FFmpeg Video Editor

Edit video, audio, and images from the terminal using ffmpeg and the fftools.py helper script.

CLI Tool

All operations go through the helper script. Find it relative to where you cloned the repo, or at:

SCRIPT="~/.claude/skills/ffmpeg/scripts/fftools.py"

All output is JSON.

Commands Reference

# INFO — Get file details (duration, resolution, codecs, size)
python3 $SCRIPT info input.mp4

# CUT — Extract a segment
python3 $SCRIPT cut input.mp4 output.mp4 --start 00:01:30 --end 00:02:45
python3 $SCRIPT cut input.mp4 output.mp4 --start 00:01:30 --duration 30

# MERGE — Concatenate multiple videos
python3 $SCRIPT merge video1.mp4 video2.mp4 video3.mp4 -o merged.mp4
python3 $SCRIPT merge video1.mp4 video2.mp4 -o merged.mp4 --reencode  # different codecs

# EXTRACT AUDIO — Rip audio track
python3 $SCRIPT extract-audio input.mp4 output.mp3
python3 $SCRIPT extract-audio input.mp4 output.opus --bitrate 128k

# SUBTITLES — Burn subtitles into video
python3 $SCRIPT subtitles input.mp4 output.mp4 --srt subs.srt

# RESIZE — Change resolution
python3 $SCRIPT resize input.mp4 output.mp4 --width 1280 --height 720
python3 $SCRIPT resize input.mp4 output.mp4 --width 1920 --height -1  # auto height

# GIF — Video to animated GIF (two-pass, high quality)
python3 $SCRIPT gif input.mp4 output.gif --fps 15 --width 480
python3 $SCRIPT gif input.mp4 output.gif --start 00:00:05 --duration 3

# THUMBNAIL — Extract single frame
python3 $SCRIPT thumbnail input.mp4 thumb.jpg --time 00:00:30

# THUMBNAILS — Extract N evenly-spaced frames
python3 $SCRIPT thumbnails input.mp4 --count 10 --output-dir ./thumbs/

# COMPRESS — Reduce file size
python3 $SCRIPT compress input.mp4 output.mp4 --crf 28 --preset medium
python3 $SCRIPT compress input.mp4 small.mp4 --crf 32 --preset fast  # aggressive

# SPEED — Change playback speed
python3 $SCRIPT speed input.mp4 fast.mp4 --factor 2.0    # 2x faster
python3 $SCRIPT speed input.mp4 slow.mp4 --factor 0.5    # half speed (slow-mo)

# WATERMARK — Add image overlay
python3 $SCRIPT watermark input.mp4 output.mp4 --watermark logo.png --position bottom-right

# TEXT — Burn text overlay
python3 $SCRIPT text input.mp4 output.mp4 --text "Chapter 1" --position center --size 72 --border
python3 $SCRIPT text input.mp4 output.mp4 --text "Sale!" --position top-left --start 5 --end 15

# CONVERT — Change format
python3 $SCRIPT convert input.avi output.mp4
python3 $SCRIPT convert input.mp4 output.webm --video-codec libvpx-vp9

# FRAMES — Extract frames as images
python3 $SCRIPT frames input.mp4 --output-dir ./frames/
python3 $SCRIPT frames input.mp4 --output-dir ./frames/ --every 30  # every 30th frame

# AUDIO REPLACE — Swap audio track
python3 $SCRIPT audio-replace input.mp4 output.mp4 --audio music.mp3

# VOLUME — Adjust audio level
python3 $SCRIPT volume input.mp4 output.mp4 --level 2.0      # double volume
python3 $SCRIPT volume input.mp4 output.mp4 --level 0.5      # halve volume
python3 $SCRIPT volume input.mp4 output.mp4 --level 10dB     # boost 10dB

# FADE — Add fade in/out effects
python3 $SCRIPT fade input.mp4 output.mp4 --fade-in 2 --fade-out 3

# STABILIZE — Fix shaky video (requires vidstab filter)
python3 $SCRIPT stabilize input.mp4 output.mp4 --shakiness 7 --smoothing 15

# TRANSCRIBE — Speech-to-text WITH speaker labels (offline; see Transcription section)
python3 $SCRIPT transcribe meeting.mp4                       # → meeting.json/.srt/.txt
python3 $SCRIPT transcribe meeting.mp4 --speakers 3          # tell it the speaker count
python3 $SCRIPT transcribe meeting.mp4 --no-diarize          # transcript only, no speakers

Transcription & Analysis — understand what's said in a video

Use the transcribe command to turn a meeting recording, interview, or any speech video/audio into a speaker-labelled, timestamped transcript. This is the tool for questions like "what was discussed?", "who said what?", "summarise this meeting", or "find where they talked about X".

