video-edit

name

video-edit

description

Use this skill to CUT, TRIM, CLEAN, EDIT existing video footage. Triggers: 'schneide das Video', 'edit/cut/trim this video', 'remove silence/fillers', 'concat videos', 'add subtitles', 'edit these into a launch video', 'make a recap from these clips', any task that takes existing video file(s) as input and produces a polished edited output. Audio-first approach: transcribe → identify cuts on word boundaries → FFmpeg-based EDL render → self-evaluate cut points. Do NOT use for: (1) creating animated videos from scratch — use the `video` skill (Remotion). (2) understanding/classifying video content (describe scenes, extract timestamps for visual events, classify) — use a multimodal LLM directly. (3) raw single-command FFmpeg ops where you already know the exact filter chain — just run FFmpeg directly.

Video Editing Skill

Audio-first AI-driven editing of existing video footage. Inspired by browser-use/video-use. The premise: read transcripts, don't dump frames — keeps token cost manageable and lets the LLM reason on speech boundaries, fillers, and silence.

When to use

• You have one or more raw video files and want a clean cut version.
• Remove "umm / uh / false starts" or long pauses.
• Concat multiple takes into a single narrative.
• Add subtitles burned in or as SRT sidecar.
• Light color correction per segment.
• Generate B-roll-inserts at silence gaps from existing clip pool.

Core workflow

1. Inventory + transcribe

# Inventory: list each source clip with duration, fps, codec
for f in *.mp4 *.mov; do
  ffprobe -v error -show_format -show_streams "$f" -of json
done > inventory.json

# Transcribe with word-level timestamps. Options:
#   - The Atlas built-in `stt` skill (CPU-based, no API key needed)
#   - ElevenLabs Scribe (best-in-class for diarization + word timestamps)
#   - OpenAI Whisper API (cheap, word-level via verbose_json format)
stt --language de input.mp4 > transcript.txt

The choice depends on what you need:

• Just text + rough timing: stt skill (free, runs locally).
• Per-word timestamps + diarization: ElevenLabs Scribe API.
• Per-word timestamps, no diarization: OpenAI Whisper with response_format=verbose_json and timestamp_granularities=["word"].

2. Pack transcripts to takes_packed.md

Combine all transcripts into a single human-readable markdown file (~5–15 KB). Schema:

# clip_01.mp4  (12.4 s, 1080p30, h264)
[00:00.00] Hallo, mein Name ist Max,  ◀ speaker 1
[00:01.34]   (umm)                    ◀ filler
[00:01.89] und ich zeige euch heute,  ◀ speaker 1
...

Why pack into markdown? The LLM reads this file as primary source of truth — frames are too expensive (1080p30 × 60s = 1800 frames × ~258 tokens each ≈ 460k tokens). Markdown is ~5k.

3. LLM proposes EDL (Edit Decision List)

Prompt the LLM with takes_packed.md + user goal ("edit these into a 60-second launch video"). Expected output:

[
  {"src": "clip_01.mp4", "in": 0.00, "out": 1.30,  "fade_in_ms": 0,  "fade_out_ms": 30},
  {"src": "clip_01.mp4", "in": 1.89, "out": 12.40, "fade_in_ms": 30, "fade_out_ms": 30},
  {"src": "clip_03.mp4", "in": 2.10, "out": 8.50,  "fade_in_ms": 30, "fade_out_ms": 30, "subtitle": "Was wir gebaut haben"}
]

EDL cuts MUST land on word-boundaries from the transcript — never mid-word.

4. Render EDL via FFmpeg

ffmpeg -f concat -safe 0 -i playlist.txt \
       -vf "subtitles=subs.srt:force_style='FontSize=20'" \
       -c:v libx264 -preset slow -crf 18 \
       -c:a aac -b:a 192k \
       final.mp4

For per-segment color grade or 30ms audio fades at every cut, see references/ffmpeg-edl-render.md.

5. Self-evaluate

After render, the LLM should visually + auditively check each cut point. Generate small composite (3 frames before + 3 frames after the cut):

# 6 frames around timestamp $t
ffmpeg -ss $((t-0.1)) -i final.mp4 -vframes 6 -vf "fps=30" cut_check_$t_%d.png

If a cut looks visually jarring (large jump-cut without B-roll cover) or audio pops, adjust the EDL and re-render. Cap at 3 self-eval iterations.

Environment

Required:

• ffmpeg ≥ 6.0
• A transcription path (built-in stt skill, or ElevenLabs/OpenAI API key)

Optional:

• yt-dlp for online video sources

Output convention

Put edits in <source_dir>/edit/:

<source_dir>/
├── clip_01.mp4
├── clip_02.mp4
└── edit/
    ├── takes_packed.md       # transcripts
    ├── edl.json              # cut decisions
    ├── playlist.txt          # FFmpeg concat-demuxer input
    ├── subs.srt              # subtitles
    └── final.mp4             # the deliverable

Common pitfalls

1. Mid-word cuts — audio sounds clipped, fix EDL to use only word_end timestamps.
2. Frame-dump anti-pattern — extracting all frames for the LLM blows token budget. Read transcript instead.
3. Subtitle font collision — use force_style='FontName=Arial' to override system defaults.
4. Concat demuxer needs same codec — re-encode mismatched clips first with -c:v libx264 -c:a aac.
5. 30ms fades on every cut mask micro-pops. Standard for production cuts.

See also

• video skill — to CREATE animated content from scratch (Remotion)
• stt skill — for raw transcription
• references/ffmpeg-edl-render.md — full FFmpeg EDL-render recipes
• browser-use/video-use — reference implementation, MIT license