00

洋流大爆發，大叔也跟風來了

水源市場旁的長江一號雄記蔥抓餅，最近洋流大爆發，大叔也跟風來探店。看到報導，阿姨為國家犧牲這麼大，一定要來支持一下。

先承認一件事：
雖然阿姨的雄記蔥抓餅在水源市場旁賣了這麼久，
雖然大叔從念書就在這一帶走跳，
雖然知道他們家有了米其林必比登、五百碗，
大叔其實很少來，主要是：
蔥抓餅其實不太是大叔的菜（也沒有不喜歡啦）。
要那種油鍋炸（煎）出來，加半熟蛋，再抹上店家自製甜辣醬的焦香蔥油餅，才是南部小孩念念不忘的味道。

攤上掛著米其林的牌子，旁邊貼著 500 碗。

阿姨 20 多年前是「少康專案」的關鍵牽線人。沈伯洋公館掃街時，阿姨招手請他吃了一塊餅，才引發了這一連串的排隊風潮。

排隊：等到天長地久

這天傍晚四點才開攤，你看就排成這麼長。

水源市場旁的隊伍。

大叔最不喜歡排隊了，今天拚了。水源市場旁能排成這樣，我真的從來沒看過。經過了天長地久，終於輪到了。

阿姨中途還特地來跟排隊的朋友們說明，今天前面的客人買得很多，所以才這麼久，好客氣。

先由招牌試起，一定要加玉米。九層塔免費，不加簡直是犯罪行為。看著煎台上鍋鏟飛舞，完全停不下來。

煎台上的鍋鏟，完全停不下來。

菜單與價位，九層塔免費。

但這個蔥抓餅真的不賴。餅皮鬆軟，有 Q 度有嚼勁，帶著淡淡的麵香。融化的起士混著甜甜的玉米蛋，吃的時候小心，剛起鍋還會燙舌頭。還在坐公車的路上，一下子就把整張厚厚的餅嗑光了。

招牌加玉米。

這個比永康公園那家好吃多了，會再訪。

蜂蜜芥末醬也要加。辣椒醬選小辣就好，他們家自己炒的，有後勁的。

大叔私評：整體：4.5｜餐點：4.5｜服務：4.5
推薦菜色：招牌蔥抓餅加玉米（醬都要加）

竟然錯過這攤這麼久，大失誤。下回我還要試試沈伯洋的全餐。

店家資訊

店名：雄記蔥抓餅
地址：台北市中正區羅斯福路四段 108 巷 2 號（公館夜市內）
Google 評分：4.2 星（992 則，2026-09-23）
菜單與價位直接看上面的圖

攤上貼的黛黛茶合作海報。

同樣在公館、大叔從大學吃到現在的必比登，還有這一家：公館藍家割包｜從大學吃到現在的必比登老店

English version

No Bot, Just Physics: The Mac Meeting Recorder I Built and Open-Sourced

2026-09-20T10:00:00.000+08:00

MyParrot records each side of a Mac call into its own audio channel, so nothing has to guess who said what, and no bot joins the meeting. Free, MIT licensed, signed and notarized: https://github.com/tingyulu/MyParrot

Recording meetings had me stuck for a long time.

Most days I'm on Zoom or Google Meet, coaching people on their careers, interviewing candidates, and sitting in product meetings as the consultant in the room. After a call I often want the transcript: to check a detail, or to pull together the advice I owe the other person. It's a simple need, and every tool I tried had one thing I couldn't live with.

Two approaches, two costs I didn't want to pay

Meeting recorders mostly work one of two ways.

The first sends a bot into your meeting. It records and transcribes for you, which is convenient. For my work it has one fatal flaw: the client opens the call and finds a stranger in the participant list called "Something Notetaker." That costs trust before anyone has said a word. They're about to talk to me about their career, sometimes about their most private worries, and an unknown recorder is sitting in the room. Some meetings don't let bots in at all.

The second skips the bot and records the whole call as one track. Once everyone's voice is mixed together, "who said this?" is left to the software to guess. When it guesses wrong, a whole paragraph lands on the wrong person. Reading the transcript afterward, I first had to work out which lines were mine and which were theirs.

One costs trust, the other costs accuracy, and I didn't want either.

If a bot is out, how about physics?

After a while, something clicked.

The other person's voice comes out of my Mac as system audio. My voice goes in through the microphone. Those two signals are separate to begin with.

