去玩遊戲

「我在司改會的時間，是從讀不完的資料開始的，最後卻變成親手做出那些曾經只寫在紙上的世界。」

一個完全陌生的名字

實習的第一天，很興奮的我拿到的是一個看起來再平常不過的任務：「上網查一下蘇建和案」。

那是四個月前的事，當時我對台灣的司法制度幾乎一無所知，這個名字對我來說完全陌生，我甚至不知道該從哪裡查起。我在網路上不停地滑，關鍵字從「蘇案」換到「蘇建和」，最後留在腦中的，只有案子裡的五個名字；除了名字之外，我什麼也不明白。

接著，我被交付了新的任務：讀這個案子的判決。整整三個星期，我的時間全部花在那一百三十頁判決上，一頁一頁地讀、分析、做註記。讀完最後一頁時，我確實有一種踏實的收穫感，但那個三十五年前的案子，還是那麼的遠，它是文字，是描述，也是別人的事。

AI做遊戲，是白日夢的幻想嗎？

暑期實習生陸續加入司改會，我們的計畫正式啟動。我不再是那個埋頭讀判決的實習生，而是要帶著他們認識這個案子、做出一個遊戲。

我們從30年前的手寫筆錄開始，有人負責翻閱數百份剪報，有人負責讀十幾年前寫成的鑑定報告，也有人負責比對筆錄裡那些互相對不上的自白、整理所有曾經參與過的人的照片。就這樣，我們把原始材料一件一件湊了出來。

當其他暑期實習生們埋在三十多年的資料裡時，我在做的事情是跟 Claude 聊天。

我每天花五個小時以上的時間跟它對話，不停地丟問題給它，想搞清楚它的極限與可能性：它真的做得出我想像中那個完整的互動式網頁遊戲嗎？還是這終究只是一場 AI 的白日夢？

當時我們手上除了一堆還沒整理好的素材，什麼都沒有。我們給 Claude 的第一道指令是：做一個 RPG 風格的房間。它交回來一個空蕩蕩的房間、一扇簡單的門、一片沒有顏色的磁磚地板，還有一個可以從畫面這頭走到那頭的小人。第二道指令，把我們零散的材料全部丟給它，包括未經整理的圖片和一疊筆記，測試它能不能從一團亂裡整理出頭緒。

Claude 最初給的答案，我們一次次推翻、一次次要求重來，直到它終於替這個遊戲打下地基，做出了最初的六個房間。在一道又一道的指令之間，我慢慢學會了 AI 真正聽得懂的說法：把尺寸、位置、限制與目標全部講清楚，不留任何模糊空間。

難的不是遊戲設計，是如何讓遊戲機制自己說故事

接下來是佈置每一個房間。椅子要排成整齊的一列、桌子的大小要和房間的比例相符、警察角色要換上制服並配上警棍，每個房間都被反覆調整，直到它符合實際的樣子。

法庭的配置必須和檢察官實際的座位一致，法袍的顏色要一一對應，看守所的燈光要調得更暗，那份壓迫感才成立。從顏色到家具，每一個細節，都是我們不斷向 Claude 下指令的成果。事實上，我們用得太兇，額度常常用光，只能乾等它重置。（這是整個過程裡我們最不喜歡的部分⋯⋯）

除了畫面之外，Claude 也照著我們的每一道指令，把案件資料改寫成對白。遊戲裡的每一句話都出自我們蒐集到的材料：先取自法律文件，再把那些密度極高的文字改寫成能推著敘事走的對話。一個字一個字推敲的時候，Claude 只負責抓我們的錯字；哪些素材要留、哪些要拿掉，如何取得義務律師團的同意，終究還是得靠我們自己。

遊戲的機制是整個過程中最吃力的另一個部份，因為機制沒有辦法直接從文件裡搬過來，只能靠我們自己設計。

我們一幕一幕慢慢琢磨，不斷調整玩家「可以做什麼」和「不可以做什麼」。每一個決定都是刻意的，從「不看完桌上那張紙就打不開房門」到「限制玩家能走動的範圍」，都是為了讓玩家自己走到真實的故事前面。

我們希望每一個房間的每一個遊戲機制，自己說出曾經發生的一切。

建和，一個再也普通不過的朋友

四個月過去，我的實習即將結束。寫下這段文字的時候，建和就坐在我對面。

這個遊戲，是為了建和而做的。在這之前，我對建和的全部認識，都來自網路、來自判決。

我聽過他沒完沒了的碎念：他最喜歡的菜、每天好奇地問我午餐吃了什麼、堅持要我去試試的那攤鹹酥雞。不知不覺地從我們每一次的日常閒聊裡，橫在我和這個案子之間的距離，變成了一段溫暖的友誼。

建和，成為了我的朋友。

這是一種很奇怪的感覺。看不完的判決、找不完的網路資料，沒有任何一個字提到他退休以後想開一間小吃攤、他在汐止最喜歡的鐵板燒，也從來沒有提到他煮給我們好喝得不像話的酸梅湯和紅豆冰，更沒有提到建和是一個體貼、善良的同事、前輩，是一個再普通不過的，和你我沒有任何不同的人。

希望這個遊戲能像一台時光機，把我自己花了四個月和一段友誼才走到的地方，讓大家也走一次，讓大家認識蘇建和的故事。

---

I Built a Time Machine with AI

“My time at JRF began as what seemed like endless reading to being able to create what was once written on the papers I read.”

