lazy-video

Drop any video into input/, tell Claude Code "edit this", get branded 9:16 reels in output/<name>/. Zero manual editing.

New here? Open Claude Code in this folder and say "set me up" — it detects your OS, installs the tools (with your permission), and configures everything. Prefer to do it by hand? See Requirements + Setup below.

Built for portrait screen-recording tutorials, but it takes any input: non-9:16 video gets cropped to portrait first, and a clip with no lesson markers becomes a single reel. Speech in any language works — the pipeline preserves code-switched / mixed-language captions instead of translating them.

The intelligence (crop window, where to cut, fixing captions, hook titles) is the LLM. Remotion is only the renderer — it does not decide cuts.

Pipeline

Stage 0   Claude crops non-9:16 input to portrait (HandBrakeCLI)   [only if needed]

uv run python run.py input/my-video.mp4
  ├─ 1 transcribe  faster-whisper (cuda, cpu fallback)    -> transcript.json
  ├─ 2 EDL         mechanical draft + LLM final pass       -> edl.json  [you approve]
  ├─ 3 cut         ffmpeg slice, remove flubs/silence      -> clips/
  └─ 4 render      Remotion: motion + karaoke captions + swipe CTA -> reels

Run once → pauses for EDL approval → run the same command again → finished reels (output/<name>/NN_<slug>.mp4).

Requirements

• ffmpeg + HandBrakeCLI on PATH
• Python 3.12 + uv
• Node + Remotion deps in remotion/
• NVIDIA GPU for whisper_device: cuda (set it to cpu in config.json otherwise)

Setup

uv sync                        # python deps (faster-whisper + cuda libs)
cd remotion && npm install     # remotion deps
npx remotion browser ensure    # headless chrome (one-time)

Config

All tunables live in config.json:

Key	Meaning
marker_phrases	spoken lesson boundary — the phrase you say between lessons (boundary = end of phrase)
flub_word	spoken re-take trigger; flubbed take cut, clean take kept
silence_threshold_sec	trim dead-air gaps longer than this
default_intensity	gentle | medium | hype retention motion
whisper_model / whisper_device	e.g. large-v3, cuda/cpu
output.*	1080x1920 H.264 30fps
brand.*	captions, CTA, intro, background music (below)

Captions: karaoke word-by-word reveal, ALL CAPS in a configurable font (default Anton), max 4 words on screen, positioned by brand.caption_position (raw "NN%" supported). Reels start cold (no intro sting) and end on a swipe-left CTA card (MDI icon). Set brand.bg_music to an audio file to loop it under the voice at brand.bg_music_volume (empty = no music).

Drop-and-go

See the skill: drop a video, say "edit this", approve the EDL, get reels.

Repo layout

src/        python pipeline (transcribe, edl, cut, render)
remotion/   renderer (Reel, Captions, motion)
assets/     caption font
input/      drop videos here
output/     finished reels per video
config.json single source of tunables
run.py      one command, all stages