Video To Context

Video-to-Context: Extract full context from videos

Turn screen recordings into readable context (transcript + key frames) so you can understand what happened without watching the video.

When to use

• User shares a screen recording from a Jira card, Slack, or bug report
• You need to understand both what's shown on screen AND what's being said
• Video is in .mov, .mp4, .webm, or similar format

Requirements

• ffmpeg (frame extraction)
• whisper-ctranslate2 with CUDA (speech-to-text) -- see setup below
• NVIDIA GPU with CUDA drivers (CPU fallback possible but slow)
• uv package manager

Setup (one-time)

Run this to install the transcription stack:

#!/bin/bash
set -euo pipefail

VENV="$HOME/.local/share/video-tools"

# Create isolated venv
uv venv "$VENV" --python 3.12

# Install whisper + CUDA libs
VIRTUAL_ENV="$VENV" uv pip install whisper-ctranslate2 nvidia-cublas-cu12 nvidia-cudnn-cu12

# Create wrapper that sets CUDA library paths
mkdir -p "$HOME/.local/bin"
cat > "$HOME/.local/bin/whisper-gpu" /dev/null && echo "Ready." || echo "Setup failed."

Ensure ~/.local/bin is in PATH, or create a symlink:

sudo ln -sf ~/.local/bin/whisper-gpu /usr/local/bin/whisper-gpu

Extraction process

Step 1: Transcribe audio

whisper-gpu "$VIDEO" \
  --model large-v3 \
  --language en \
  --device cuda \
  --compute_type int8_float16 \
  --output_format all \
  --output_dir "$OUTDIR" \
  --word_timestamps True \
  --vad_filter True \
  --batched True \
  --batch_size 8

This produces .txt (plain transcript), .srt (timestamped), and .json (word-level timestamps).

For non-English videos, change --language or remove it for auto-detection.

Step 2: Extract scene-change frames

ffmpeg -y -i "$VIDEO" \
  -vf "select='gt(scene,0.3)',scale=1280:-1,drawtext=text='%{pts\:hms}':fontsize=24:fontcolor=white:box=1:boxcolor=black@0.7:x=w-tw-10:y=h-th-10" \
  -vsync vfr -q:v 3 \
  "$OUTDIR/frames/scene_%04d.jpg"

The scene>0.3 threshold catches major visual changes (page navigations, modal opens, screen switches). Timestamps are burned into each frame's bottom-right corner.

Step 3: Extract interval frames (every 5s)

ffmpeg -y -i "$VIDEO" \
  -vf "fps=1/5,scale=1280:-1,drawtext=text='%{pts\:hms}':fontsize=24:fontcolor=white:box=1:boxcolor=black@0.7:x=w-tw-10:y=h-th-10" \
  -q:v 3 \
  "$OUTDIR/frames/interval_%04d.jpg"

Interval frames fill gaps where no scene change was detected but something relevant might still be happening.

Step 4: Read the output

1. Read the .txt transcript first for the full narrative
2. Read specific frames (.jpg) when you need to see what was on screen at a given moment
3. Cross-reference .srt timestamps with frame timestamps to match speech to visuals

Gotcha: macOS filenames with Unicode spaces

macOS screen recordings use U+202F (NARROW NO-BREAK SPACE) instead of regular spaces in filenames. If your file isn't found, try:

# Rename to regular spaces
find "$DIR" -maxdepth 1 -name "Screen*Recording*" -exec bash -c '
  NEW=$(echo "$1" | sed "s/\xe2\x80\xaf/ /g")
  [ "$1" != "$NEW" ] && mv "$1" "$NEW"
' _ {} \;

Or add a glob fallback in your script:

if [ ! -f "$VIDEO" ]; then
  RESOLVED=$(find "$(dirname "$VIDEO")" -maxdepth 1 \
    -name "$(basename "$VIDEO" | sed 's/ /*/g')" -print -quit 2>/dev/null)
  [ -n "$RESOLVED" ] && VIDEO="$RESOLVED"
fi

Output structure

video_context/
  context.md          # Transcript + frame list (read this first)
  recording.txt       # Plain transcript
  recording.srt       # Timestamped transcript
  recording.json      # Word-level timestamps
  frames/
    scene_0001.jpg    # Scene-change frames (timestamps burned in)
    scene_0002.jpg
    interval_0001.jpg # Every-5s frames (timestamps burned in)
    interval_0002.jpg

Quick reference

| What | Command/Flag | Notes | |------|-------------|-------| | Whisper model | large-v3 | Best accuracy. Use medium for speed. | | Quantization | int8_float16 | Good for 8GB VRAM GPUs (e.g. RTX 4060) | | Scene threshold | gt(scene,0.3) | Lower = more frames. 0.3 is a good default. | | Frame interval | fps=1/5 | Every 5 seconds. Change denominator as needed. | | CPU fallback | --device cpu | Works but 10-20x slower than CUDA | | Output formats | --output_format all | Gets txt, srt, json, vtt, tsv at once | | VAD filter | --vad_filter True | Skips silence, improves accuracy | | Batched mode | --batched True | Faster on GPU, uses more VRAM |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: sennaBruno
• Source: sennaBruno/claude-skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.