openclaw-video-editor
A focused, honest video editing skill for OpenClaw. It wraps ffmpeg, ffprobe, and small Python helpers around real, working workflows. No misleading claims, no external API calls, no hidden network behavior.
What this skill does
This skill helps the agent run common video edits using only locally installed tools (ffmpeg, ffprobe, python3). It ships:
- A subtitle generator (
scripts/generate_srt.py) that converts Whisper / Deepgram / AssemblyAI / generic word-timing JSON into .srt, .vtt, or .ass.
- A scene-detected highlight reel builder (
scripts/highlight_reel.py) that picks key moments using ffmpeg scene-change scores.
- A color grading helper (
scripts/apply_lut.py) that applies .cube LUT files or named filter presets.
- A dependency check (
scripts/check_deps.sh) so the agent can confirm ffmpeg/ffprobe/python3 exist before running any command.
- A small library of correct, tested ffmpeg one-liners below.
This skill does not perform AI-based subject masking, voice cloning, or generative video. If you want those, use a dedicated tool — this skill is intentionally limited to honest local processing.
What this skill does not do
To set expectations clearly:
- It does not include AI background removal. The blur workflow below is a
boxblur filter, not subject segmentation.
- It does not include a transcription model. It only formats word-level timings into subtitle files. Pair it with a transcription tool to get the JSON input.
- It does not call any external service. All commands run locally on the machine where
ffmpeg is installed.
Required binaries
Before running any workflow, the agent should confirm dependencies:
bash scripts/check_deps.sh
This returns a non-zero exit code if ffmpeg, ffprobe, or python3 is missing. The agent should surface a clear error to the user instead of attempting commands that will fail.
Workflows
1. Generate subtitles from a transcript
python3 scripts/generate_srt.py transcript.json subtitles.srt
python3 scripts/generate_srt.py transcript.json subtitles.vtt
python3 scripts/generate_srt.py transcript.json subtitles.ass --font Helvetica --fontsize 28
Tunable flags: --max-chars, --max-words, --max-duration. Useful for non-English languages where line lengths differ.
2. Burn subtitles into a video
ffmpeg -i input.mp4 -vf "subtitles=subtitles.srt:force_style='Alignment=2,MarginV=30,Outline=2,Shadow=1'" -c:a copy output.mp4
Alignment=2 is bottom-center. Use Alignment=8 for top-center when the bottom of the frame has important action.
3. Background blur (not segmentation)
A real shallow-depth-of-field look without green screen requires AI segmentation, which this skill does not include. What it does provide is honest full-frame blur, useful as a privacy filter:
ffmpeg -i input.mp4 -vf "boxblur=20:1" -c:a copy blurred.mp4
If you already have a binary alpha matte (mask.mp4, where the subject is white and the background is black, frame-aligned to input.mp4), you can composite a blurred background behind the original subject:
ffmpeg -i input.mp4 -i mask.mp4 \
-filter_complex "[0:v]boxblur=20:1[bg];[0:v][1:v]alphamerge[fg];[bg][fg]overlay=format=auto" \
-c:a copy composited.mp4
This only works if a per-frame alpha matte already exists. Producing the matte itself is out of scope.
4. Color grading
Two options.
Honest filter presets (no LUT file required):
# Warmer, more saturated look
ffmpeg -i input.mp4 -vf "eq=contrast=1.1:saturation=1.2,colorbalance=rs=0.1:gs=0.0:bs=-0.05" -c:a copy graded.mp4
# Desaturated black & white
ffmpeg -i input.mp4 -vf "hue=s=0,eq=contrast=1.2:brightness=-0.02" -c:a copy bw.mp4
Real LUT (when you have a .cube file):
python3 scripts/apply_lut.py input.mp4 lut.cube graded.mp4
# or directly:
ffmpeg -i input.mp4 -vf "lut3d=lut.cube" -c:a copy graded.mp4
.cube LUT files are the only way to get accurate cinematic looks. The filter-based "presets" above are approximations.
5. Audio normalization (broadcast loudness)
Two-pass loudnorm targeting EBU R128 (-23 LUFS):
# Pass 1: measure
ffmpeg -i input.mp4 -af "loudnorm=I=-23:LRA=7:tp=-2:print_format=json" -f null - 2> loudnorm.log
# Pass 2: apply with measured values
# Read the input_i / input_tp / input_lra / input_thresh / target_offset values from loudnorm.log
ffmpeg -i input.mp4 \
-af "loudnorm=I=-23:LRA=7:tp=-2:measured_I=<input_i>:measured_TP=<input_tp>:measured_LRA=<input_lra>:measured_thresh=<input_thresh>:offset=<target_offset>:linear=true" \
-c:v copy normalized.mp4
For online platforms, -14 LUFS is the common spec instead of -23 LUFS.
6. Highlight reel via scene detection
python3 scripts/highlight_reel.py input.mp4 highlight.mp4 --duration 30 --threshold 0.4
Uses ffprobe/ffmpeg scene-change scores to pick clip boundaries, then assembles a short reel up to the requested duration.
7. Watermark
# Bottom-right text watermark
ffmpeg -i input.mp4 -vf "drawtext=text='@yourhandle':x=w-tw-20:y=h-th-20:fontsize=24:fontcolor=white@0.7:box=1:boxcolor=black@0.4:boxborderw=8" -c:a copy watermarked.mp4
# Image overlay (logo)
ffmpeg -i input.mp4 -i logo.png -filter_complex "[0:v][1:v]overlay=W-w-20:20" -c:a copy watermarked.mp4
8. Format conversions and resizes
# 1080p H.264 with reasonable quality
ffmpeg -i input.mp4 -vf "scale=-2:1080" -c:v libx264 -preset medium -crf 20 -c:a aac -b:a 160k output_1080p.mp4
# Vertical 9:16 crop for shorts
ffmpeg -i input.mp4 -vf "crop=ih*9/16:ih,scale=1080:1920" -c:a copy vertical.mp4
Safety notes
- All workflows are local. No part of this skill calls a remote API or sends data anywhere.
ffmpeg filtergraphs in this skill never use subprocess from inside Python with shell-evaluated user input. The Python helpers shell-quote arguments via shlex and reject paths containing shell metacharacters.- The skill never modifies system configuration, environment variables, or other plugins. It only reads input files and writes output files at paths the user provides.
Limitations
- The skill assumes inputs are valid video/audio files reachable on the local filesystem.
- Two-pass loudness normalization is described but not automated end-to-end; the agent must read the measurement output and pass values into the second pass.
- The blur compositing workflow only works when a per-frame matte already exists. Generating the matte requires a separate segmentation tool.
License
MIT. See LICENSE for the full text.
