我如何使用 LLM
從結構到創作的工作方法論How I Use LLMs
A Methodology from Structure to Creation

王新仁 ｜ Aluan Wang

圖片提供：明日製作所
攝影：林家賢

Ileivoivm——「aluan wang」經 Affine Cipher 加密後的產物。 Ileivoivm — the result of running "aluan wang" through an Affine Cipher.

INPUT   aluan wang
CIPHER  Affine (slope=5, intercept=8)
OUTPUT  ileiv oivm → Ileivoivm

生成式藝術家 創意程式設計師 akaSwap 共同創辦人 北藝大新媒體藝術碩士 台灣首位 Art Blocks 藝術家 台灣首位 verse.works 長篇藝術家

Generative Artist Creative Coder Co-founder, akaSwap MFA, New Media Art, TNUA First Taiwanese artist on Art Blocks First Taiwanese artist on verse.works

1982 年生於台中。超過 15 年數位與生成藝術經驗。作品展出於國立台灣美術館、尊彩藝術中心、新加坡藝術科學博物館等。兩度獲台北數位藝術節首獎（2012、2015）。

Born 1982, Taichung. 15+ years in digital and generative art. Exhibited at National Taiwan Museum of Fine Arts, Liang Gallery, ArtScience Museum Singapore. Two-time Grand Prize winner at Taipei Digital Art Festival (2012, 2015).

更多背景 ↗More about me ↗

從取樣開始，從濾波結束Start with Sampling, End with Filtering

早期用 Pure Data 創作。Pd 是圖像化的程式語言，其中有個功能叫陣列（array）。

I started making work with Pure Data. Pd is a visual programming language with a core feature: the array.

一排格子，每格存一個數值。關鍵在於怎麼讀：

A row of cells, each holding a number. The key is how you read it:

後來我做的很多作品，本質上都在做同一件事：替同一份資料發明新的讀法。

Most of my work since then has been doing the same thing: inventing new ways to read the same data.

作品背後都是結構Behind Every Piece Lies Structure

同一份 hash，可以被讀成視覺，也可以被讀成樂譜。格子是陣列，索引是時鐘——序列機讀音高，鼓機讀觸發。結構沒變，讀法決定它是旋律還是節奏。

The same hash can be read as visuals — or as a musical score. Cells are arrays, the index is a clock — the sequencer reads pitch, the drum machine reads triggers. Same structure, different reading.

Paths 系列：資料決定歷史的樣子The Paths Series: Data Shapes History

從田野錄音到衛星影像，從波函數塌縮到摩斯電碼。四件作品，同一條線索：
你選擇收集什麼資料、怎麼 mapping，就決定了歷史長什麼樣子。

From field recordings to satellite imagery, from wave function collapse to Morse code. Four works, one thread:
what data you choose to collect, how you map it — that determines what history looks like.

Chaos 三部曲：作品吃作品The Chaos Trilogy: Art Feeding on Art

GeoPunk 取樣地球。Chaos 取樣自己。
作品以前作為原料，層層遞迴，生成新的混沌。

GeoPunk sampled the Earth. Chaos sampled itself.
Each piece consumes its predecessors, recursing layer by layer into new chaos.

一個人不會真的消失A Person Never Truly Disappears

幾年前，一個朋友走了。

A few years ago, a friend passed away.

有天我在 Google 上查一個技術問題，搜尋結果指向一個論壇。
點進去，看見他的回答。

One day I googled a technical question. The results led to a forum.
I clicked in — and saw his answer.

他留下了一個錨點。知識被索引，他就沒有消失。

He had left an anchor. His knowledge was indexed. He hadn't disappeared.

JSON 作為記憶層JSON as the Memory Layer

LLM 的上下文有限，記憶會斷。
JSON 和 .md 就是錨點。

LLM context is finite. Memory breaks.
JSON and .md are the anchors.

《修仙-七玄關》：故事的外部記憶Seven Gates: External Memory for a Novel

一個橫跨宗教、犯罪與超自然的長篇小說。角色在第三章說的話不能跟第一章矛盾，伏筆埋了就要收。

A long-form novel spanning religion, crime, and the supernatural. What a character says in chapter three can't contradict chapter one. Every foreshadowing must pay off.

{
  "chapter_1": {
    "segments": {
      "segment_1": {
        "title": "序曲－雨中的電話",
        "summary": "林承祐在深夜接到命案報告...",
        "content": "（完整文本）",
        "analysis": {
          "setting":        "嘉義市區",
          "characters":     ["林承祐", "小陳"],
          "emotional_tone": "孤獨、麻木、恐懼",
          "foreshadowing":  "宗教案件的陰影重現"
        }
      }
    }
  }
}

每個段落附帶 analysis。寫第五章時回溯前面所有 analysis，情緒曲線連貫，伏筆正確延續。

Every segment carries an analysis. When writing chapter five, I trace back through all prior analyses — emotional arcs stay coherent, foreshadowing threads continue correctly.

PolyPaths：觀眾不知道自己正在被記錄PolyPaths: The Audience Doesn't Know They're Being Recorded

觀眾在介面上畫路徑，以為只是在互動。他們不知道每一筆都被記下來了。

Visitors draw paths on the interface, thinking they're just interacting. They don't realize every stroke is being recorded.

路徑壓縮成 hash，塞進 URL。任何人拿到這串 URL，就能還原完整的繪製過程——然後長出一棵植物。

Paths get compressed into a hash, embedded in the URL. Anyone with that URL can replay the entire drawing process — and grow a plant from it.

↗ 掃描互動

InkField：AI 用 JSON 畫水墨InkField: AI Paints Ink Wash via JSON

AI 可以畫一幅水墨。不是操作滑鼠，不是生成像素——是輸出一份 JSON，引擎還原成完整畫作。

AI can paint ink wash. Not by moving a mouse, not by generating pixels — by outputting a JSON file that the engine renders into a complete painting.

↗ 掃描互動

錄製系統只記錄事件（筆觸參數 + 種子值），不記錄像素。雙種子確定性系統，99.9% 一致性。一幅作品 < 100KB。

The recording system captures events (stroke parameters + seed values), not pixels. Dual-seed deterministic system, 99.9% consistency. One painting < 100KB.

Claude Code 開發，CLAUDE.md 管理規範，doc/ 保留每次迭代的技術文檔。

Built with Claude Code. CLAUDE.md governs project specs. doc/ preserves technical docs from every iteration.

InkField Gallery：過程可以被繼承InkField Gallery: Process Can Be Inherited

每幅作品的繪製過程都是一份 JSON。任何人可以下載、研究、重播，然後用同樣的引擎畫出自己的版本。

Every painting's process is a JSON file. Anyone can download it, study it, replay it, then use the same engine to paint their own version.

不是複製成品，是繼承過程。你拿到的不是一張圖，是一套筆法。

It's not copying a finished piece — it's inheriting a process. What you get isn't an image, it's a set of brushstrokes.

我的 LLM 工具鏈My LLM Toolchain

最後一層：用 .md 建立數位分身The Final Layer: Building a Digital Twin with .md

我把 Facebook、X、Threads 上的所有發言下載下來，整理成一份「寫作風格指南」.md：

I downloaded all my posts from Facebook, X, and Threads, and distilled them into a "writing style guide" .md:

留下錨點，讓一切可以被 indexLeave Anchors. Make Everything Indexable.

延伸閱讀Further Reading