Roblox 风格方块恐怖短片分镜

<inputs>
先问我小镇的规则（“在这个小镇，如果你的手机在 3:00 响起，千万别接”）、谁违反了规则，以及反转是什么。如果我没有想法，就提出 3 个方案让我选。然后写出 8 个镜头和所有字幕，每次 1 到 3 个词，并标注时间。开始制作前先给我看。
</inputs>

<direction>
一段 1080x1920 的 Roblox 风格 3D 恐怖短片，看起来像是在夜晚的方块游戏中拍摄。
世界：绿色带凸点的底板、带黄色虚线的黑暗道路、灰色人行道、8 栋玩具色彩的房屋，配有山墙屋顶、门廊、白色栅栏和发光窗户；方块松树、真实点光源效果的路灯、钟楼、月亮、星星和蓝色夜雾。
角色：经典方块化身（圆柱形头部、方形躯干，手臂和腿通过肩部与髋部转轴连接，以便坐下和行走），画出的面部表情随节拍变化。实体：全身漆黑，身高 1.75 倍，发光的白眼睛，只有最后一个镜头才露出白色咧嘴笑容。
手机特写：一张 2D 来电界面从 3D 镜头上滑出（先显示“Unknown”，然后显示“Mom 🏠”），铃声脉冲，Decline / Accept 按钮带有点击波纹效果。
字幕：每次显示 1 到 3 个词，置于中央，使用厚重圆润字体，白字配粗黑描边，弹入时带有过冲效果。关键词用黄色，危险词用红色。手机显示时，将字幕移到顶部。
摄影机：每隔 3.5 到 4 秒切换一个镜头，每次切镜时闪白，缓慢推进，并在铃声和惊吓处加入抖动。
禁止：真实游戏标志或界面、素材库素材、静止不动的摄影机。
</direction>

<structure>
0-3.6 秒，航拍小镇：“IN THIS TOWN, / IF YOUR PHONE / RINGS AT 3:00, / DON'T ANSWER.” → 3.6-7.2 秒，钟楼敲响 3:00，钟声，画面抖动 → 7.2-11 秒，孩子站在门廊上，手机发光 → 11-14.6 秒，手机特写：“Unknown”，他拒接 → 14.6-18.4 秒，街道，窗户一扇扇熄灭，手机再次响起：“Mom 🏠” → 18.4-22.4 秒，他家里每扇窗都一片漆黑：“BUT MOM / WAS ASLEEP / UPSTAIRS.” → 22.4-26.4 秒，他接起电话，心跳加速，身后的前门吱呀一声打开：“SHE WHISPERED / "TURN AROUND."” → 26.4-29.4 秒，高个身影就站在他身后门廊上，咧嘴笑着 → 29.6 秒，黑屏，“PART 2?” + “FOLLOW FOR PART 2”。
</structure>

<build>
1。工具：Python、Playwright 搭配 Chromium、FFmpeg、numpy、scipy、npm 安装的 three.js。缺什么就安装。
2。一个 HTML 页面：用 three.js 画布以 1080x1920 渲染，再绘制到 2D 画布上添加字幕、手机界面、闪白和暗角。无头 Chromium 需要 --use-angle=swiftshader，因此不需要 GPU。
3。render(t) 仅根据 t 设置每个位置、灯光、表情和摄影机（是时间的纯函数），因此每一帧都精确确定。
4。完整渲染前，先将 8 个关键帧渲染为静帧，检查取景、光线和字幕是否重叠。
5。完整渲染：Playwright 将 t 设为 frame/30，并把 JPEG 帧传给 ffmpeg -c:v libx264 -crf 18 -pix_fmt yuv420p。
6。用 numpy 制作音效：低沉铺底音、风声、音乐盒前奏、钟表滴答声和 3 次钟声、伴随手机震动的木琴铃声、按键点击声、窗户轻响、带气声的 2 音节耳语、门吱呀声、逐渐加速的心跳声、渐强音、惊吓处的重击声，以及每次切镜时的呼啸声。
7。合并视频和声音，检查能否播放，然后发给我。
</build>

<inputs>
Ask me for the town's rule ("in this town, if your phone rings at 3:00, don't answer"), who breaks it and the twist. If I have none, pitch 3 and let me pick. Then write 8 shots and every caption, 1 to 3 words at a time, with timings. Show me before you build.
</inputs>

<direction>
A 1080x1920 roblox-style 3D horror short that looks filmed inside a blocky game at night.
World: a green studded baseplate, a dark road with yellow dashes, grey sidewalks, 8 toy-coloured houses with gable roofs, porches, white fences and glowing windows, blocky pine trees, street lamps that are real point lights, a clock tower, a moon, stars and blue night fog.
Characters: classic blocky avatars (cylinder head, box torso, arms and legs on shoulder and hip pivots so they sit and walk), a drawn face that changes per beat. The entity: all black, 1.75x tall, white glowing eyes, a white grin only in the last shot.
Phone close-ups: a 2D incoming-call screen slides up over the 3D shot ("Unknown", then "Mom 🏠"), ring pulse, Decline / Accept buttons with a tap ripple.
Captions: 1 to 3 words in the centre, heavy rounded font, white with a thick black outline, pop-in with overshoot. Key words yellow, danger words red. Move them to the top while the phone is on screen.
Camera: a new shot every 3.5 to 4 s, white flash on every cut, slow push-ins, shake on the bell and the scare.
Banned: real game logos or UI, stock assets, a camera that stands still.
</direction>

<structure>
0-3.6 s drone over the town: "IN THIS TOWN, / IF YOUR PHONE / RINGS AT 3:00, / DON'T ANSWER." → 3.6-7.2 s the clock tower hits 3:00, bell, shake → 7.2-11 s the kid on his porch, phone glowing → 11-14.6 s phone close-up: "Unknown", he declines → 14.6-18.4 s the street, windows going dark, it rings again: "Mom 🏠" → 18.4-22.4 s his house, every window dark: "BUT MOM / WAS ASLEEP / UPSTAIRS." → 22.4-26.4 s he picks up, heartbeat speeds up, the front door creaks open behind him: "SHE WHISPERED / "TURN AROUND."" → 26.4-29.4 s the tall figure is standing on the porch right behind him, grinning → 29.6 s black, "PART 2?" + "FOLLOW FOR PART 2".
</structure>

<build>
1. Tools: Python, Playwright with Chromium, FFmpeg, numpy, scipy, three.js from npm. Install what's missing.
2. One HTML page: a three.js canvas at 1080x1920 drawn into a 2D canvas that adds captions, the phone screen, flashes and a vignette. Headless Chromium needs --use-angle=swiftshader, so no GPU is needed.
3. render(t) sets every position, light, face and camera from t alone (pure function of time), so every frame is exact.
4. Render 8 key frames as stills and check framing, lighting and caption overlap before the full render.
5. Full render: Playwright sets t = frame/30 and pipes JPEG frames into ffmpeg -c:v libx264 -crf 18 -pix_fmt yuv420p.
6. Sound in numpy: a dark pad, wind, a music box intro, clock ticks and 3 bell strikes, a marimba ringtone with phone buzz, button taps, window clicks, a breathy 2-syllable whisper, a door creak, a heartbeat speeding up, a riser, a hard hit on the scare, a whoosh on every cut.
7. Mux video + sound, check it plays, send it to me.
</build>