A Slow Week
This week, progress noticeably slowed down. In A Car Ride, Design Markups, and a Strong AI, I wrote about why this blog took just seven days to launch, and how I handed my shelved Instagram project over to Claude to turn into a neighborhood-by-neighborhood walking guide for @love.in.hongkong. But this website is a completely different beast. The volume of content is massive, and while Claude can generate the code and build the layout, I quickly realized how much of the heavy lifting still falls entirely on me.
The site still needs time before it is ready to launch, but here is a quick peek at how the @love.in.hongkong walking guide is taking shape. I am sorting the reels by neighborhood across Hong Kong, so visitors can click into a specific area and find curated walking routes and local highlights:
My Boss and Legal Counsel
Back in Without a Cheat Sheet, I mentioned that Claude is a serious AI. That seriousness showed up in a realm I rarely think about: copyright. It turned out to be remarkably strict, catching details I had never paid attention to. The clearest example is my walking map.
For my walking maps, I had always simply drawn routes directly onto Google Earth aerial imagery. That is what people routinely do for personal passion projects, where copyright rarely crosses your mind (or where you honestly don't realize there is an issue to begin with). But now that @love.in.hongkong is becoming an independent standalone website, it technically counts as a commercial site. After carefully reviewing all the assets, Claude flatly "forbade" me from using Google maps. It checked Google’s terms of service and pointed out that commercial websites cannot simply display Google maps as static graphic images. Claude caught a legal blind spot I had sailed right past.
The exact same thing happened when we built this blog. At the time, I wanted to use an illustration I found on Pinterest as the hero banner at the top of the homepage. Claude never outright refused. Instead, it kept that hero banner on the to-do list before launch, methodically analyzed the legal risk if the original artist discovered it and demanded payment, and even drafted a polite permission email for me to send. I got the message loud and clear: if I didn't sit down and compose my own image, this blog was never going live. When I told my colleague Sylvia about this, she laughed and said, "Claude is your boss now." Looking back at our debates over map licensing, Claude isn't just my boss; it has become my in-house legal counsel.
When Aesthetics Take Over
Claude framed the issue around legal terms, but when I stepped back and looked at the actual interface, my own eyes arrived at the exact same conclusion for a completely different reason.
The color palette of Google's map is simply too saturated and loud. Sitting directly beside an Instagram reel, the map steals the reel's thunder. In my design hierarchy, the map is supposed to be a quiet supporting graphic; the hero must always be the video. The moment the map demands attention, the video loses its focus.
In the end, what truly doomed the map was not legal anxiety or Claude's cautionary warnings. It was my own aesthetic standard as a designer.
Placing the map directly next to the video player made that visual conflict obvious right away:
The Unsettled Map
Once we agreed the map had to go, Claude did an admirable job playing the role of a seasoned senior designer, hunting across the web for free, open-source alternatives. But my designer's eye refused to settle for whatever happened to be free.
The government's free basemaps look terribly dull and dated. Government aerial photos looked foggy and dirty next to crisp, high-resolution video footage. On the other hand, the prettier, more aesthetic maps I tested were stripped too clean: they were not clear enough to provide the right information.
Claude eventually came up with a middle ground. Taking my notes, it went in and modified the government's free map itself, adjusting the color balance. While it was a noticeable improvement, in my gut I still wasn't one hundred percent satisfied with it. Rather than letting the whole project stall, I decided to shelve the map for now. It remains unsettled.
Moving Postcards and the Blue House
With the map on pause, I turned my attention to the video content. On my Instagram feed, I have a collection of "moving postcards"—a short-form reel with beautiful footage and music. The footage is beautiful, but when I originally posted them, they had no voiceover scripts and very little text. My plan was to put them on the website with travel guide notes in the captions first, buying time before I can produce full travel-guide reels later on.
The moment we started, however, I discovered that without my voiceover script to guide it, the AI simply cannot write anything that grasps the human experience. Both Claude and the other AI agent failed completely.
