skillZs
★ LIVE SKILL TAGS ★
>>> LIVE SKILLS INDEX <<<
* OPEN SOURCE *
NO LOGIN, NO TRACKING
※ REAL INSTALL DATA ※
← back to all skills
agricidaniel/claude-video94 installs

claude-video-shorts

Longform-to-shortform video pipeline that automatically extracts viral-ready short clips from long videos. V3 pipeline with VLM-powered visual analysis (Gemini 2.5 Flash), auto-detects screen recordings vs talking-head, uses "Framed" layout for screen content (padded top/bottom with centered captions), face-tracked crop for talking-head. Transcribes with WhisperX, scores by visual interest + transcript quality, adds karaoke captions, hook text, normalizes audio. Use when user says "shorts", "clips", "highlight", "shortform", "reels", "tiktok clips", "extract clips", "viral clips", "repurpose", "longform to shortform", "auto clip", "find best moments", or "screen recording shorts".

How do I install this agent skill?

npx skills add https://github.com/agricidaniel/claude-video --skill claude-video-shorts
view source ↗

Is this agent skill safe to install?

  • Gen Agent Trust Hubpass

    The skill implements a video processing pipeline using standard open-source libraries and local scripts. It is generally safe but includes an inherent low-risk attack surface for indirect prompt injection via the media content it analyzes.

  • Socketpass

    No alerts

  • Snykpass

    Risk: LOW · No issues

What does this agent skill do?

claude-video-shorts — Longform to Shortform Pipeline (V3)

One-Command Pipeline

V3 Screen-Aware Pipeline (recommended)

For screen recordings, tutorials, and mixed content — uses VLM for visual analysis:

source ~/.video-skill/bin/activate && bash scripts/screen_shorts_pipeline.sh "$INPUT" \
  [--count N] [--min-duration S] [--max-duration S] [--style STYLE] \
  [--platform PLATFORM] [--skip-captions] [--skip-vlm] [--hook-text "Your hook here"] \
  [--mode screen|talking-head|auto]

V1 Audio-Only Pipeline (legacy)

For simple talking-head content where face tracking is sufficient:

source ~/.video-skill/bin/activate && bash scripts/shorts_pipeline.sh "$INPUT" \
  [--count N] [--duration S] [--style STYLE] [--platform PLATFORM]

V3 Pipeline Architecture

INPUT (any video) → Auto-detect mode (screen vs talking-head)
        │
  [1] WhisperX transcription (word-level timestamps)
  [2] PySceneDetect scene boundaries
  [3] Extract keyframes at scene cuts + intervals
  [4] Auto-detect: face rate <60% → screen mode
  [5] VLM analysis (Gemini 2.5 Flash) → visual interest + zoom regions
  [6] Topic segmentation + multi-modal scoring
  [7] Per-topic: extract → reframe → caption → normalize
  [8] Export for platform → manifest JSON

Framed Layout (Screen Mode)

The validated layout for screen recordings (1080x1920):

+--------1080px--------+
|                       |
|  HOOK TEXT (200px)    |  "SEO Health Score: 57/100"
|  Compelling data pt   |  "What Claude found in seconds"
|                       |
+=======================+
|                       |
|  SCREEN CONTENT       |  Cropped from source, scaled to 1080x1372
|  (1372px)             |  VLM-selected crop position
|                       |  ~63% of source resolution
|  Score tables,        |  Readable on phone
|  data, charts...      |
|                       |
+=======================+
|                       |
|  KARAOKE CAPTIONS     |  Impact 72pt, yellow sweep fill
|  (348px)              |  Dark box background (BorderStyle=4)
|  Centered in zone     |  MarginV=240, word-by-word \kf timing
|                       |
+-----------------------+

FFmpeg Filter Chain

crop={W}:{H}:{X}:0 → scale=1080:1372 → pad=1080:1920:0:200:black
→ drawbox (chrome mask) → drawtext (hook) → ass (captions)

Crop Position Logic

VLM analyzes each frame and suggests zoom regions. The crop position (X) varies:

  • Left-aligned content (tables, lists): x_pct ~0.22
  • Center-aligned content (title pages, score circles): x_pct ~0.35
  • Right-aligned content (sidebars, panels): x_pct ~0.55

Multiple zoom regions within a clip create jump cuts (not smooth pan) — this matches natural scroll/section transitions in screen recordings.

Multi-Modal Scoring (V3)

FactorWeightSourceDescription
Visual interest0.30VLMCharts, tables, data = high score
Content completeness0.25VLM + transcriptComplete data shown, full thoughts
Audio engagement0.20TranscriptEmphasis words, questions, emotion
Hook potential0.15TranscriptFirst sentence strength
Standalone coherence0.10TranscriptWorks without context

Caption Style (V3)

The validated caption parameters for the framed layout:

Font: Impact, 72pt, Bold
Primary: White (&H00FFFFFF)
Secondary/Karaoke: Yellow (&H0000FFFF)
Outline: Black, 4px
Background: Semi-transparent black (&HC0000000)
BorderStyle: 4 (opaque box behind text)
Alignment: 2 (bottom-center)
MarginV: 240 (centered in 348px bottom zone)
Letter spacing: 2
Animation: \kf word-by-word karaoke sweep
Words per line: 3

Flags Reference

FlagDefaultDescription
--count N5Number of shorts to extract
--min-duration S20Minimum clip length in seconds
--max-duration S55Maximum clip length in seconds
--style STYLEboldCaption style: default, bold, minimal, neon, shadow
--platform PLATshortsExport: tiktok, reels, shorts, all
--skip-captionsfalseSkip caption burn-in
--skip-vlmfalseSkip VLM analysis (audio-only scoring)
--hook-text TXTautoCustom hook text for all shorts
--language LANGenTranscription language
--output-dir DIRinput dirOutput directory
--vlm-model MODgemini-2.5-flashVLM model for frame analysis
--max-frames N40Max frames for VLM analysis
--mode MODEautoForce: screen, talking-head, auto

Requirements

Core (always needed)

  • FFmpeg with libass support
  • WhisperX (pip install whisperx in venv, 6GB VRAM)
  • PySceneDetect (pip install scenedetect)

VLM Analysis (screen mode)

  • GOOGLE_API_KEY environment variable
  • pip install google-genai in venv
  • Cost: ~$0.01-0.03 per video (~40 frames via Gemini 2.5 Flash)

Face Tracking (talking-head mode)

  • MediaPipe (pip install mediapipe)
  • OpenCV (pip install opencv-python)
  • CPU-only, no VRAM required

Platform Duration Limits

PlatformMax DurationRecommendedAspect
TikTok10 min30-60s9:16
Instagram Reels90s30-60s9:16
YouTube Shorts60s30-55s9:16

Safety Rules

  1. Always run bash scripts/preflight.sh "$INPUT" "$OUTPUT" before writes
  2. Never overwrite source video — all shorts are new files
  3. Confirm before processing videos longer than 60 minutes
  4. Clean up temp files after pipeline completes
  5. Report estimated processing time before starting

Reference

Load references/shorts-pipeline.md for detailed algorithms, scoring formulas, and tuning parameters.

Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.

<a href="https://skillzs.dev/skills/agricidaniel/claude-video/claude-video-shorts">View claude-video-shorts on skillZs</a>