shorts

claude-shorts

Interactive longform-to-shortform video creator powered by Claude Code. Extracts viral-ready vertical clips from long videos using Claude as the intelligent orchestrator with Remotion-rendered premium animated captions.

How It Works

Claude Code guides you through a 10-step interactive pipeline:

1. Preflight - Validates input video, checks disk space, detects GPU
2. Transcribe - GPU-accelerated transcription with word-level timestamps (faster-whisper)
3. Detect Content - Auto-classifies: talking-head, screen recording, or podcast
4. Analyze - Claude reads the full transcript and scores 8-12 candidate segments
5. Present - Shows candidates in a formatted table with scores, hooks, and rationale
6. Approve - You pick segments, adjust timecodes, choose caption style and platform
7. Snap Boundaries - Aligns cut points to word boundaries, sentence endings, and audio silences
8. Prepare - Extracts clips (FFmpeg stream copy) and computes reframe coordinates
9. Render - Remotion renders 1080x1920 vertical video with animated captions
10. Export - Platform-optimized encoding (YouTube Shorts, TikTok, Instagram Reels)

Demo

Demo video/GIF coming soon — showing the full pipeline from input to rendered short with Bold-style captions.

Features

• Claude-powered segment scoring - 5-dimension rubric (hook strength, coherence, emotion, value density, payoff) with weighted scoring. No heuristic keyword matching - Claude understands narrative arcs.
• 3 caption styles - Bold (ALL CAPS, yellow highlights), Bounce (bouncy spring, rotating colors), Clean (minimal fade-in)
• Cursor tracking - For screen recordings, detects mouse cursor via frame differencing and smoothly pans the crop to follow it
• Audio-aware boundary snapping - Never cuts mid-word or mid-sentence. Extends to natural sentence endings and silence points.
• Remotion rendering - React-based single-pass rendering with spring animations, word-level karaoke highlighting, hook text overlays, and progress bars
• GPU acceleration - CUDA for transcription, NVENC for export encoding (falls back to CPU gracefully)

Prerequisites

• FFmpeg (system package)
• Python 3.10+
• Node.js 18+
• Claude Code (CLI)
• NVIDIA GPU recommended (for CUDA transcription + NVENC encoding)

Installation

# Clone the repository
git clone https://github.com/AgriciDaniel/claude-shorts.git
cd claude-shorts

# Install Python + Node.js dependencies
bash setup.sh

# Install as a Claude Code skill
bash install.sh

Windows

claude-shorts requires Unix tools (FFmpeg, bash). On Windows, use WSL 2:

# Inside WSL
git clone https://github.com/AgriciDaniel/claude-shorts.git
cd claude-shorts
bash setup.sh
bash install.sh

What setup.sh does

• Creates a Python virtual environment at ~/.shorts-skill/ (or reuses ~/.video-skill/ if it exists)
• Installs faster-whisper, mediapipe, numpy, opencv-python, and PyTorch (CUDA or CPU variant)
• Runs npm install in the remotion/ directory
• Checks for system dependencies (FFmpeg, jq)

Usage

In Claude Code, invoke the skill:

/shorts

Then provide your video file when prompted. Claude will:

1. Transcribe the video
2. Present scored segment candidates
3. Ask which segments to render, caption style, and target platform
4. Render and export the final shorts

Example Interaction

You: /shorts ~/Videos/my-talk.mp4
Claude: [Transcribes, detects content type, scores segments]

| # | Time          | Dur  | Score | Hook                           |
|---|---------------|------|-------|--------------------------------|
| 1 | 04:22 - 05:01 | 39s  | 87    | "Nobody talks about this..."  |
| 2 | 12:45 - 13:28 | 43s  | 82    | "Here's the exact framework." |
| 3 | 08:11 - 08:52 | 41s  | 79    | "I tested this for 6 months." |

Claude: Which segments? Caption style? Platform?
You: 1 and 3, bounce style, youtube

Claude: [Snaps boundaries, extracts clips, renders, exports]
Output: shorts/short_01_yt.mp4, shorts/short_03_yt.mp4

Project Structure