The clearest example was the reel of the Blue House in Wan Chai under the Mid-Autumn lanterns. I had filmed a moving postcard there in the past, and more recently recorded a voiced version. With only the silent footage to work from, the AI drafts were painfully robotic. It merely described what was visible on screen: people looking at lanterns, someone taking photos, people sitting under the lanterns on the benches.
When I sat down to write my own voiceover script for that reel, my line was: "the courtyard turns into a cozy neighborhood living room where people actually slow down, sit on the benches, and catch up under the warm glow."
The AI cataloged the objects in the frame; my script captured what it feels like to exist in that space, and why the place draws people in. Putting the two side by side made the gap unmistakable. I let out a heavy sigh. There are no shortcuts here. I cannot be lazy; every single line with warmth and soul will have to be written with my own hands.
So I took a step back and asked the AI to write the factual background for each location. Objective facts do not require personal warmth—surely that was safe territory? But even the facts came back dead on arrival, reading "too wikipedia." When it drafted a description for East Coast Park, it kept droning on about the breakwater, repeating boring background facts instead of explaining what fun facilities the park has and what would actually attract people to visit. When I searched online, I found those exact sentences in press releases (which only shows just how boring those press releases really are).
Tedious Work and Plain Facts
So that is where the project stands right now.
The current phase is about patience: methodically filling in each neighborhood's content, one place at a time. I have to dive into every detail, rewriting and polishing most of the captions myself so they carry a human pulse. Claude assists by assembling the verifiable facts—historical background and opening hours—giving me raw material I can use as a base draft to revise.
Building this Instagram site is not as simple as the blog website. It is much more tedious and will take much more time.
So Claude can be my coder, my senior designer, my boss, even my legal counsel. But the writer? That job is still mine. And it's going to take a lot longer than seven days.