So why wait for them to be mixed into one track, then spend effort guessing who's talking? Record each into its own channel instead: them on the left, me on the right. The two sides stay apart from the moment they're captured, with no bot in the call and no model trying to tell voices apart. The "who said this?" problem gets sidestepped by a physical fact.

Once I saw that, I handed the idea to Claude Code. Half a day later I had a working prototype, and I called it MyParrot.

The main window, with the transcript labeled by channel: Them and You. Every meeting in these screenshots is a fictional demo voiced by macOS text-to-speech, not a real client call.

What it does today

MyParrot is at v0.3.1 now.

The other side is captured with Core Audio's Process Tap, available since macOS 14.4. There's no driver, kernel extension, or virtual audio device to install, and it never asks for Screen Recording permission.
Your side comes from the microphone. MyParrot watches for mic and audio-route changes and rebuilds the capture on its own, so if you swap mics mid-call, the file stays continuous.
A live transcript appears while you record. You choose the engine: Apple's SpeechAnalyzer on macOS 26, or a fully local Whisper engine (whisper.cpp on Metal, large-v3 family). When you stop, it re-transcribes at higher precision.
Exports include a plain .txt and a timestamped .srt where every line starts with who said it.
Recording and transcription stay on your Mac. The app only goes online for an optional update check against GitHub Releases, which you can switch off, and for Whisper model downloads you start yourself.

Whisper was the upgrade I cared about most. I switch between Mandarin and English constantly, and Apple's built-in recognizer often stumbles on that. Whisper does noticeably better, and it runs entirely on your machine. It still makes mistakes, and I expect better on-device engines to arrive soon.

The bug I never thought to spec

One edge case caught me completely off guard.

If a Bluetooth headset connected halfway through a recording, macOS switched the default output device. An early build kept recording, but the other side's track came out blank, with no warning at all. You thought you had the call. You played it back and the other person was gone.

MyParrot now detects the switch and rebuilds the capture on its own, so you don't have to restart the recording.

A related tip: listen through a Bluetooth headset if you like, but don't talk into its mic. The moment any app opens a Bluetooth microphone, the headset drops to a narrowband call profile, and the other side's audio falls to phone quality too. Speak into the built-in or a wired mic. MyParrot's automatic mic selection already skips Bluetooth mics.

A month of daily use later

Once recording was reliable, my problem moved from "did it record?" to "what do I do with it now?" The painful part of a one-hour recording isn't getting a transcript. It's finding the three minutes you need when you don't know where to start listening.

Version 0.3.0 went after that.

A waveform you can read. Playing a recording draws a stereo waveform right in its row: the top half is them, the bottom half is you. Because the file is already split by speaker, the waveform shows who spoke when, with no guessing. Click anywhere to jump there.

It also caught one of my own mistakes. One recording showed sound for the first quarter and a flat line after that. I had finished the call, forgotten to press stop, and recorded more than eighty minutes of nothing. Before, I would only have found out by listening.

Shortcuts printed on the buttons. ⌘R starts, ⌘P pauses, ⌘E stops, ⌘T transcribes. There's nothing to memorize, because each key is on its button. ⌘1 cycles through three window sizes, down to a floating mini bar you can park in a corner during a call.

A level meter that tells the truth. The old meter was linear, so normal speech lit two or three segments and looked dead. More than once I thought my mic was broken. It now uses a dB scale: speech lights more than half the bar, silence stays at one segment. I calibrated it against the noise floor of my own Yeti mic, so on other mics the exact count will differ.

One meter per side, so you can see that both the other person and your own mic are coming through.

In mini mode you still see the timer, both meters, and the pause and stop buttons, without the full window in the way.

Limits worth knowing

Mac only, macOS 14.4 or later. Live transcription works best on macOS 26.1 or later.
In meetings longer than about 30 minutes, the two capture clocks slowly drift apart. Sample-accurate sync is planned.
Echo cancellation is an offline post-process, off by default, and it writes a separate file without touching the original.
Recording a call without consent is illegal in many places. Tell people you're recording, and get consent where the law requires it.

Why open source

Two reasons.

The first is privacy. Recording and transcription run entirely on your Mac, Whisper included, so your calls don't go to anyone's server. With the code in the open, anyone can check how it handles your audio.

The second is that I doubt I'm the only one with this problem. If you spend a lot of time in meetings, want notes you can actually use, and would rather not send your calls to a cloud service you know nothing about, give it a try.

Download

MIT license, free to use. The app is signed and notarized, so it opens without a fight with macOS security settings. Grab the latest .zip from Releases, unzip it, and drag MyParrot.app into Applications.

