video-edit

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.