Communitygithub.com

phone-screen-horror

制作"模拟恐怖"类竖屏手机屏幕叙事短片(观众看主角的手机界面操作推进剧情)。界面用 HTML 界面套件渲染或用图像模型生成,做出带点击反馈、拼音输入、滚动、应用开合、通知、锁屏日期推移的逐镜头动画,合成分轨声音(环境/键盘/UI/人声/配乐/警笛),导出 MP4,可选剪映专业版草稿。用户提到手机屏幕叙事、模拟恐怖、手机界面短片、聊天记录短片、剪映草稿时使用。

phone-screen-horror 是什么?

phone-screen-horror is a Claude Code agent skill that 制作"模拟恐怖"类竖屏手机屏幕叙事短片(观众看主角的手机界面操作推进剧情)。界面用 HTML 界面套件渲染或用图像模型生成,做出带点击反馈、拼音输入、滚动、应用开合、通知、锁屏日期推移的逐镜头动画,合成分轨声音(环境/键盘/UI/人声/配乐/警笛),导出 MP4,可选剪映专业版草稿。用户提到手机屏幕叙事、模拟恐怖、手机界面短片、聊天记录短片、剪映草稿时使用。.

兼容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/doublerabbit-star/phone-screen-horror/tree/HEAD

在你喜欢的 AI 中提问

打开一个已预加载此 Agent Skill 的新对话。

文档

Phone Screen Narrative Short Film (Analog Horror) Production

Tools are located in scripts/ and ui_kit/, with detailed usage in references/. An example project "Hiring Terminated" (6 min 45 sec, 141 shots, 11 scenes) has its scene scripts in examples/yingpin/. For new projects, read this file first, then refer to references/ as needed.

Guidance for Zero-Experience Users

When the user says "I don't know how," "first time," or only provides a single sentence of story, follow these steps to guide them through:

  1. First confirm it can actually be made: Check local file reading, command execution ability, Python, required packages, ffmpeg/ffprobe, browser, and Chinese fonts. Checking paths and system can be done by the tool; don't repeatedly ask the user. If the tools cannot run, clearly state that you can only organize the story now, not claim a video was generated.
  2. Lower the story input barrier: A one-sentence idea is enough to start; don't send a long form first. If the user lacks an idea, give three short topic suggestions; if they already have an idea, organize it into a short paragraph of plot and main on-screen text. Only ask crucial questions that would change the story's direction; fill in details with reasonable defaults and explain them.
  3. Clearly state the next step: For example, "I will first turn this story into a playable preview with sound. After you watch it, tell me where it's too fast or not tense enough." Do not dump script names, parameters, and directory structure on the user, nor make them choose technical solutions.
  4. First exercise controls complexity: If the user has not specified scale, suggest about one minute, a few scenes, telling a simple story through chat, notifications, and lock screen. Default to delivering only MP4. Adjust based on actual reading time; don't speed up text to fit a duration. If the user already has a complete story, preserve its scope; don't force it into a practice clip.
  5. Minimize asset dependencies: Prioritize HTML interfaces, text avatars, and simple graphics. Only list specific material needs if the story requires photos. If image tools in the current environment can help, assist with creation. If generation is not possible, provide prompts and ask the user to provide. Don't start by asking a beginner to prepare a large number of screenshots.
  6. Actually advance production: After the story direction is clear and the necessary environment is available, produce a full audio preview from start to end, then collect feedback. The user only needs to describe modifications using feelings like "too fast to read," "sound too loud," "ending needs to be slower," no need to provide shot numbers.
  7. Give specific help when user action is needed: Handle Python package installations within your authorization. For system software installation or authorization requiring user action, explain what's missing, its impact, and the next step; handle only the current obstruction at a time, don't dump the entire environment checklist on the user.

You can start by saying: "Just tell me a story idea, and I'll turn it into chats, notifications, and sounds on a phone. First, check the production environment, then make a playable preview for you to watch." If the user has already given the story, proceed directly to organizing; don't ask for it again.