Repo: https://github.com/tingyulu/MyParrot
Download: https://github.com/tingyulu/MyParrot/releases/latest

I record my own coaching sessions, candidate interviews, and product advisory calls with it every day. If you hit a problem, tell me in a GitHub issue.

Ideas come in unlimited supply. Tools that actually ship don't.

I'm Eric Lu ("Uncle Eric"), a product consultant, headhunter, and career coach. MyParrot is the recorder behind my own calls.

出門工作不降級：飯店房間裡的四螢幕行動辦公室

2026-09-19T11:00:00.000+08:00

在家兩三個螢幕排開，什麼都攤在眼前。出了門只剩一台筆電，同樣的工作就得一直切視窗，切著切著就忘了誰跑到哪一步。

這次住 Hyatt Place，我在房間裡把行動辦公室升級成四個畫面。會排成這樣有個很實際的理由：手上同時有好幾個 AI 在跑，螢幕不夠真的會漏事情。

四個畫面，各管一件事

飯店的電視：監看牆。

電視不適合當工作的主螢幕，距離遠、角度也不對，接了傷眼又傷脖子。但它就擺在那裡，不用也可惜。所以它不幹活，改當監看牆：正在跑的東西全部攤在上面，我不必盯著它，餘光掃一眼就知道哪個停了、哪個還在動。

MacBook Pro：主工作區。

右邊 ASUS ZenScreen Touch：第二螢幕。

隨身型的，收起來跟一本書差不多厚。它負責接住「正在讀的那份東西」：資料、文件、要對照的稿子。有觸控，偶爾直接用手滑比翻回筆電快。

左邊 iPad：儀表板。

固定開著數據儀表板，不參與工作，只負責被看。它的價值在於「不必開」，數字一直在那裡。

真正改變的地方

以前是我在等機器跑完。現在倒過來，AI 在等我看完。

這件事第一次意識到的時候有點意外。它們跑得比我看得快，於是瓶頸從機器移到了我的眼睛。多一個畫面換來的，是少問一次「剛剛那個好了沒」。

所以這套配置解的不只是效能問題，是注意力問題。四個畫面不會讓任何一件事跑得更快，但它讓我不必為了確認進度反覆切視窗，每切一次都要多燒一些腦細胞。

代價很實在

桌面一長大，人就更難離開座位。

以前筆電闔上就算下班，現在四個畫面攤在那裡，隨時瞄得到「還有東西在跑」。工具越順手，越要自己定一個下班時間，因為它不會自己叫你停。

另外就是行李。多一塊螢幕就多一份線材與重量，大叔這年紀，背包背不動了～

我自己的配法

帶出門的只有三塊，電視是房間附的。真的少不了的是兩塊：筆電，加那台隨身螢幕。有了第二塊，「正在做的」跟「正在看的」才不必搶同一塊空間。

AI 給方案，我們給判斷。

大家出門時都怎麼配？有沒有什麼好方法可以分享呢？

Eleven Years for Vertical Tabs: Why Good Features Stall in a Market Leader's Backlog

2026-09-18T10:00:00.000+08:00

This July, a Chrome update finally brought vertical tabs to my browser. Google had started rolling them out that April.

Vivaldi had them in 2015. Edge followed in 2021. Arc made the sidebar its core design in 2022. Eleven years later, the browser with the biggest market share finally came around.

Safari on the Mac still has no vertical tab layout. Apple's own documentation lists two layouts, "Separate" and "Compact," and both are horizontal.

The usual explanations go something like this: the audience is too small, the change is too expensive, and extensions can fill the gap.

None of them holds up well here.

Vertical tabs are an option. Nobody is taking horizontal tabs away. If you don't want them, you never turn them on, so nobody's muscle memory gets hurt.

Chrome extensions can't change the tab strip, which is the browser's own territory. So "there's already a workaround" doesn't work either.

And once you have twenty-plus horizontal tabs open (hardly unusual these days), each one shrinks to a tiny icon and you can't tell them apart. Nothing fixes that. At that point the feature simply stops working.

An optional feature with no workaround and an obvious pain point still waited eleven years.

I can't see inside Google, so this is a bold guess: the real answer is much simpler, and much harder to say out loud. Someone with decision rights just didn't like it.

In a big company's backlog, nobody writes "I don't like it." It gets translated into priorities, resource allocation, and risk assessment, then filed in a document that sounds perfectly reasonable. In my experience, the people in the room usually know that isn't the real reason, and nobody needs to say so.

