Mini Explainer Workbench
Short-Video Outlier Radar, Multimodal Deconstruction & 9:16 Render Pipeline
A local full-stack short-video workbench closing the P3 Outlier Discovery → P1 Multimodal Deconstruction → P2 Multi-Asset 9:16 Rendering loop. Integrates 6 open-source creator skill modules including Humanizer-zh and marketingskills.
- Python
- FastAPI
- Gemini 2.5
- yt-dlp
- OpenCV
- edge-tts
- Pillow
- FFmpeg
- Plate
- PLATE 05
- Role
- Primary Author
- Status
- Local Full-Stack Workbench · End-to-End
- Year
- 2026
- Scope
- Primary Author: Independently built the FastAPI workbench, P3 3-funnel outlier radar, Gemini 2.5 video deconstruction pipeline, 6-skill prompt orchestration, and Pillow/OpenCV/FFmpeg 9:16 render engine.
Interactive HTML Slide Deck
Navigate with controls, keyboard arrows, slide indicators, or Fullscreen presentation mode.
Quantifies candidate topics through Outlier Multiplier, Save Potential, and Localization Angle scorecards.
- Ranks breakout videos by outlier view multiplier
- 3-funnel scorecards expose localization upside
- One-click handoff into the P1 analysis queue
- P1-P3 Full Pipeline Loop
- 6 Open-Source Skill Modules
- 9:16 Vertical Code-Driven Video
- 4.2/s char/s · Natural Pacing Target
Context & Engineering Constraints
Short-form educational creators struggle with guesswork topic selection, tedious manual frame-by-frame video breakdown, stiff AI-sounding narration scripts, and repetitive multi-format visual production.
System Architecture
- 01
P3 Outlier 3-Funnel Discovery Radar
Scores candidate videos across three funnels—Outlier Multiplier, Save Potential, and Localization Angle—to surface repeatable breakout topics.
- 02
P1 Multimodal Deconstruction & 6 Open-Source Skills
Extracts keyframes with yt-dlp and OpenCV for Gemini 2.5 pacing analysis, then rewrites scripts via 6 open-source skills including Humanizer-zh (4.2 char/s pacing) and marketingskills.
- 03
P2 Multi-Asset Studio & FFmpeg 9:16 Render Engine
Compiles one structured script into 6 dynamic slide templates, 3:4 social cards, edge-tts voice tracks, and final 9:16 vertical videos rendered via Pillow, OpenCV, and FFmpeg.
Engineering Highlights
- Connects P3 discovery, P1 deconstruction, and P2 rendering inside one reproducible local FastAPI pipeline.
- Chains 6 open-source skills including Humanizer-zh and marketingskills with a 4.2 char/s pacing constraint to produce natural spoken narration.
- Replaces repetitive timeline layout with code-driven templates outputting slide frames, 3:4 social cards, and synced 9:16 FFmpeg renders.
- Runs cleanly as a modular local Python stack free from closed SaaS video-editor lock-in.