Default Settings (Adopt directly; only ask for truly missing information that affects production)

  • The user may not know how to program. You run all commands; don't ask the user to install dependencies, run scripts, or change code. If missing Python packages, run pip install yourself. If missing ffmpeg, browser, or CapCut, tell the user in one sentence what to install and where to download it.
  • The user provides the story; you are responsible for the interface, animation, sound, and export. If the user already has screenshots (e.g., made in GPT), use theirs. If not, use the HTML UI kit to make them; for avatar, wallpaper, face, and other photo-type assets, ask the user to generate them with an image model; you provide the prompts (references/image_prompts.md).
  • CapCut is optional. Without CapCut, still deliver MP4 and six separate WAV tracks.
  • System: Windows or macOS. Voice uses system TTS (Windows OneCore, macOS say). Only actually verified on Windows + CapCut Professional 11.8. For Mac and CapCut, explain to the user that it's untested.

Hard Rules

  1. The deliverable is a playable video with sound. Scripts, screenshots, silent tracks, and plans are not considered completion.
  2. If not listened to, write "not listened"; if not measured, write "not measured"; if the CapCut draft was not actually opened, write "pending verification." Do not forge projects or check results.
  3. Do not repeatedly ask about already confirmed settings. If there is doubt about asset versions, check the screen or checksum yourself.
  4. The screen shows only the phone screen: no phone frame, fingers, click circles, mouse, or camera moves. No stretching. 1170×2532, 30fps.
  5. Horror comes from everyday life: no supernatural elements, glitch text, jump scares, narration, or explanatory subtitles. When involving death or suicide, don't show the method or process; only imply through unreturned responses, time passing, and distant sirens.
  6. Do not submit text-dense screens to video generation models; render them all deterministically (PIL). Only small parts like face windows can be animated.
  7. Never overwrite a CapCut draft the user has opened; new versions get new names (e.g., _roughCutV4, _final2).

Production Workflow

1. Brief

First, accept the user's one-sentence or paragraph description of the story. When organization is needed, use the short entry from references/brief_template.md; only use the full template when the user wants detailed planning. Characters, on-screen text, scenes, and sounds are organized by you; fill in missing parts with reasonable defaults and briefly explain the assumptions. references/brief_example.md is the complete brief for "Hiring Terminated" (the duration "about 90 seconds" was later overturned by feedback).

2. Set Up Project

<project>/project.json        Copy from templates/project.json, fill in title, voices
<project>/00_ui/              HTML interfaces
<project>/01_final_assets/    All images (engine finds by filename, filenames unique across project)
<project>/03_project/tools/   Copy all from scripts/ + ui_kit/ folder; scenes.py, tap_rects.py also written here

02_audio, 04_cache, 05_exports are created automatically by the tool. Run all commands in 03_project/tools. Fonts, browser, TTS, CapCut draft directory are all found automatically; if not found, specify them in project.json (field descriptions in references/pipeline.md).

3. Interface Assets

  • HTML UI Kit (preferred, see references/ui_kit.md): One HTML per interface, placed in 00_ui/S01/…. python ui_render.py renders to 1170×2532 PNG, automatically generates _blank/_text layers for typing, click areas .taps.json, and face window positions .ui.json. Chinese must be accurate; keyboard must match keyboard.py keys.
  • Image Model Screenshots: Place in 01_final_assets/S01/…. For typing interfaces, use engine.mask_layers to split layers. For click areas, use sheet.py to generate a coordinate grid mosaic, measure them, and write them into tap_rects.py. Verify Chinese characters in images one by one.
  • After rendering, use Read to check a few PNGs, confirming text, layout, and day/night theme are correct.

4. Write Scenes

In tools/scenes.py, each scene is a function with the name S + two digits (S01…S99). Template and API in templates/scenes_template.py:

from scene import *

def S01():
    c = Scene('S01', mood='day_bright', k=1.5, music='calm')
    s = c.shot('wake'); fade_in(s, img('lock.png')); s.hold(img('lock.png'), 2.0 * c.k); c.add(s)
    c.hold('inbox', 'inbox.png', 1.6, 'notify', ('tap', 1.2))
    c.save()

if __name__ == '__main__': run(globals())

typ() for typing (visible keyboard, pinyin, space to commit), V() for voice, apptrans.open_app/close_app for app open/close animation, engine.scroll/face/timer/xfade, etc. Every click must have a TAPS entry; otherwise there will be no press feedback (rendering will print [tap] no rect for …).

