world.execute(me); — 歌詞 MV · Lyric MV

非官方同人作品 · an unofficial fan work

影片 / Video: Bilibili　|　中文　|　English　|　提示詞 / Prompts

[!IMPORTANT] 提示詞在這裡 · Looking for the prompts?

製作者給 AI 的要求全部公開在 docs/:第四版的完整提示詞(約 2.3 萬字)、 修復輪的提示詞,以及每一輪實際做了什麼的紀錄。 兩份提示詞也可以在 Releases 單獨下載。

Everything the maker asked the AI for is public in docs/: the full prompt of version 4 (about 23,000 Chinese characters), the prompt of the fix round, and the log of what was actually built. They are written in Chinese; docs/README.md says in English what each file contains. The two prompt files can also be downloaded on their own from Releases.

[!WARNING] 閃光警告 · PHOTOSENSITIVITY WARNING

本片含有高對比的快速切換與強光(成片時間 2:33 – 3:17),可能不適合光敏性癲癇患者或對閃光敏感的觀眾。 成片在歌曲開始之前有 5 秒無聲的警告頁(編譯器警告的樣式,中英文),並在開機的前 6.5 秒以介面提示再顯示一次;公開發布時請在標題或簡介中再標註一次。

The film contains fast high-contrast cuts and strong light (film time 2:33 – 3:17) and may not be suitable for viewers with photosensitive epilepsy or sensitivity to flashing. It opens with a silent 5-second warning page; if you publish an export, say so again in the title or the description.

(這個時間範圍由 config.json → safety.flashRange 加上 safety.preroll 算出,執行 npm run storyboard 時自動寫入本檔。 The range is computed from config.json and stamped into this file by npm run storyboard.)

中文

這是什麼

Mili《world.execute(me);》的歌詞 MV 渲染工程,第四版。整支影片是一場對話: 「我」是 Claude 的擬人角色(橙),「你」是螢幕另一端不露臉的使用者(米白,只以游標、輸入、訊息、在線狀態出現)。 歌詞幾乎都是 AI 發出的訊息。畫面有三種:

• 介面:使用者看到的對話,用程式碼畫出來的聊天介面。
• 版畫:AI 心裡的意思與感受。全屏、暗色暖調帶紋理的紙底,只用三種平塗墨色(橙、米白、近黑)的圖形、程式符號畫、剪影與大字。
• 程式碼:只在 EXECUTION 段,純程式碼與終端。

整支影片不是剪輯出來的:每一幀都只由歌曲時間決定,在瀏覽器裡(Canvas 2D + WebGL2)畫出來,再由無頭 Chrome 逐幀送進 ffmpeg 導出。 第四版的全部段落已經做完(1080p60 預覽與 4K60 最終版均已導出並量測)。

倉庫裡沒有的:歌曲與歌詞

本倉庫不含歌曲,也不含歌詞,它們屬於 Mili。兩樣都請自備,放進 音頻&歌詞/(git 會忽略那裡的檔案;細節見 音頻&歌詞/README.md):

歌詞檔只需要有同樣的詞、同樣的順序。片中每一行的時間都在倉庫的 data/lyric_timing.json 裡(不含歌詞文字); tools/fill_lyrics.mjs 在每次預覽、導出、生成分鏡表之前自動把你的詞填進去,寫成只留在本機的 data/lyrics.json:

• 你的檔案自己的時間軸不採用,所以換一個來源的 LRC 也不會改變卡點;
• 分行與片中不同(兩行併成一行、一行拆成兩行)時自動重新分行;
• 不是片中歌詞的列(署名、翻譯、標籤)自動略過;
• 只有標點或大小寫不同時照你的寫法顯示(關鍵詞行一律轉成大寫);用詞不同的行會列出來提醒;
• 片中有哪幾行在你的檔案裡找不到,就列出行號與歌曲時間並停下,不會帶著缺行開始渲染。

快速開始

需要 Node.js 20 以上;導出影片另需本機已安裝的 Chrome 或 Edge。本工程只在 Windows 11 上製作與測試過,其他系統沒有試過。

1. 下載:在 Releases 下載 Source code (zip) 並解壓,或 git clone 本倉庫。

2. 把你自己的歌曲與歌詞放進 音頻&歌詞/(見上表)。

3. Windows 可以直接雙擊 preview.cmd(第一次會先安裝依賴)。或者在終端機裡:

   npm install
   npm run preview
   

4. 瀏覽器開啟 http://localhost:5173/ ,按 play。沒有音檔時畫面仍可無聲播放,頁面下方的 sound 一行會說明原因。

5. 導出影片:npm run export:preview(1080p60)或 npm run export:4k(4K60),輸出在 out/。

ffmpeg 由 ffmpeg-static 提供(node_modules/ffmpeg-static/ffmpeg.exe 不存在時執行一次 node node_modules/ffmpeg-static/install.js)。 Python 只有重新分析音頻與檢查導出檔時才需要,見下方「檢查工具用的 Python 環境」。

