Ultimate Video Maker

Claude Code makes the whole reel: script, hook, voices, music, motion, mix and caption. You post it.

▶ Live showcase: https://mbgrao.github.io/ultimate-video-maker/

Ultimate Video Maker is a set of Claude Code skills and a production pipeline for short vertical videos (Instagram Reels, TikTok, Shorts). It has two engines:

Engine	Use it for	Where
Remotion pipeline	Ads, explainers and story reels with real voices, composed music and frame-exact motion	ai-video-skills/
/brag	Fast launch videos made from a project's own code and site (HyperFrames)	skills/brag · BRAG.md

Showcase

Every video below was made end to end by Claude Code with this repo. The sources, scripts and timelines are in git. The videos and other media stream from the media server (MEDIA.md).

Reel	Language	Length	Watch	Project
You Probably Don't Need an App (localhostlabs)	English	42.7 s	▶ watch	localhostlabs-noapp-20261109
You Probably Don't Need an App (Hindi cut, same visuals)	Hindi / Hinglish	52.8 s	▶ watch	same project
This is not a love story (MealMates)	English	51.3 s	▶ watch	mealmates-reel2-20261107
Deal 2 ki, banda 1 (MealMates, every line voiced)	Urdu	51.9 s	▶ watch	mealmates-ad1-20261107
Deal 2 ki, banda 1 (MealMates, v1)	Urdu	35.0 s	▶ watch	same project
Staff nahi, system chahiye (localhostlabs)	Roman Urdu	53.7 s	▶ watch	localhostlabs-reel-20261105
Website vs App vs Software: Day 1 (localhostlabs)	English + Roman Urdu	51 / 57 s	▶ EN · ▶ PK	localhostlabs-30day-20261106
Who owns your code?: Day 2 (localhostlabs)	English + Roman Urdu	48 / 54 s	▶ EN · ▶ PK	same project
localhostlabs launch reel, made with /brag	Roman Urdu	44.2 s	▶ watch	examples/localhostlabs-reel

How a reel gets made

1. Brief → script. The script is written to a reel format (hook → turn → points → CTA). Hooks are scored with hookscore.py from the Instagram skills, and claims are checked against the live site.
2. Voices. ElevenLabs eleven_v3 with acting tags. Each language gets its own native voice, picked by audition and Whisper read-back. Character timestamps give word-exact timing.
3. Timeline. One scene per beat. Every effect is pinned to a spoken word (timeline.json): the slam lands on "app", the record scratch on "don't".
4. Pictures. OpenAI images for photoreal scenes. Character references keep the cast consistent across shots. Everything else is built as animated UI in Remotion, so every frame is a pure function of time.
5. Music and SFX. ElevenLabs music is composed with a plan cut to the reel's own sections. SFX land on their words.
6. Mix and QA. The voice-first mix levels each line, ducks the music under speech and loudness-normalises to −14 LUFS. Whisper transcribes the final mix, and contact sheets catch layout bugs before the full render.
7. Deliver. MP4, a poster baked in as frame 0, an .srt, and an Instagram caption that passes the ig-caption linter.

See ai-video-skills/PIPELINE.md for the full step order. Each step has a free option and a paid upgrade.

Quick start

git clone https://github.com/MBGrao/ultimate-video-maker && cd ultimate-video-maker/ai-video-skills
../scripts/fetch-media.sh                  # optional: showcase videos, voices, music, images (~280 MB)
./scripts/restore.sh                       # clones the pinned third-party skills + the 419-repo catalog (~28 GB)
cp config/.env.example config/.env.local   # add only the keys you buy (ElevenLabs, OpenAI, ...)
export PATH=/opt/homebrew/opt/node@22/bin:$PATH   # Remotion wants Node 22

Then ask Claude Code something like "make a 40-second English reel for my-site.com, voice and music included". The ai-video-pipeline skill routes the request through PIPELINE.md.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