That's how a market leader's blind spots grow, one layer at a time. It's rarely one big mistake. It's a long run of "not now" and "there's something more urgent."

There's a line that gets passed around a lot online. Supposedly Nokia's CEO said it at the press conference announcing the Microsoft acquisition: "We didn't do anything wrong, but somehow, we lost."

I couldn't find an original source for that quote. It may well have been made up.

But it spread for a reason. It hits a nerve: a company can write a sensible document for every decision, run each one through the proper process, and still lose. And there's no next chapter.

Because "we didn't do anything wrong" hides a lot of things that never got written down.

So I'm not that surprised that Google's AI lineup today, Gemini, NotebookLM and the rest, feels bolted together from separate parts. (To be clear, I'm not bashing Google. I still love NotebookLM.)

Have you run into this in your own product? Which feature got blocked, by whom, what reason went on paper, and what happened next?

And in your career, was there a path you didn't take because someone didn't like it?

香港灣仔藍瓶子｜離開香港前，再去三樓窗邊吃份 Brunch Plate 配手沖

2026-09-17T08:30:00.000+08:00

離開香港前，還是想再去一次灣仔的藍瓶子。三樓窗邊擺上電腦，早餐配手沖，這個數位遊牧的早上很可以。

三樓還是一樣寬敞舒服。前一組客人剛離開，我立刻換到窗邊；快中午的陽光曬進來，拿出電腦，邊工作邊吃這頓遲來的早餐。

Brunch Plate

第一次點 Brunch Plate。鬆餅配上香濃鮮奶油，我覺得比淋糖漿好，也不會膩；沙拉給得大方，培根有點厚度，又沒煎到焦，這盤可以。

手沖 House Blend

咖啡當然選手沖，今天喝 House Blend。大家都在拼淺焙的時候，這杯的香氣和厚實感還是有到位。Good，人完全醒了。

啊咧，離開後才想起要補貨蛋糕濾紙……這個年紀，記性都沒了～

大叔私評

大叔私評：整體：4.5｜餐點：4.5｜服務：4｜氣氛：4.5
推薦菜色：Brunch Plate、手沖 House Blend

自助式服務少了一點被照顧的感覺，服務分數給得保守；咖啡好喝，窗邊也坐得舒服，整體還是 4.5。

在香港工作的早上，你會挑哪家咖啡館坐下來？

店家資訊

店名：藍瓶咖啡灣仔店（Blue Bottle Coffee Wan Chai）
位置：香港灣仔聖佛蘭士街 15 號
Google 評分：4.2（395 則，2026-08-30 讀取）
Google Maps：https://maps.google.com/?cid=6688312824928939641&hl=zh-TW

（出了台灣吃過的店和咖啡館，也收在大叔の餐桌。）

Teaching AI to Take Notes Before It Forgets: MyR2D2, 12 Open-Source Claude Skills

2026-09-16T10:00:00.000+08:00

Use AI tools long enough and you end up stuck between two bad options.

Feed it the whole backstory, and you burn through tokens as the context grows, and every reply gets slower. Don't feed it, and it wakes up with amnesia every day. The rules you agreed on, the traps you already fell into, the next step it promised to take: all gone. Tomorrow it leads you straight back into the same mistake.

Claude, ChatGPT and Gemini all offer some kind of memory now, and it helps a little. In my experience, though, a new session still only knows what got written down, and the working details of a task rarely are.

I can't cure the amnesia. So I taught the AI to write itself notes before it forgets. Those notes grew into a set of small tools, and I've open-sourced them as MyR2D2. It's at v0.7.3 now, with 12 skills.

What amnesia actually costs

"AI has no long-term memory" sounds abstract. In practice it looks like this.

You spend an hour with it untangling a problem and agreeing on a plan. The next day you open a new chat and it knows nothing. You explain everything again, it hands you a slightly different version of yesterday's plan, and then it recommends the exact route you ruled out yesterday.

The second failure is sneakier. To avoid the first one, you paste in all the background every time. Conversations get longer and slower, your quota drains, and the three sentences that really matter end up buried under two thousand words of context, where the model can easily overlook them.

So the real question isn't whether to give AI a memory. It's whether there's a third way between burning tokens and losing everything.

My answer: stop asking the AI to remember. Make it write down what matters, on disk, before it forgets.

What's in the box

The 12 skills fall into four groups. The cheat sheet below is the one-page version.

