gbro Collage B-roll

把一句约 5 秒的口播压成一个 sharp visual idea，再做成高级编辑风纸拼贴组装动画。

When to use this skill

Use this skill whenever:

User asks for "collage b-roll", "paper collage B-roll", "halftone collage", "纸拼贴 b-roll", "半调拼贴", or "gbro-collage-broll".
Converting voiceover lines or abstract concepts into editorial halftone paper-collage visual metaphors.
Generating assemble-from-empty 9:16 vertical video clips via Gemini Omni Flash with strict 3-gate human approval.

When NOT to use this skill:

User requires frame-by-frame exact layer manipulation, timeline editing, or transparent assets (use HyperFrames / HTML-based video renderers instead).
User only wants video prompts without rendering final MP4 clips.
User requests realistic human or product photoreal commercial footage.

Instructions

1. Prerequisites & Environment Check

Before Gate 1, run the setup validator script:

bash .agent-skills/gbro-collage-broll/scripts/check_setup.sh

Ensure GEMINI_API_KEY, ffmpeg/ffprobe, Python >= 3.10, and google-genai >= 2.10.0 (in ~/hyperframes-projects/.omni-venv) are ready.

2. 3-Gate Approval Workflow

Gate 1: Visual Metaphor Approval

Compress the ~5s script into one clear visual thesis (core meaning, emotion, action verb, 3–6 key objects, background/accent colors, assembly sequence).
Output the metaphor proposal without generating images/videos, then STOP and await user approval.

Gate 2: Still Frame Approval

After Gate 1 approval, write visual-spec.json and generate the 9:16 still frame via Codex image_gen.
Generate contact sheet still-contact-sheet.jpg and write gate2-qa.md. STOP and await user approval before calling any video model.

Gate 3: Video Generation & QA

After Gate 2 approval, prepare empty color first frame (first-frame.png) and target last frame (last-frame.png).
Execute batch Gemini Omni Flash video generation using scripts/generate_video.py (defaulting to gemini-omni-flash-preview):

~/hyperframes-projects/.omni-venv/bin/python \
  .agent-skills/gbro-collage-broll/scripts/generate_video.py \
  --batch <project>/omni-jobs.json \
  --concurrency 3

Convert output MP4 to silent format (final-5s-noaudio.mp4), run frame extraction contact sheets, and write gate3-qa.md.

Examples

Example 1: Single Line B-roll Request

Input:

collage b-roll: 很多人以为 AI 是来替你思考的，其实它更像一面镜子，会把你问题里的漏洞照出来。

Workflow:

Gate 1 proposal:
- 核心意思: AI 是照出漏洞的镜子而非替代思考
- 隐喻: 放大镜与撕纸镜面透出光芒照亮排版漏字
- 色彩: 芥末黄底，黑白半调，深红点色
Gate 2 generation & contact sheet review.
Gate 3 Omni Flash video generation & silent MP4 delivery.

Best practices

Strict 3-Gate Discipline: Never skip Gate 1 or Gate 2. Stopping for human approval saves API cost and aligns aesthetic direction before running video models.
Color Palette Diversity: Do not default all clips to cobalt blue. Use orange, mustard yellow, dark purple, teal, or forest green based on semantics.
Assemble-from-Empty Timing: Ensure Image 1 is an empty color field and Image 2 is the exact completed frame so Omni Flash performs crisp stop-motion assemble animation.
No Text/UI Clutter: Keep still frames clean from readable letters, logos, watermarks, or UI artifacts.

References

Project repo: https://github.com/pyang5166/gbro-collage-broll
Default Video Model: gemini-omni-flash-preview