Video Viral Breakdown and Copywriting Generation

Problem

Short-video virality often comes from hooks, pacing, conflict, CTAs, and the timing of visuals, speech, and on-screen text. These signals are scattered across audio, frames, and OCR output, making manual breakdown slow and hard to reuse for new scripts. This skill turns a proven short video into structured data and then supports rewrite copy.

Workflow

• Input: accepts local video files or share links from Douyin, Bilibili, Xiaohongshu, WeChat Channels, and YouTube; links are downloaded before parsing.
• Breakdown: pipeline.py uses ffmpeg, faster-whisper, and easyocr to extract audio, keyframes, and subtitles, producing a time-ordered timeline with speech, frame OCR, and representative frames.
• Copywriting: reads analysis.json and references/prompt_templates.md, distills the original video’s viral formula, and generates multiple rewrite drafts from different angles.
• Runtime: first run creates an isolated environment, installs dependencies, and downloads models; later runs use local cache, and CPU fallback is available without a GPU.

Boundaries

It fits short-video structure analysis and derivative copy, not pure audio transcription, precise editing, or frame-level CV detection. Long videos, music-only clips, and complex caption styles can hit GPU memory, OCR misses, or Whisper hallucinations, so tune model size, frame sampling, and device selection.