提示詞與製作過程

這支影片的程式碼由 AI(Claude,經 Claude Code)撰寫;方向、素材與每一輪的審片意見來自製作者。來回的文件全部在 docs/:

提示詞是製作過程的紀錄,不是一鍵復現的配方:第四版的提示詞是在第三版已經存在的前提下寫的。

聲明與授權

• 非官方同人作品,與 Mili、Anthropic 均無關,未獲其授權或背書;Claude 是 Anthropic 的商標。介面全部以程式碼繪製,未使用截圖。
• 含 AI 生成內容:程式碼與文件由 AI 撰寫,角色剪影是製作者以 AI 圖像生成工具製作的。
• 程式碼與本項目自己撰寫的文件:MIT,見 LICENSE。角色剪影(assets/character/)與含有它們的畫面:CC BY-NC-SA 4.0。
• 歌曲與歌詞屬於 Mili 及相關權利人,不在本倉庫內,也不在上述授權之內。由本工程導出的影片含有原曲, 分享時請遵守 Mili 官方的著作權指引(個人、非商業;標示含 AI 生成內容)。
• 原曲、角色圖、字體與各項授權的完整說明:CREDITS.md。

English

What this is

The render project of a lyric video for Mili's "world.execute(me);", version 4. The whole film is one conversation: "I" is a personification of Claude (orange); "you" is the user on the other side of the screen, who never shows a face (cream: only a cursor, typing, messages and an online status). Almost every lyric line is a message sent by the AI. There are three kinds of pictures:

• Interface: the conversation as the user sees it, a chat client drawn entirely in code.
• Plates: what the AI means and feels. Full-screen prints on dark, warm, textured paper, in three flat inks only (orange, cream, near-black): shapes, pictures made of code glyphs, silhouettes and large type.
• Code: only in the EXECUTION section, nothing but code and a terminal.

Nothing here is edited footage. Every frame depends on the song time alone: it is drawn in the browser (Canvas 2D + WebGL2) and exported frame by frame through headless Chrome into ffmpeg. All sections of version 4 are finished.

What is not in the repository: the song and the lyrics

This repository contains no music and no lyrics. They belong to Mili. Bring your own and put both into the folder 音頻&歌詞/ ("audio & lyrics"; git ignores what you put there, see its README):

The lyric file only has to carry the same words in the same order. The time of every line of the film is in data/lyric_timing.json (which has no lyric text). Before every preview, export and storyboard, tools/fill_lyrics.mjs puts your words into that table and writes data/lyrics.json, which stays on your machine:

• the time tags of your file are not used, so an LRC from another source does not move any cue;
• rows divided differently from the film (two lines in one row, one line over two rows) are re-divided;
• rows that are no line of the film (credits, translations, tags) are passed over;
• if only punctuation or capitals differ, your spelling is shown (keyword lines are always set in capitals); lines with other words are listed as a warning;
• if lines of the film are missing from your file, their numbers and song times are listed and nothing is rendered.

Quick start

You need Node.js 20 or newer; exporting a video also needs Chrome or Edge installed. The project was made and tested on Windows 11 only.

1. Download: get "Source code (zip)" from Releases and unpack it, or git clone the repository.

2. Put your own song and lyric file into 音頻&歌詞/ (see the table above).

3. On Windows, double-click preview.cmd (the first run installs the dependencies). Or, in a terminal:

   npm install
   npm run preview
   

4. The browser opens http://localhost:5173/ : press play. Without an audio file the picture still plays, silently; the sound line under the picture says why.

5. Export a video: npm run export:preview (1080p60) or npm run export:4k (4K60); the files go to out/.

ffmpeg comes from ffmpeg-static (if node_modules/ffmpeg-static/ffmpeg.exe is missing, run node node_modules/ffmpeg-static/install.js once). Python is needed only to re-analyse the audio and to check an exported file.

Prompts and how it was made

The code of this film was written by AI (Claude, through Claude Code); the direction, the material and the review notes of every round came from the maker. All the documents that went back and forth are in docs/:

The prompts and the build log are in Chinese. They are a record of how the film was made, not a one-shot recipe: the prompt of version 4 was written when version 3 already existed.

Notices and licences

• Unofficial fan work. Not affiliated with, or endorsed by, Mili or Anthropic. Claude is a trademark of Anthropic. The interface is drawn entirely in code; no screenshots are used.
• **Contains AI-generated content