claude-video — CLI-Native Video Editing for Claude Code

Quick Reference

| Command | What it does | |---------|-------------| | /video | Interactive mode — describe what you want in plain English | | /video setup | Install all dependencies (faster-whisper, auto-editor, scenedetect, yt-dlp, mediainfo) | | /video edit | Trim, cut, split, merge, speed, crop, overlay, stabilize, transitions | | /video transcode | Convert codecs, compress, GPU-accelerated encoding, two-pass | | /video audio | Normalize loudness, reduce noise, mix, extract, remove silence | | /video caption | Transcribe speech → animated word-by-word subtitles (Whisper + ASS) | | /video analyze | Inspect with ffprobe, measure quality (VMAF/SSIM/PSNR), detect scenes | | /video export <platform> | One-command export: youtube, tiktok, instagram, linkedin, web, gif, podcast | | /video download <url> | Download video via yt-dlp with format selection | | /video create | Programmatic video creation via Remotion (React-based motion graphics) | | /video shorts | Longform → shortform pipeline: transcribe, score, crop 9:16, caption, export | | /video image | AI image generation (Gemini 3 Pro Image, FLUX.2 local, SD 3.5, OpenAI, rembg) | | /video generate | AI video generation (Veo 3.1, Runway Gen-4, local SVD) | | /video screenshot | Web capture via Playwright (screenshots, recordings, Ken Burns) | | /video enhance-audio | AI audio (Demucs separation, DeepFilter denoise, TTS, diarize, AudioSR) | | /video enhance | AI video (Real-ESRGAN upscale, RIFE interpolation, CodeFormer, rembg) | | /video promo | Stock footage + Remotion promo videos with contrast-aware text overlays | | /video setup --ai | Install AI dependencies (PyTorch, WhisperX, FLUX.2, Real-ESRGAN, etc.) |

Orchestration Logic

Command Routing

When the user provides a specific command, load the matching sub-skill:

/video edit or intent is trim/cut/split/merge/speed/crop/overlay/stabilize/transition → Read skills/claude-video-edit/SKILL.md
/video transcode or intent is convert/compress/encode/codec/format change → Read skills/claude-video-transcode/SKILL.md
/video audio or intent is audio normalize/noise/mix/extract/silence/volume/LUFS → Read skills/claude-video-audio/SKILL.md
/video caption or intent is transcribe/subtitle/caption/whisper/SRT/ASS → Read skills/claude-video-caption/SKILL.md
/video analyze or intent is inspect/info/quality/VMAF/metadata/scene detect/ffprobe → Read skills/claude-video-analyze/SKILL.md
/video export or intent is export for youtube/tiktok/instagram/linkedin/web/gif/podcast → Read skills/claude-video-export/SKILL.md
/video download or intent is download video from URL → Read skills/claude-video-download/SKILL.md
/video create or intent is create/generate programmatic video/motion graphics/title card → Read skills/claude-video-create/SKILL.md
/video shorts or intent is shortform/clips/TikTok shorts/reels from longform → Read skills/claude-video-shorts/SKILL.md
/video image or intent is generate image/AI image/FLUX/stable diffusion/remove background → Read skills/claude-video-image/SKILL.md
/video generate or intent is generate video/AI video/Veo/Runway/text-to-video → Read skills/claude-video-generate/SKILL.md
/video screenshot or intent is screenshot/web capture/screen recording/Ken Burns → Read skills/claude-video-screenshot/SKILL.md
/video enhance-audio or intent is separate vocals/denoise AI/diarize/TTS/voice clone/upsample audio → Read skills/claude-video-enhance-audio/SKILL.md
/video enhance or intent is upscale/super resolution/interpolate/face restore/remove background video → Read skills/claude-video-enhance/SKILL.md
/video promo or intent is promo video/stock footage/marketing video/stock clips with text → Read skills/claude-video-promo/SKILL.md
/video setup → Run bash scripts/setup.sh
/video setup --ai → Run bash scripts/setup.sh --ai

Interactive Mode

When user says /video without arguments or describes a task in natural language:

Run bash scripts/check_deps.sh to verify tools are available
Run bash scripts/detect_gpu.sh to know encoding capabilities
Identify intent from the user's description
Route to the appropriate sub-skill
If intent is ambiguous, ask the user to clarify

Multi-Step Pipelines

For complex requests that span multiple sub-skills (e.g., "download this YouTube video, remove silence, add captions, export for TikTok"), execute sub-skills sequentially:

Download → edit → caption → export
Pass output of each step as input to the next
Use temp files in /tmp/claude-video/ for intermediate outputs
Clean up temp files after final output is produced

Safety Rules — MANDATORY

Always use -n flag (no-overwrite) unless user explicitly requests overwrite
Never let output path equal input path — run bash scripts/preflight.sh "$INPUT" "$OUTPUT" before any write
Never delete source files — all operations produce new files
Run preflight before every write operation: bash scripts/preflight.sh <input> <output>
Confirm before: batch operations (>3 files), downloads, operations estimated >5 minutes
Auto-execute without confirmation: ffprobe analysis, thumbnail extraction, GPU detection, metadata reading
Temp file cleanup: use trap 'rm -f "$TEMP"' EXIT INT TERM pattern in all scripts
Disk space check: run bash scripts/estimate_size.sh before large encodes

AI Sub-Skill Routing Guide

"separate vocals" / "isolate dialogue" → claude-video-enhance-audio (Demucs, AI model)
"reduce noise" / "denoise" → claude-video-audio (FFmpeg afftdn) for quick fix, claude-video-enhance-audio (DeepFilterNet3) for AI quality
"remove background" from image → claude-video-image (rembg), from video → claude-video-enhance (rembg batch + alpha assembly)
"upscale" / "super resolution" → claude-video-enhance (Real-ESRGAN)
"slow motion" (basic) → claude-video-edit (setpts), (AI frame interpolation) → claude-video-enhance (RIFE)
"generate image/video" → claude-video-image / claude-video-generate
"shorts" / "clips from long video" → claude-video-shorts

Python Virtual Environment

AI scripts require an isolated venv with PyTorch nightly cu128 (for RTX 5070 Ti Blackwell sm_120):

# Install the AI venv (one-time)
bash scripts/setup.sh --ai

# Venv location
~/.video-skill/bin/python3
~/.video-skill/bin/activate

All Python scripts in scripts/ that use AI models (segment_scorer.py, face_tracker.py, image_generate.py, video_generate.py, audio_enhance.py, video_enhance.py, web_capture.py) require activation: source ~/.video-skill/bin/activate

VRAM Management (16GB Budget — RTX 5070 Ti)

Before loading any GPU model, check free VRAM:

nvidia-smi --query-gpu=memory.free --format=csv,noheader,nounits

| Tier | VRAM | Models | Rule | |------|------|--------|------| | Exclusive (>10GB) | 12-13GB | FLUX.2 klein, Bark TTS | Must be ONLY GPU model loaded | | Heavy (5-10GB) | 5-8GB | Demucs, WhisperX, AudioSR | One heavy at a time | | Light (<5GB) | 1-4GB | rembg, CodeFormer, RIFE, pyannote | Can coexist if total <12GB | | NVENC | ~500MB | Hardware encoder | Always available | | CPU only | 0 | MediaPipe, PySceneDetect, Playwright, librosa | No VRAM needed |

NEVER run two heavy/exclusive models simultaneously. Load → process → unload → next model.

Encoding Defaults

GPU-First Strategy

Run bash scripts/detect_gpu.sh at start of any encoding task
If NVENC available: use h264_nvenc/hevc_nvenc/av1_nvenc with -preset p5 -tune hq
If no GPU: fallback to libx264/libx265/libsvtav1 with -preset medium
Always add -movflags +faststart for MP4 output

Codec Selection Priority

Stream copy (-c copy): when no re-encoding needed (trim, r

claude-video

Como adicionar

Cole no README do seu repo

Skills relacionadas

frontend-design

mcp-builder

webapp-testing

web-artifacts-builder

Receba novas skills de Design e Frontend toda segunda