第四週：一張被禁用的地圖、客廳般的中庭、我的法務顧問
當用聊天寫程式慢了下來，才發現真正重要的部分只能自己動筆
慢下來的一週
這個星期，進度明顯慢了下來。在〈第三週：那段車子裡的對話、設計修改、強大的 AI〉裡，我才聊到為什麼這個部落格只花七天就順利上線，以及我把擱置已久的 IG 網站交給 Claude 接手，替 @love.in.hongkong 打造按地區分類的散步指南。然而，這個網站的內容實在太多了。雖然 Claude 能幫忙寫程式搭架構，但我發現，有太多事情得由我自己完成。
雖然網站還需要一些時間完成，但這裡可以先讓大家瞥一眼 @love.in.hongkong 散步網站的雛形。我正把短影音依香港各區分類歸位，未來只要點進特定區域，就能瀏覽那裡相關的散步路線與介紹：
老闆與法務顧問
在〈不用小抄〉裡，我曾說過 Claude 是個嚴肅的 AI。這份嚴肅在版權上體現得淋漓盡致。它非常嚴格，注意到許多我過去從未在意的細節。最明顯的例子，就是我的散步地圖。
以前畫散步地圖，我都是直接在 Google Earth 的空拍圖上標註。畢竟那只是我個人的興趣專案，不用在乎版權這件事（或者說，我根本不知道這也有版權問題）。但現在 @love.in.hongkong 要做成獨立網站，性質上算商業網站。Claude 仔細審查所有素材後，嚴格地「禁止」我繼續使用 Google 地圖。它查了 Google 的服務條款，指出商業網站不能直接展示 Google 地圖。它注意到我忽略的問題。
同樣的事在做這個部落格時也發生過。當時我想用 Pinterest 上找到的插畫當首頁主視覺。Claude 從來沒有直接說不行，但它把這張圖列為網站上線前的待辦清單，還跟我分析萬一被原作者發現、要求付費的風險，甚至連徵求同意的授權信都幫我擬好了。我立刻聽懂了它的弦外之音：如果我不自己動手做一張，它就不會幫我把這個部落格弄上線。我和同事 Sylvia 聊起這件事，她忍不住大笑說：「Claude 現在是妳的老闆了。」回想它對地圖版權的盤查，我覺得它不只是老闆，根本兼任了我的法務顧問。
最終還是審美說了算
雖然 Claude 搬出了版權條款，但當我靜下心來看整個排版時，我也得出了相同的結論，只是理由截然不同。
Google 地圖的色彩實在太鮮豔了。當這張地圖直接放在直式短影音旁邊時，它搶了影片的風采。在我的設計裡，地圖應該是安靜陪襯的輔助圖片，主角永遠是影片。地圖一旦搶眼，影片就失去了焦點。
所以最後決定換掉地圖的關鍵，並不是法律規定或 Claude 的警告，而是我作為設計師的審美標準。
把地圖放在影片旁，那個視覺打架的感覺很明顯：
懸而未決的地圖
既然要換，Claude 也展現了資深設計師的幹練，努力在網路上搜尋免費合法的開放圖資。但我身為設計師的審美，實在沒辦法隨便將就。
政府的免費地圖底色非常沉悶，看起來很舊，在清晰高畫質的影片旁，顯得更加灰濛濛、甚至髒髒的。至於其他比較美觀、有氣質的地圖，圖面卻不夠清楚，無法傳達我想要給的資訊。
Claude 後來幫我做出一個折衷方案。它照我的修改意見，自己動手調整了政府免費地圖的色調。雖然效果改善了不少，但我心裡依然覺得沒有很滿意。為了不讓進度卡死，我決定先把它擱在一邊。這張地圖目前依然懸而未決。
動態明信片與藍屋中庭
地圖放一邊後，我轉頭去處理短影音。我的 IG 上有不少「動態明信片」（moving postcard，一種結合攝影美感與時間流動感的短片形式），畫面很漂亮，但當初發布時沒有旁白配音，也沒有寫詳細介紹。我的想法是，在有時間錄製攻略影片前，先把這些動態明信片放到網站上，並在說明欄附上導覽文字。
然而一動手，我就發現少了我的旁白腳本，AI 根本寫不出具有人情味與真實體驗的句子。無論是 Claude 還是另一個 AI 工具，全都失敗了。
最明顯的例子，是灣仔藍屋中秋燈籠那支短影音。我曾經替那個場景拍過動態明信片，最近也做了一支有旁白配音的短片。而 AI 在只看到動態明信片的狀態下，寫出來的稿子非常機械化。它只是描述畫面上看得見的事物：有人在看燈籠、有人在拍照、有人坐在燈籠下方的長凳上。
但當我為那支影片寫旁白腳本時，我的句子是：「庭院變成了街坊的客廳，大家慢下腳步，坐在長凳上，襯著燈籠透出的溫暖光線，聊著彼此的近況。」
AI 記錄了鏡頭前的物件，而我的腳本寫出了那個空間帶給人的感受，以及這個地方為什麼會吸引人。兩者放在一起，落差一目了然。我忍不住嘆了一口氣。這件事沒有捷徑，我沒辦法偷懶，所有帶有溫度的文字終究得由我來完成。
所以我退一步，讓 AI 幫我寫景點的背景資料。客觀事實不需要人的溫度，這樣總該辦得到了吧？結果我發現，它寫出來的文字一樣無比枯燥，讀起來「太維基百科了」。例如它在描述東岸公園時，通篇都在提防波堤，淨是些無聊的背景，卻完全沒提大家真正關心的特色設施與吸引人的亮點。後來我上網一查，才發現那些句子跟新聞稿裡寫的一模一樣（這也代表那些新聞稿有多麼無聊）。
繁瑣的日常與客觀事實
這就是目前的真實進展。
現階段的工作，就是耐著性子把每個地區的內容逐步填滿。我必須親自深入每個細節，修改並重寫大部分的文案，賦予它們應有的溫度。至於那些可以查證的客觀事實，像是歷史背景與開放時間，先由 Claude 幫忙補齊，讓我有資料可以當底稿來改。
這個 IG 網站製作過程，完全不像當初做部落格那麼簡單。它瑣碎得多，也勢必需要花費漫長得多的時間。
所以，Claude 可以是我的工程師、我的資深設計師、我的老闆，甚至是我的法務顧問。但是，寫作者？這份工作還是我的。而且，這次絕對不只七天。