guozhi-video-breakdown

Extract evidence from Xiaohongshu, Douyin, TikTok, Reels, Shorts, Bilibili, or local talking-head short videos, dissect viral hooks, captions, retention, editing, and conversion design, then generate

SKILL.md

GitHub

果汁视频拆解

Turn a short-form talking-head sample into evidence-backed viral teardown and an original script package for Xiaohongshu or Douyin.

Operating Principles

  • Evidence first: inspect frames, contact sheets, subtitles, transcript, title, cover, and visible comments before judging why it worked.
  • Separate raw facts, inferred intent, and reusable pattern.
  • Learn mechanisms, not expressions: never copy exact wording, creator persona, watermark, private screenshots, face, voice, or proprietary assets.
  • Optimize for Chinese short-video口播: first 3 seconds, subtitle rhythm, retention turns, trust proof, save/comment triggers, and a script that can be recorded.
  • Do not bypass login walls, paywalls, private posts, platform restrictions, or rate limits. If a link cannot be downloaded, ask for a local file, transcript, screenshots, or exported subtitles.

Workflow

  1. Clarify the task only when needed.

    • If the user already provides a video/link/transcript and wants拆解/复刻, proceed.
    • If the output target is missing, default to a 45-60 second Xiaohongshu口播 script.
    • Ask only if both the sample and target niche/topic are missing.
  2. Build the evidence packet.

    • For a local video or public URL, read references/evidence-extraction.md and run scripts/extract_video_evidence.sh when the environment has yt-dlp, ffmpeg, and whisper.
    • For transcript/subtitles only, skip visual claims and mark visual evidence as unknown.
    • For screenshots/contact sheets only, analyze visual rhythm and caption style but mark口播 as unknown.
  3. Dissect the viral mechanism.

    • Read references/viral-autopsy-framework.md.
    • Cover: target audience, emotion, topic promise, title/cover, first 3 seconds, segment structure, retention nodes, proof, gold lines, visual materials, editing rhythm, conversion trigger, and replication boundary.
  4. Generate an original replication package.

    • Read references/replication-output-template.md.
    • Produce: title candidates, cover text, 3 hook variants, full口播稿, timeline/storyboard, subtitle/flower-text plan, materials list, editing notes, CTA/comment trigger, and risk notes.
  5. Validate before final delivery.

    • Confirm whether contact_sheet/frames were inspected when available.
    • Confirm whether transcript/subtitles were read when available.
    • Mark missing evidence explicitly.
    • Check the new script changes audience, story, proof, wording, and assets enough to avoid copying.

Resource Routing

  • Read references/evidence-extraction.md when handling a video file, URL, audio file, subtitle, transcript, screenshot set, or contact sheet.
  • Read references/viral-autopsy-framework.md for any爆款拆解,黄金3秒, retention, or口播机制 analysis.
  • Read references/replication-output-template.md whenever the user wants复刻脚本, publish-ready口播稿, title candidates, or a production plan.
  • Read references/test-prompts.md when validating trigger behavior or reviewing this Skill.
  • Run scripts/extract_video_evidence.sh only for local files or public URLs that the user is allowed to analyze.

Validation

For Skill structure:

python3 /Users/guozhi/.codex/skills/standard-skill-maker/scripts/validate_skill.py <this-skill-folder>

For the extraction script:

bash -n scripts/extract_video_evidence.sh

For a real run, test with a short local sample first:

scripts/extract_video_evidence.sh "/absolute/path/to/sample.mp4" outputs/sample-evidence 1 small zh