5. Low-Resolution Full Film with Sound

YP_RES=proxy (default, 586×1268) render all scenes → python build_audio.py → python export.py <name> proxy. Start by making a full film with sound covering the entire piece, not a polished 30-second clip first. export.py prints duration, number of shots, and loudness (LUFS / peak).

6. Self-Check

  • python peek.py S04 generates a thumbnail mosaic of a scene's frames. Use Read to check scene by scene: text, pressed states, keyboard, transitions, black frames.
  • Check the loudness printed by export. You can only measure numerical values for sound; when reporting, must state you have not listened.

7. Local Modifications Based on Feedback

  • Pacing: Modify k or individual shot seconds; only re-render affected scenes (python scenes.py S05 S07), then build_audio + export.
  • Only change sound (music, sound effects, voice settings): Don't re-render visuals; directly build_audio + export.
  • Interface text or layout: After modifying HTML, run python ui_render.py (re-render only changed HTML), then re-render scenes that use it.
  • Each new MP4 uses a new name (export.py does not overwrite files with the same name). For long projects, maintain a RECOVERY.md in the project (current version, commands, to-do) so you can continue after context interruption.

8. HD and CapCut

After user approval, set YP_RES=hd and re-render all scenes (1170×2532, much slower than low-res, can be parallelized per scene) → build_audio.py → export.py <name> hd → If user wants CapCut, build_jianying.py <name> hd.

  • build_jianying.py always creates an engineering package (05_exports/<name>_工程包/: assets, six separate tracks, README.txt); only writes a draft if pyJianYingDraft is installed; only copies it into the CapCut draft directory if found; refuses to run if a draft with the same name already exists.
  • On a new computer or new CapCut version, first run jy_roundtrip_test.py, ask the user to open the test draft and tell you if it plays correctly.
  • Drafts are always written as "pending verification" before the user opens them. On Windows, winshot.py can be used to capture the CapCut window for auxiliary confirmation.

Feedback from the First Project (Apply by default to new projects)

See references/lessons.md for details.

  • Pacing should be slow: The first version at 94 seconds was rated "too fast"; the final 6 minutes 45 seconds satisfied the user. Hold duration should be based on "can read it through"; especially slow for information-dense pages like forms and assessments. Prioritize readability first; don't be bound by the total duration written in the brief.
  • Normal speed: Typing about 7.5 keys/second, pause 0.22s after each word is committed. Voice at normal speech speed; no speeding up or cutting off.
  • Every click requires a visual change on the interface (button darkens or fades, line highlights), otherwise the audience doesn't know where they clicked.
  • Typing must show the keyboard: Key presses popping up, pinyin with underline growing letter by letter, candidate bar changes, space to commit the entire word. Characters cannot appear out of thin air.
  • Going from desktop to app must have an opening animation (icon zooming out and expanding), and returning to desktop must have a closing animation.
  • Scrolling sound must be noticeable.
  • Music is instrumental: gentle and melodic early on, gradually accumulating discomfort scene by scene, becoming "uncomfortable enough" by the end. Rely on material changes, not just volume.
  • Ending must show passage of time: lock screen date, time, battery charge changing day by day, notifications piling up, battery turning red, finally distant sirens (example in examples/yingpin/scripts/lockdays.py).

Reference Files

  • references/pipeline.md: Usage of each script, project.json fields, sound events, CapCut notes.
  • references/ui_kit.md: Writing the HTML UI kit, components, layers, click areas.
  • references/image_prompts.md: Prompts for GPT and other image models, plus how to split layers and measure click areas after getting images.
  • references/brief_template.md: Brief template for the user to fill out.
  • references/lessons.md: Feedback, changes, and technical pitfalls for each version of the first project.
  • references/brief_example.md: Original brief for "Hiring Terminated."
  • examples/yingpin/scripts/: Scene scripts for "Hiring Terminated" (scenes_v2.py includes its own typ/V, equivalent to scene.py's writing style) and specialized tools. Contains only scripts, no original screenshots, and cannot be run directly; used as reference for shot writing.

Reporting Format

Use Chinese, plain speech, no jargon. Clearly state: what changes were made; output file path and duration; loudness values and note "not listened"; CapCut draft name and whether verified; text baked into the screen cannot be changed in CapCut; only lens re-rendering can change it.

相关技能