Local Video Clip Engine with AI

# Local Video Clip Engine — SoloStack Skill File
> Drop this file into your SoloStack repo and Claude Code will use it as
> context to run this workflow automatically.
---
## Overview
Build a local Opus Clip replacement — Whisper for transcription, FFmpeg for rendering, ASS karaoke captions, zero subscriptions.
**Apify Actor:** `N/A — fully local pipeline`
**Pricing:** $0 (runs entirely on your machine)
## Capabilities
1. **Word-by-Word Karaoke Captions** — ASS subtitle \kf tags progressively fill each word with brand yellow (#FEBB02) as it's spoken — the exact same visual effect as Opus Clip, with full control over fonts, colors, and positioning.
2. **Automatic Clip Selection** — Heuristic scorer slides a window across the transcript and ranks segments by engagement signals: word density, sentence boundaries, questions, keyword triggers like 'here's the thing' and 'the key is'.
3. **Hardware-Accelerated Rendering** — FFmpeg's VideoToolbox encoder uses Apple Silicon hardware to render clips 10-20x faster than software encoding. A 60-second clip renders in ~3 seconds at 8Mbps quality.
4. **Flexible Output Formats** — Crop to vertical (9:16 for Reels/TikTok), square (1:1 for Instagram feed), or keep original aspect ratio. One command flag controls the output format.
5. **Fully Local & Private** — Nothing leaves your machine. No cloud uploads, no API keys for transcription, no render queues. Your raw podcast footage stays on your disk.
6. **Human-Editable Subtitles** — Captions are standard ASS subtitle files — open them in any text editor to fix typos, adjust timing, or change styling before the final render.
## Pipeline
1. **Probe video** — FFprobe extracts resolution, duration, fps, and codec info from your source video in milliseconds
2. **Transcribe with Whisper** — OpenAI Whisper runs locally on your Mac — word-level timestamps, no API calls, no cloud upload, ~1x realtime
3. **Score & select clips** — Sliding window scorer finds the most engaging 30-90s segments based on word density, questions, keyword triggers, and pause patterns
4. **Generate karaoke captions** — ASS subtitle generator creates word-by-word animated captions with \kf karaoke timing — identical to Opus Clip's style
5. **Render with FFmpeg** — FFmpeg center-crops to 9:16, burns captions, and encodes via VideoToolbox hardware acceleration — ~3 seconds per 60s clip
## Use Cases
- **Podcast clip repurposing:** Turn a 60-minute podcast episode into 5-10 vertical clips with animated captions, ready for TikTok, Reels, and YouTube Shorts — in under 2 minutes of render time.
- **Interview highlight reels:** Score and extract the most quotable moments from recorded interviews. The heuristic scorer catches questions, keyword triggers, and natural pause boundaries.
- **Course content snippets:** Chop online course recordings into bite-sized clips for social media promotion. Each clip gets word-by-word captions that make it watchable on mute.
- **Webinar repurposing:** Extract the best segments from hour-long webinars for LinkedIn, Twitter, and email campaigns. Square format for feed posts, vertical for Stories/Reels.
## Warnings
- Whisper's medium model downloads ~1.5GB on first run. After that, it's cached locally and runs offline.
- Transcription runs at ~1x realtime on M-series Macs — a 60-minute podcast takes ~60 minutes to transcribe. Use the 'small' model for faster (but less accurate) results.
- Clip selection uses heuristics, not AI — it finds likely-good segments, but you should review the candidates and pick your favorites. The selection is 80% as good as AI-powered tools.
- Center-crop works for most talking-head content. If your video has important action at the edges, you may need to adjust the crop offset manually.
## Sample Output
```json
{
  "segments": [
    {
      "start": 12.04,
      "end": 13.12,
      "text": "Here's what most people get wrong about content repurposing.",
      "words": [
        { "word": "Here's", "start": 12.04, "end": 12.28, "probability": 0.97 },
        { "word": "what", "start": 12.30, "end": 12.45, "probability": 0.98 },
        { "word": "most", "start": 12.48, "end": 12.62, "probability": 0.96 },
        { "word": "people", "start": 12.65, "end": 12.88, "probability": 0.99 },
        { "word": "get", "start": 12.92, "end": 13.01, "probability": 0.95 },
        { "word": "wrong", "start": 13.04, "end": 13.12, "probability": 0.97 }
      ]
    }
  ]
}
```
Download the complete skill file with full implementation details,
agent prompts, and ready-to-run scripts.