Communitygithub.com

wwwmktg75-crypto/short-video-producer

テーマやメモから原稿、合成ナレーション、シーン別画像、BGMを制作・編集し、縦型ショート動画を完成させる。録音済みラジオの編集ではなく、新規ショート動画の制作依頼に使う。

short-video-producer 是什么?

short-video-producer is a Codex agent skill that テーマやメモから原稿、合成ナレーション、シーン別画像、BGMを制作・編集し、縦型ショート動画を完成させる。録音済みラジオの編集ではなく、新規ショート動画の制作依頼に使う。.

兼容平台~Claude Code✓Codex CLI~Cursor
npx skills add wwwmktg75-crypto/short-video-producer

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

Short Video Production Agent

Produces a complete MP4 from a theme. Codex handles the script, direction, and asset creation, while included scripts composite images and BGM according to the actual measured duration of the voiceover. This is a production skill invoked via chat, not a resident service.

Input and Default Settings

  • Accepts a theme, notes, a URL to references, or an existing script. If no theme is provided, it asks for one.
  • Unless specified otherwise: Japanese language, 9:16 aspect ratio, 1080×1920, 30fps, approximately 45 seconds, a friendly conversational tone. Respect any explicit instructions about duration, voice, visual style, or publishing platform.
  • Does not automatically insert program names, AI character personalities, radio greetings, or lead-ins to the main content.
  • Output is gathered under output/shorts/<datetime>-<theme>/. Revised versions go into a new folder.

Production

  1. Narrow down the theme to a single key message. Write a script with a hook at the beginning, followed by a concrete example/explanation, and a brief summary. Aim for roughly 120–150 characters for 30 seconds as an initial target; finally adjust based on the actual measured length of the generated voiceover. Verify the latest information and numerical values from references and primary sources, and record the references.
  2. Divide the script into scenes by meaningful chunks. For each scene, set the spoken text, image instructions, and on-screen short text. Proceed with production unless a structure review is requested.
  3. Check which image generation tools and their skills are available, then generate scene images. Unify the visual style, characters, and colors, and be mindful of the vertical safe zone. Do not embed long script text inside images. If the required generation tool is unavailable, communicate the situation and proceed with feasible steps such as using user-provided materials. Do not treat placeholder images as finished assets.
  4. Create scene-by-scene narration using an available TTS system. If using an API, check its current official specification and receive authentication credentials via environment variables, etc. For local macOS voices, you can use say -v Kyoko -f narration.txt -o narration.aiff (first verify availability with say -v '?'). Record the voice method used. If the desired voice is unavailable, do not silently replace it with a different voice. Check for misreadings, unnatural pauses, and duration; if too long, shorten the script and regenerate.
  5. Choose BGM from user-provided tracks, sound sources with verifiable usage terms, or available music generation tools. Record the track name, source, and usage terms in credits.md. If no BGM is specified and no source is available, the included make_bgm.py can create a simple original instrumental backing track. This is a simple accompaniment synthesized from note data; do not describe it as output from an AI music generator. Lower the volume during narration.
  6. Save a manifest.json in the following format and run render.py. Paths to images, audio, and BGM should be relative to the manifest or absolute. Do not speed up or truncate narration; create each scene with the audio duration plus about 0.2 seconds. The standard renderer supports still image transitions and attached SRT subtitles. If motion or burned-in subtitles are needed, implement them with Remotion or other tools before delivery. Do not claim that attached SRT alone constitutes a captioned video.
  7. Verify the resolution, audio/video sync, duration, and overall decode of the MP4. Use extracted frames to check for image cuts or readability issues. If preview listening is possible, check volume, pronunciation, and balance with BGM. If preview is not possible, clearly note that limitation. If the final duration deviates significantly from the requested length, adjust the script and audio.
  8. Return the finished MP4, script, SRT, manifest, and asset sources as links. Distinguish between actually completed steps and unchecked points. Only post the video if specifically requested.

Composition Commands

Required environment: Python 3, ffmpeg, ffprobe. No additional Python packages are necessary.

{
  "bgm": "bgm.wav",
  "scenes": [
    {"image": "scene-01.png", "audio": "scene-01.aiff", "text": "This is the script to be read."},
    {"image": "scene-02.png", "audio": "scene-02.aiff", "text": "This is the script for the next scene."}
  ]
}
python3 <skill-dir>/scripts/make_bgm.py /absolute/output/bgm.wav --seconds 60
python3 <skill-dir>/scripts/render.py /absolute/output/manifest.json --output /absolute/output/render-v1

render-v1/final.mp4, captions.srt, and timing.json are output. Scene subtitles use scene-level timestamps, not word-level synchronization. On failure, reuse already completed assets, fix the cause, and re-run with a new output destination.

相关技能