raw-video-processing

SKILL.md

Skill: Raw Video Processing

Post-process raw screen recordings to improve pacing — remove silent segments, then speed up the result.

Prerequisite: FFmpeg and uv must be installed.

When to Use

The user has recorded a screencast and wants to clean it up before publishing. Typical issues in raw recordings:

Long pauses / dead air while thinking or waiting for loading

Keyboard typing sounds and other low-level background noise that should be treated as silence

Overall pacing feels slow and could benefit from a slight speed boost

Default Workflow
When the user provides a raw video file, run both scripts in sequence by default:

Step 1: Remove Silent Segments

uv run --python 3.12 /path/to/skills/raw-video-processing/scripts/remove_silence.py <input.mp4> -t="-20dB" -d 0.5

This detects and cuts out silent portions (including keyboard sounds), producing <input>_nosilence.mp4.

Always pass these parameters (tuned for screen recordings with keyboard noise):

-t="-20dB" — aggressive threshold that filters out keyboard typing and background noise (use = syntax to avoid argparse treating negative values as flags)

-d 0.5 — remove short silences too (0.5s minimum)

-p 0.2 — seconds of breathing room kept around speech boundaries (default, usually no need to pass)

The script prints a detailed summary: number of silent segments found, total silence removed, and all kept segments with timestamps. Review this output to confirm the result looks reasonable.

Step 2: Speed Up the Video

uv run --python 3.12 /path/to/skills/raw-video-processing/scripts/speed_video.py <input>_nosilence.mp4

This applies a speed multiplier to the silence-removed video, producing <input>_nosilence_1.2x.mp4.

Default parameters:

--speed 1.2 — 1.2x playback speed (a subtle boost that doesn't feel rushed)

Script Options

remove_silence.py

Flag
Default
Description

-o, --output
<input>_nosilence.mp4
Custom output path

-t, --threshold
-30dB
Silence threshold in dB (higher = more aggressive). Always use -20dB for screencasts — pass as -t="-20dB" to avoid argparse issues with negative values

-d, --duration
0.8
Minimum silence duration in seconds to remove. Use 0.5 for screencasts

-p, --padding
0.2
Padding kept around non-silent segments

--dry-run
off
Only print detected segments, don't export

speed_video.py

Flag
Default
Description

-o, --output
<input>_<speed>x.mp4
Custom output path

-s, --speed
1.2
Playback speed multiplier

Custom Scenarios

Only remove silence — run just Step 1.

Only speed up — run just Step 2 directly on the input file.

Conservative cleanup — use -t="-30dB" -d 0.8 if the default is cutting too much speech.

Extra aggressive cleanup — use -t="-15dB" -d 0.3 and --speed 1.5 for maximum compression.

Preview before committing — use --dry-run on remove_silence.py to see what would be cut without creating a file.

Custom output name — use -o on either script to control the output path.

Important Notes

Always run remove_silence before speed_video. Silence detection works on the original audio; speeding up first would alter the audio characteristics and make silence detection less accurate.

For long videos (>30 min), the silence removal step may take a few minutes as it processes each segment individually.

Both scripts preserve video quality — remove_silence uses stream copy (no re-encoding), while speed_video re-encodes with FFmpeg defaults.