video-pipeline

name	video-pipeline
description	End-to-end video production: design system → TTS → scene verification → final render. Uses ui-ux-pro-max for design, Fish-TTS for voice, video-gen (Remotion) for rendering. Actions: make video, create video, 영상 만들어, 샘플 영상, video sample. Always verify scenes before final render.

Video Pipeline — Design → Voice → Verify → Render

Full video production pipeline integrating 3 skills with a mandatory verification phase.

When to Use

When user requests video creation, sample videos, or "영상 만들어줘":

• Product demos, explainers, social content
• AI/tech concept visualizations
• Personal brand content
• Any YAML-scripted video

Pipeline Overview

Phase 1: UI/UX Pro Max → Design System (style, colors, fonts)
Phase 2: YAML Script → Fish-TTS Audio + Scene Config
Phase 3: Scene Verification → Render stills, review each scene ⚠️ MANDATORY
Phase 4: Final Render → Only after verification approval

Dependencies

Skill	Location	Purpose
ui-ux-pro-max	~/.claude/skills/ui-ux-pro-max/	Design system generation
fish-audio	~/2lab.ai/skills/fish-audio/	Voice reference files
video-gen	~/2lab.ai/skills/video-gen/	Remotion-based video rendering

Step-by-Step Workflow

Phase 1: Design System

Generate a design system using UI/UX Pro Max based on the video's topic:

python3 ~/.claude/skills/ui-ux-pro-max/scripts/search.py \
  "<topic keywords>" --design-system -p "<Project Name>"

Output gives you:

• Recommended style (Cyberpunk, Minimal, etc.)
• Color palette (primary, secondary, CTA, bg, text)
• Typography pairing (heading + body fonts)
• Key effects (neon glow, glassmorphism, etc.)

Map to video-gen theme: Match the recommended style to the closest video-gen theme:

UI/UX Pro Style	video-gen Theme
Cyberpunk UI	neon
Dark Mode OLED	default or bigbrain
Minimalism	minimal
Glassmorphism	default
Retro-Futurism	evangelion
Warm/Organic	warm
Nature	nature
Elegant/Serif	cinematic or manuscript
Notion-style	notion

Phase 2: Write YAML Script + Generate TTS

Write a YAML script file. See ~/2lab.ai/skills/video-gen/SKILL.md for full scene types.

Available voices: elon, iu, karina, egirl (in ~/2lab.ai/fish-tts/voices/)

Available scenes: hero, list, grid, code, flow, chat, stat, quote, timeline, comparison, emoji, image, bigtext, terminal, tokenPredict, matrixRain, progressBar, radar, countdown, map, callout, splitScreen, typewriter, pyramid, reveal, infographic, trainingData, systemPrompt, conversation

Run Phase 1+2 together:

bash ~/2lab.ai/skills/video-pipeline/scripts/pipeline.sh <script.yaml> <output.mp4>

This stops after Phase 3 (verification) and waits for approval.

Phase 3: Scene Verification ⚠️ CRITICAL

This phase is MANDATORY. Never skip.

After TTS + metadata generation, render each scene as a still PNG:

bash ~/2lab.ai/skills/video-pipeline/scripts/verify-scenes.sh <workspace_dir> /tmp/scene-verify

Verification workflow:

1. Script renders PNG for each scene's midpoint frame
2. Images saved to /tmp/scene-verify/scene-001.png, scene-002.png, etc.
3. Read each PNG with the Read tool to visually inspect
4. Check for:
   • Text readability (not cut off, correct content)
   • Color/contrast issues
   • Layout problems (overlapping elements)
   • Theme consistency
   • Data display correctness (stats, lists, code)
5. If issues found → modify video-config.json in workspace and re-verify
6. If all scenes pass → proceed to Phase 4

Sending stills to user:

• Use mcp__send-file__send_photo to send each scene PNG to user
• User can approve or request changes per scene

Phase 4: Final Render

Only after all scenes are verified:

WORKSPACE=<workspace_dir> bash ~/2lab.ai/skills/video-pipeline/scripts/render-final.sh <output.mp4>

Or if using the pipeline script with verification skipped:

SKIP_VERIFY=1 bash ~/2lab.ai/skills/video-pipeline/scripts/pipeline.sh <script.yaml> <output.mp4>

Quick Reference: One-Shot (No Verification)

For rapid prototyping when verification isn't needed:

bash ~/2lab.ai/skills/video-gen/scripts/build-from-script.sh <script.yaml> <output.mp4>

Quick Reference: Full Pipeline (Recommended)

# 1. Design system
python3 ~/.claude/skills/ui-ux-pro-max/scripts/search.py "AI tech dark" --design-system -p "My Project"

# 2. Write YAML (use design system recommendations)
# 3. Run pipeline (stops at verification)
bash ~/2lab.ai/skills/video-pipeline/scripts/pipeline.sh script.yaml output.mp4

# 4. Review stills
# Read /tmp/scene-verify/scene-001.png ... scene-N.png

# 5. Final render (after approval)
WORKSPACE=/tmp/video-pipeline-xxx bash ~/2lab.ai/skills/video-pipeline/scripts/render-final.sh output.mp4

Content Tips

• Don't be generic — Use real quotes, real data, real stories
• Vary scene types — Don't repeat the same scene type twice in a row
• Terminal scenes — Great for "hacker" feel, real commands
• Code scenes — Show actual algorithms, not pseudocode
• Stat scenes — Use real numbers with units
• Quote scenes — Real quotes hit harder than fabricated ones
• Flow/Timeline — Good for processes, history, concept progressions

File Structure

~/2lab.ai/skills/video-pipeline/
├── SKILL.md                    # This file
└── scripts/
    ├── pipeline.sh             # Full 4-phase pipeline orchestrator
    ├── verify-scenes.sh        # Phase 3: render stills per scene
    └── render-final.sh         # Phase 4: final render after approval