vstoos/vlog-film

An agent skill that edits raw trip and car vlog footage into a finished HDR (HLG 10-bit) YouTube
film: a teaser, jump cuts on speech, dash-cam road inserts under music, photo and still inserts
with captions, an animated route map, licence-plate blurring, levelled voice, loudness-normalised
final render, chapters, a thumbnail, face blurring for passers-by, subtitles in several languages, and dubbed
audio tracks in your own cloned voice with automatic glitch checks. Then it runs review
rounds: you watch, send notes with times ("12:22 the end of the phrase is cut"), and the agent
checks the source, fixes the edit and rebuilds.
It works with any agent that reads the Agent Skills SKILL.md format and
can run shell commands: Claude Code, OpenAI Codex, Hermes Agent and others.
Built on, and tested with, one real film: a 24-minute Ukrainian-language farewell to a 2013 Toyota
Yaris (iPhone 10-bit HLG clips, Sony dash-cam timelapse, photos, an OBD Torque screen recording).
references/example_yaris.py is that film's complete edit.
> Українською: це скіл для агентів (Claude Code, Codex, Hermes), який монтує відео з поїздок у
> готовий HDR-фільм для YouTube: тизер, склейки по мовленню, вставки з відеореєстратора під музику,
> фото з підписами, анімована карта маршруту, розмиття номерів, вирівняна гучність, розділи,
> обкладинка і субтитри кількома мовами. Встановлення нижче: агент може зробити його сам.
## Maintenance
This is a personal project shared as-is. I may not have time to review pull requests or answer
issues. Please fork or clone it and adapt it to your needs; ideas and wishes are welcome as issues,
with no promise of a reply.
## What you need
| | Tested with | Why |
| OS | Windows 11 | macOS / Linux should work with small changes (fonts, encoder), untested |
| GPU | NVIDIA RTX 3090 (24 GB) | WhisperX speech recognition (CUDA), hevc_nvenc encoding, libplacebo (Vulkan) |
| Python | 3.11 | scripts; a venv with WhisperX, torch (CUDA 12.8), onnxruntime-gpu, OpenCV, pillow-heif |
| ffmpeg | 9.0 "full" build (gyan.dev) | needs libplacebo, Vulkan, hevc_nvenc or libx265 |
| Shotcut | 26.x (its melt) | the final render of the MLT project; you can also open the project in Shotcut |
| Disk | ~10 GB | venv with torch and models, plus renders |
| Music | YouTube Audio Library tracks | safe for monetised videos |
Speech recognition models (WhisperX large-v3, ~3 GB) download on first use. The licence-plate
detector is morsetechlab/yolov11-license-plate-detection
(AGPL-3.0), and faces are found with OpenCV's YuNet
(MIT); both are downloaded by the setup script, not included here.
## Install
Easiest: open your agent and say "install the vlog-film skill from https://github.com/vstoos/vlog-film".
It can follow the steps below; the setup script checks everything first and installs only after
you agree.
By hand:
bash
git clone https://github.com/vstoos/vlog-film.git C:/Data/vlog-film
cd C:/Data/vlog-film
python setup/setup.py # check: what is there, what is missing
python setup/setup.py --install # install missing parts (winget on Windows), create .venv, download the plate model
python setup/install_skill.py # make the skill visible to Claude Code, Codex, Hermes, ...