Communitygithub.com

HUANGCHIHHUNGLeo/claude-real-video

Watches videos for Leo. When Leo shares a video link (YouTube/IG/TikTok) or local video file, and wants a summary, analysis, breakdown, or benchmarking — since Claude cannot directly ingest video, this tool extracts key frames, transcripts, camera movement rhythm, and audio/tone/gesture timelines for reading.

What is claude-real-video?

This skill, named claude-real-video / crv Pro, is designed for developers and AI agent users who need to analyze video content that Claude cannot natively process. It extracts key frames, transcripts, camera movement patterns, and perception timelines (audio events, tone, emotion, gestures, scene labels) from YouTube, Instagram, TikTok, or local video files. The skill supports multiple analysis modes: 'watch' for summaries and fact-checking, 'creator' for deconstructing why a video went viral or how it was made, and 'full' for comprehensive analysis. It outputs a structured MANIFEST.txt with timestamps and confidence scores, a grid of frames, and a perception.json for low-confidence events. Critical usage rules include reading the full transcript before reporting to avoid omissions, cross-referencing perception labels with frames and transcript for verification, and using motion rhythm data for cloning or recreating video styles. It runs entirely locally, requires no uploads, and can save results to a knowledge base.

Works with✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/HUANGCHIHHUNGLeo/claude-real-video/tree/HEAD/personal-skill-backup

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

claude-real-video / crv Pro — Watches Videos for Leo

Leo's local installation is the Pro version 0.5.0 (~/Projects/crv-pro/.venv/bin/crv-pro, or simply crv-pro if the PATH includes it).

When to Use

When Leo shares a video (URL or file path) and asks about its content, wants a summary, a clone/reference breakdown, or questions like "How did they say it?", "How was it shot?", or "Why did the audience buy it?"

How to Choose the Right Command (0.5.0)

What Leo SaysCommand
"What is this about?" / "Summarize it" / "Is what they said true?"crv-pro "<src>" -o out --mode watch
"How was it made?" / "Deconstruct it for me" / "Why did it go viral?" / "Can I learn from this?"crv-pro "<src>" -o out --mode creator
"Full analysis" or unclear intentcrv-pro "<src>" -o out --mode full

If Leo has a specific question, add --why "<their original phrasing>".

  • --senses = Full sensory analysis: sound events, tone curve, emotion, gestures/expressions, scene labels (first run downloads ~2.7GB model; after that, an 86-second video takes about 40 seconds)
  • For a single sense: --prosody / --audio-events / --emotion / --gesture / --scene-labels
  • --lens content|creator|both (default both) determines which analysis direction the MANIFEST instructs the AI to follow
  • For long videos, add --max-frames 60; for videos without speech, use --no-transcribe
  • For platforms that require login (IG/TikTok): first convert the cookie JSON to Netscape format, then use --cookies (tested 2026-07-05 can directly fetch IG reels — this is the proper way):
    python3 -c "import json; ck=json.load(open('/Users/leo/Projects/auto-post/browse-cookies/kanisleo328-threads.json')); cs=ck if isinstance(ck,list) else ck['cookies']; print('# Netscape HTTP Cookie File'); [print('\t'.join([c['domain'],'TRUE' if c['domain'].startswith('.') else 'FALSE',c.get('path','/'),'TRUE',str(int(c.get('expirationDate',2e9))),c['name'],c['value']])) for c in cs if 'instagram' in c.get('domain','')]" > /tmp/ig-cookies.txt
    crv-pro "<reel url>" --cookies /tmp/ig-cookies.txt --mode full
    
    If that also fails, fall back to [[reference_ig_post_full_read]]'s internal IG API manual download.

Reading Order for Output

  1. MANIFEST.txt — Starts with a lens instruction block to follow; includes frame list, motion block, perception timeline (sound/tone/emotion/gesture/scene events, each with timestamps and confidence scores), and transcript
  2. grids/ — Continuous nine-grid frames; for details, open frames/*.jpg (scale down to within 1500px before reading)
  3. perception.json — Complete perception data (low-confidence events hidden by thresholds are here)

⚠️ Cardinal Rules (Set by Leo QA on 2026-07-05)

  • The full transcript must be read from beginning to end before producing a report (2026-07-16 Leo caught sampling-based reading leading to incorrect attributions and missing the three most interesting details of the entire video): read the entire transcript.txt before writing the content breakdown; sampling may only be used to locate timestamps, never to substitute understanding.

  • Perception labels are clues, not conclusions: cross-verify with frames and transcript before citing (as taught at the start of MANIFEST); the emotion model may misread "normal forceful tone" as angry — non-neutral emotion with confidence <0.75 is already hidden; even exposed labels require verification.

  • Before cloning/building a replica, always check the motion rhythm table (cuts/min, median shot length, camera movement distribution) — don't rely on intuition (see [[reference_reel_clone_pipeline]] for details)

  • Cite specific timestamps and data in answers, not generalities.

Notes

  • Runs entirely locally; the tool does not upload anything; use --kb <directory> to save into a knowledge base
  • Temporary downloaded videos and extracted frames should be cleaned up after reading (under /tmp)

Individual skills in this repo

This repo contains 2 individual skills — each has its own dedicated page.

Related Skills