It runs 100% locally and offline — audio never leaves the machine, no cloud API. Under the hood: WhisperX (large-v3) for transcription + word-level timestamps, and sherpa-onnx for speaker diarization (token-free, open ONNX models).

One-time setup (needs network once)

Run the /setup-transcribe slash command, or directly:

bash ~/.claude/skills/ffmpeg/scripts/install-transcribe.sh

By default this fetches a prebuilt bundle from the GitHub release (fast: no model gating, no torch build) and installs it fully offline; if that's unavailable (release missing, or a platform whose wheels don't match) it automatically falls back to a from-scratch build. It creates an isolated Python 3.12 venv (via uv) and all models; afterwards transcribe needs no network. If the command reports the venv is missing, run this.

Force a specific mode if you want:

bash install-transcribe.sh --release   # only the prebuilt release bundle
bash install-transcribe.sh --build     # only a from-scratch build (HF + PyPI)
HF_TOKEN=hf_xxx bash install-transcribe.sh --build   # also enable --diarizer pyannote

Air-gapped / no HuggingFace access? Build a self-contained bundle once on a connected machine, host it anywhere, and install from it with no HF/PyPI access:

# On a machine where install-transcribe.sh already ran:
bash ~/.claude/skills/ffmpeg/scripts/make-bundle.sh ffmpeg-transcribe-bundle.tar.gz

The bundle packs all code, models, model caches, and Python wheels. If it exceeds 2GB it is auto-split into <name>.part-aa, <name>.part-ab, … (each <2GB) to fit hosting/upload limits. Then, on the target machine (needs only uv, ffmpeg, and python3.12):

# Single-file bundle:
bash install-transcribe.sh --bundle https://your.server/ffmpeg-transcribe-bundle.tar.gz

# Split bundle — pass the parts in order (they are concatenated automatically):
bash install-transcribe.sh --bundle \
  https://your.server/ffmpeg-transcribe-bundle.tar.gz.part-aa \
  https://your.server/ffmpeg-transcribe-bundle.tar.gz.part-ab

(To reassemble manually instead: cat ffmpeg-transcribe-bundle.tar.gz.part-* > ffmpeg-transcribe-bundle.tar.gz.)

Usage

# Full transcript + speaker labels (writes meeting.json, meeting.srt, meeting.txt)
python3 $SCRIPT transcribe meeting.mp4

# Options
python3 $SCRIPT transcribe interview.mp4 --speakers 2          # known speaker count → better accuracy
python3 $SCRIPT transcribe call.m4a --language en             # force language (skip auto-detect)
python3 $SCRIPT transcribe lecture.mp4 --format txt           # only write .txt
python3 $SCRIPT transcribe long.mp4 --model medium            # faster, less accurate
python3 $SCRIPT transcribe q4.mp4 --diarizer pyannote         # higher-quality diarizer (needs HF token)
python3 $SCRIPT transcribe notes.mp4 --no-diarize            # transcript only, no speaker separation
python3 $SCRIPT transcribe raw.wav -o ./out/session          # custom output base path

Output formats

.txt — human-readable, grouped by speaker turn. Read this to understand the content: [00:03:12] SPEAKER_01: ...
.srt — subtitles with speaker tags, for burning back in (subtitles command)
.json — full structured data: every segment + word with start/end times and speaker, for programmatic analysis (search, timelines, quoting with timestamps)

Meeting-analysis workflow

# 1. Transcribe with speaker labels
python3 $SCRIPT transcribe meeting.mp4

# 2. Read the transcript to analyse content
#    (use the Read tool on meeting.txt, or parse meeting.json for timestamps)

# 3. From here you can summarise, extract action items, find who-said-what,
#    or jump to a moment: pair a JSON timestamp with `cut`/`thumbnail`.
python3 $SCRIPT cut meeting.mp4 clip.mp4 --start 00:12:30 --end 00:14:00

Notes:

Transcription runs on CPU (CTranslate2 has no GPU path on Apple Silicon), so large-v3 is roughly 1–2× realtime. A 1-hour meeting ≈ 30–60 min to process — fine to run in the background. Use --model medium or distil-large-v3 for speed.
Passing --speakers N (when you know the count) noticeably improves diarization.