A Name That Meant Nothing to Me

From the beginning of my internship at JRF, I was tasked with a seemingly mundane and routine task- “search on the web about the Su-Chien-He case”. At that time, shortly just four months ago, my unfamiliarity with Taiwan’s legal system served me no knowledge in what the case was, the name was foreign to me, nor would I know where to begin my search. Scrolling on the web endlessly, trying key words of Su’s case related to Su-Chien-He, I ultimately left with only a shallow memory of the five names mentioned in the case with no understanding beyond their names.

Moving forward, I was handed a new task of reading over the courts' judgement of the case. Through three weeks, I spent all my time reading, analyzing and annotating the one hundred thirty pages of court judgement I was given. As I finally reached the end of the paper, although rewarding, words and descriptions of a 35 year old case still remained aloof and at a distance from me.

A Game Made by AI: Just a Daydream?

As a new group of interns came, our project finally began. I was no longer the intern reading over the courts' judgement, I was now asked to guide them into understanding and familiarizing them with the case. Our journey began with strenuous and vigorous efforts to transcribe the handwritten interrogation records into digital typed versions. Dividing the thirty five year case into different parts to be studied proved to be arduous: some were responsible to gather and browse through hundreds of newspapers, some were tasked to read and analyze forensic reports written from decades ago, others were asked to examine the interrogation records to find the misaligning differences confessions. Slowly, we began to gather the raw materials one by one.

I created a time machine using AI

While interns were busy digging through thirty years worth of materials, I was busy chatting with Claude– the most forefront and powerful AI. Spending five hours a day chatting with Claude, I persistently tossed relentless questions at Claude, trying to understand its limitations and potential– could it really create a fully interactive online game I hoped for? Or was it just a mere AI dream? With nothing other than unassembled materials, we inscribed Claude with its first order– make me an RPG themed bedroom. A plain room, a simple door, a colorless floor made of tiles, a small figure who could walk from one side of the screen to the other, Claude’s first attempt to build the game appeared. The second order followed by feeding Claude our scattered materials including unorganized images and pages of notes, we tested AI’s ability to comb through our disconnected materials. Unsatisfied with its initial responses, we scraped and demanded a new answer to every question until finally, Claude successfully laid the groundwork and built the very footing of the game containing six rooms. With each prompt given, I learned slowly to mimic the exact prompts AI responds to by specifying the dimensions, positions, constraints and goal of my instructions.

The Hard Part Wasn't Designing the Game, It Was Making It Tell the Story We Wanted to Tell

Embellishing each room– from adding the chairs in a horizontal row, making the size of the desk proportionate to the rooms’ layouts, giving police uniforms and batons to the police characters, – each room was meticulously refurbished until it matched our imagination. The court room’s layout needed to be identical to the prosecutors’ placements, the court gowns were carefully designed with its corresponding colors, the detention centre’s lighting was toned darker to match its oppressive mood; every detail (from the color to the furniture) was the result of our ceaseless instructions for Claude. In fact, our usage of Claude was so intense that we frequently ran out of usage, forcing us to wait until the token resets (this was our least favorite part of the process…).

Beyond visual enhancements of the room, Claude also followed our every instruction to rewrite the case file into dialogues. Every line in the game was built using the materials we collected, using texts from legal documents then rewriting such dense texts into a narrative dialogue for the game. Working through each word, Claude stood by to catch our spelling errors while we revised every line, evaluating what to include and omit.

The mechanics of the game were our most arduous efforts as the mechanisms could not be lifted out of a document, but rather relied on our creative designs. Shaping scene by scene carefully, we continuously adjusted what the players were allowed and not allowed to do. Each choice made was deliberate– from locking the room until the paper left on the table has been read to limiting the walking space of the players– to reflect a conclusion we wanted the players to understand from the case. By intentionally constraining and enabling certain mechanisms such as controlling the walking speed or making certain reads mandatory, we designed each room with specific conclusions to relay to the players.

Chien-ho, an Utterly Ordinary Friend

As I write this, sitting across from me is Chien-He, the person the game was built for. Just months ago, his name had only appeared on paper in which the verdict was the only information I knew about him. Through the past two months, I listened to his endless chatter about his favorite dishes and his insistent recommendations to try his favorite Taiwanese fried chicken stand. Somewhere in our daily conversations, the distance that once stood between me and the case had turned into a friendly, warm friendship with whom I considered my friend. It was an odd feeling to see how a document can seemingly appear so complete, yet provided so little of the person it was written about– nowhere did the 300 pages of texts wrote anything about his retirement dream to open a stand, his favorite teppanyaki restaurant in Xizhi, or his ridiculously tasty Sour plum drink and red bean shaved ice he cooked for us (it was the most authentic Taiwanese desserts I have tried before btw). Underlying the name I had read countlessly was an empathetic, caring and compassionate co-worker of mine, one everyone would deem as normal, just like you and me. Retelling his story, I hope to give others, through a game, some part of what took me hundreds of pages, two months and a friendship to reach.