Wrap-up and handoff

save-all: Run it before you close up or reboot. It finds everything that only lives in the conversation, writes it out, then reads it back to confirm it actually landed. It ends with either "safe to restart" or "X still isn't synced."
dropoff: When a task needs to move to another chat, another machine, or another project, it writes a handoff card: where things stand, what's next, the links, the traps. The card assumes the reader knows nothing.
pickup: dropoff's partner. A new session checks for cards addressed to it, reads them in full, claims one, and gets to work. On recent versions of Claude Code, if the receiving session is already open, dropoff can also ring its doorbell.

Work journal

mission-log: Harvests any day's session activity from the transcripts Claude Code already keeps. It costs zero tokens, because it's a plain script with no model involved.
daily-debrief and weekly-debrief: Turn that skeleton into a daily report with a reflection, then roll seven of them into a weekly. Transcripts are cleared after 30 days by default, so this is how the value gets out before they go.

Kickoff and quality

new-mission: For anything with three or more steps, or anything hard to undo. It looks around first, asks at most five questions (with numbered options, so on a phone you mostly reply with a digit), drafts a plan where every step says how it will be checked, and waits for an explicit go. "Looks fine" doesn't count. When the work is done you get a wrap-up report against the plan, including an honest ledger: what wasn't done, what wasn't verified, and what was a guess.
damage-report: Five questions the AI runs against your original request before it reports back. If there's nothing real to suggest, it's allowed to write "none."
ai-review: Sends the work to a different model family for a second opinion, by default through Codex CLI on your own ChatGPT account, with separate modes for code, public-facing copy, and research reports. It then folds the feedback into its conclusion: what it accepts, what it rejects and why, and what you need to decide. If Codex isn't installed or signed in, it says "self-review only" and lets you keep going. A tool that blocks you when you're in a hurry is a tool you stop using.
ai-search: Live web answers with citations you can check. When it can't find something, it says so instead of filling the gap from stale training data.

Budget and life

token-optimizer: Budget rules the AI follows before it spins up multiple agents: pick the right model tier, compress what comes back, and stop after three failures. On a subscription plan like mine, running out of quota stops every conversation until it resets, which hurts more than a bigger bill.
flight-to-calendar: This one is purely for me. Through a Google Calendar connector, it puts booked flights on your calendar with the time zones right, one event per leg, and a note on which side of the plane gets the sunset. If you fly often you know why: travel days are when meetings get reshuffled, and I have sprinted through more than one airport because of it.

Two design decisions

Verification over declaration. The most important rule in the whole set: writes get read back, "done" needs evidence, and a success message on its own proves nothing.

I learned that one the hard way. One evening I ran my wrap-up routine, the tool reported success all the way through, and I closed the session. The next day the file was empty, and a whole block of conclusions was gone. Since then, reading back is a hard rule. Nothing counts as saved until I can see the content on disk.

Zero dependencies by default. Handoff cards are plain Markdown files in your project folder. No database, no API, no service running in the background. Tools should still work on your worst day: new laptop, no network, some service down. A Markdown file still opens. You can plug in your own task system later, but the default only needs a project folder you can write to.

I ran the reviewer on itself

The first job I gave ai-review was reviewing ai-review. Three rounds of cross-model review caught 21 defects, and 13 of them had been introduced by my own fixes from the round before. I fixed what round one found, and round two found new problems in the exact places I had touched.

That's why the tool exists. A fix can plant a new bug, and whoever wrote the fix is the last to see it. Ask the same model to "check again" and you mostly get its original conclusion in different words. So, as of v0.7.3, ai-review ships with 41 regression tests you can run yourself, and ai-search ships with 43.

A second opinion is still just an opinion. Some of its feedback catches real problems, some is wrong because it never saw the code, and some covers things I had already weighed and rejected for good reasons. Treating the second opinion as gospel is the same mistake as treating the first one that way.

Why "MyR2D2"

R2-D2 is never the hero, but every episode depends on him. Leia hands him her message, and he carries it across a desert to Obi-Wan and plays it back word for word.

The mapping to the tools turned out tighter than I expected:

dropoff is Leia recording "Help me, Obi-Wan Kenobi."
pickup is R2 finding Obi-Wan and playing the message.
save-all is keeping the plans safe in the escape pod, so whoever comes next can pick up where you left off.

One more detail: the Chinese version of this post reached its writing session through a handoff card. I wrote the brief in the session where I was tidying the repo, then picked it up in a fresh one. For me, the real test of a tool is the day I can't work without it.

The skills are written in Chinese. You can use them in English.