Standard Workflow

Step 1 — Analyze the input

Always start by getting file info:

python3 $SCRIPT info input.mp4

This tells you duration, resolution, codecs, file size. Use this to make smart decisions about output settings.

Step 2 — Perform the operation

Use the appropriate command. For complex edits, chain multiple operations — use the output of one as input to the next.

Step 3 — Verify the result

After any operation, probe the output to confirm it worked:

python3 $SCRIPT info output.mp4

Step 4 — Show the result

Use the Read tool to display images (thumbnails, frames). For videos, tell the user the output path and file size.

Raw ffmpeg — For advanced operations

When the helper script doesn't cover your use case, use ffmpeg directly:

# Picture-in-picture
ffmpeg -y -i main.mp4 -i overlay.mp4 \
  -filter_complex "[1:v]scale=320:-1[pip];[0:v][pip]overlay=W-w-10:H-h-10" \
  -c:v libx264 -c:a copy output.mp4

# Side-by-side comparison
ffmpeg -y -i left.mp4 -i right.mp4 \
  -filter_complex "[0:v]pad=iw*2:ih[bg];[bg][1:v]overlay=W/2:0" \
  -c:v libx264 output.mp4

# Grid/mosaic (2x2)
ffmpeg -y -i a.mp4 -i b.mp4 -i c.mp4 -i d.mp4 \
  -filter_complex "[0:v]scale=640:360[v0];[1:v]scale=640:360[v1];
    [2:v]scale=640:360[v2];[3:v]scale=640:360[v3];
    [v0][v1]hstack[top];[v2][v3]hstack[bot];[top][bot]vstack" \
  -c:v libx264 grid.mp4

# Reverse video
ffmpeg -y -i input.mp4 -vf reverse -af areverse reversed.mp4

# Extract specific audio channel
ffmpeg -y -i input.mp4 -af "pan=mono|c0=FL" left_channel.wav

# Noise reduction
ffmpeg -y -i input.mp4 -af "afftdn=nf=-25" denoised.mp4

# Color correction (brightness, contrast, saturation)
ffmpeg -y -i input.mp4 -vf "eq=brightness=0.1:contrast=1.2:saturation=1.3" corrected.mp4

# Chroma key (green screen removal)
ffmpeg -y -i greenscreen.mp4 -i background.mp4 \
  -filter_complex "[0:v]chromakey=0x00FF00:0.3:0.1[fg];[1:v][fg]overlay" \
  -c:v libx264 composited.mp4

# Timelapse from image sequence
ffmpeg -y -framerate 30 -i frame_%04d.jpg -c:v libx264 -pix_fmt yuv420p timelapse.mp4

# Video from single image + audio
ffmpeg -y -loop 1 -i cover.jpg -i audio.mp3 -c:v libx264 -tune stillimage \
  -c:a aac -shortest output.mp4

# Crossfade between two videos (1 second)
ffmpeg -y -i first.mp4 -i second.mp4 \
  -filter_complex "xfade=transition=fade:duration=1:offset=FIRST_DURATION_MINUS_1" \
  -c:v libx264 crossfade.mp4

CRF Quality Guide

CRF	Quality	Use case
18	Visually lossless	Archival, professional editing
23	High quality	Default ffmpeg, good for most uses
28	Good quality	Recommended for sharing, smaller files
32	Acceptable	Social media, aggressive compression
38+	Low quality	Previews, thumbnails

Lower CRF = larger file, better quality. Each +6 roughly halves the file size.

Format Quick Reference

Container	Video codec	Audio codec	Use case
`.mp4`	H.264 (libx264)	AAC	Universal compatibility
`.webm`	VP9 (libvpx-vp9)	Opus	Web, smaller files
`.mkv`	H.265 (libx265)	Opus	Maximum compression
`.mov`	ProRes	PCM	Professional editing
`.gif`	—	—	Animations, no audio

Error Handling

"No such file" — Check the input path, use absolute paths
"Unknown encoder" — Codec not compiled in. Check ffmpeg -encoders
"Permission denied" — Output path not writable
"Avi header too large" — Corrupted input, try -fflags +genpts before input
Timeout — Video is very long. Increase timeout or use --preset ultrafast
Out of memory — Reduce resolution first, then process

beveradb/claude-code-config

claude-code-config とは？

お気に入りのAIに質問する

ドキュメント