Each skill's instructions are written in Traditional Chinese, and there's exactly one copy of each, so there are no parallel translations to drift apart. Every skill comes with trigger phrases in both languages: "about to reboot," "hand this off to X," and "anything handed off to me?" all work. Claude follows the Chinese instructions and answers in whatever language you use.

Install

MIT license. The recommended route is one line:

npx skills add tingyulu/MyR2D2

That's Vercel's skills installer, which also supports gemini-cli, codex, cursor and other agents. It installs to your project by default; add -g for a global install, or --skill to pick individual skills.

On Claude Code you can also install it as a plugin, which keeps updates in one place and puts the skills under a myr2d2: namespace so they won't collide with skills you already have:

/plugin marketplace add tingyulu/MyR2D2
/plugin install myr2d2@myr2d2

No CLI? The repo's prompts/ folder has paste-in English versions of new-mission, damage-report, ai-review and ai-search for ChatGPT or Claude on the web. The three journal skills need Claude Code, because they read its local transcripts. The README has the full compatibility table.

Repo: https://github.com/tingyulu/MyR2D2

Solve your own pain points, and don't keep the answers to yourself.

If you try it, tell me how it went: which skill you never touched, and where you got stuck. That kind of feedback is worth more than a star.

I'm Eric Lu ("Uncle Eric"), a product consultant, headhunter, and career coach. These skills are my actual daily workflow.

大叔的筆

MyR2D2's 13th skill: systems-check finds problems without fixing them

What systems-check scans

Its limits

How to install and upgrade

MyR2D2 第 13 支 skill：systems-check，只找問題不動手修

systems-check 掃哪五件事

它的邊界

怎麼裝、怎麼升級

AI 由輔助轉向主動：Grok 主動 Bot、AI Captain 和夜晚哨兵

我的 AI Captain：Chief of AI

設成主動 Bot 之後

官方怎麼說

晚上的哨兵

由輔助轉向主動

MyR2D2 Update: I Switched to Opus 5.5 and Rewrote My Subagent Dispatch Rules

MyR2D2 更新：換上 Opus 5.5，順手把派工規則改了一輪

The Complete Resume / CV Guide: Before You Start, While You Write, Before You Send

Step 1: You haven't started writing

Step 2: You're writing it

Step 3: Before you send it

When you don't have the energy for a big rewrite

If you only have five minutes

Want more

How to Start Writing a Resume / CV: Answer These 3 Questions First

Get clear first: a resume tells people three things

If it doesn't fit one of these three themes, cut it

Break it down and do it in stages

Related reading

3 Resume / CV Mistakes That Get You Screened Out in the First Round

1. Listing only what you did

2. Not answering the question

3. Being too honest

A resume isn't an autobiography. It's an exam.

3 Common Resume / CV Problems I See When Reviewing Friends' Drafts, and How to Fix Them

Update your resume regularly

Focus on achievements

The STAR method

Final thoughts

Job Hunting After a Career Gap? Start With These 3 Resume / CV Improvements

Related reading

Lanjia Gua Bao, Gongguan: Five Ways to Choose Your Pork, and a Bowl of Si Shen Soup

Five ways to choose your pork, and it was like that even back then

It still tastes the way I remember

The details

換 Opus 5.5 前先做三件事：prompt-audit、migrate、免費重置額度

價格

預設 effort 是 medium

從 Opus 5 遷移會壞的地方

模型怎麼分工

訂閱方案

Wang's Broth, Huaxi Street: A Bib Gourmand Braised Pork Rice That Fell Short

How it tasted

The details

Letian Mutton Soup by Nanmen Market: Garlic Mutton Ribs, and NT$10 for a Bowl of Rice

The braised pork rice: a fatty cut that doesn't sit heavy

The braised egg: this one passes

The garlic mutton rib soup: this is the one

The room

The details

Uncle Eric's Table: Michelin Bib Gourmand Places I've Tried

Taipei

Uncle Eric's Table: 500 Bowls Award Places I've Tried

Taipei

大叔の餐桌｜米其林必比登：大叔吃過、寫過的店

台北

大叔の餐桌｜500 碗：大叔吃過、寫過的店

台北

Hsiung Chi Scallion Pancake, Gongguan: Get the Signature with Corn, and Don't Skip the Basil

The line

What to order

The details

公館・長江一號雄記蔥抓餅｜跟著洋流排隊，招牌一定要加玉米

洋流大爆發，大叔也跟風來了

排隊：等到天長地久

店家資訊

No Bot, Just Physics: The Mac Meeting Recorder I Built and Open-Sourced

Two approaches, two costs I didn't want to pay

If a bot is out, how about physics?

What it does today