A YouTube content-creator assistant written in plain English. No framework, no app, no server: Aurora is a Claude Code skill plus four role agents. You type /aurora create "your idea" and get back a finished 10-minute video with voice, animation, subtitles, three thumbnails and a paste-ready YouTube description.
Most "AI video" tools are a pile of templates that stamp out the same video with swapped words. YouTube calls that slop, and demonetizes it. Aurora is built the other way round:
- A narrator, not a voiceover. Each channel defines a persona (name, story, lore, catchphrases) in one file. Aurora writes and narrates as that character, consistently, video after video.
- Human in the loop at the right place. One checkpoint per video: you review a draft with scene numbers burned into the corner ("scene 10: the joke is too slow"). Everything before runs unattended.
- Facts are checked. Every external claim carries a source URL or is cut or attributed. An anti-slop gate (
slop_check.py) blocks packaging when the video looks templated, repeats earlier ones, or breaks YouTube's metadata rules. - It learns. Feedback from subscribers or yourself and rules mined from your YouTube analytics are stored in your project and applied to every next video.
- It is cheap to run. A cost ladder pushes easy, bounded chores to free or small models and keeps the paid model for planning and checking.
- It is software you can read. The whole product is Markdown prompts plus a few small Python and Node helpers. Fork it, change a sentence, change the behavior.
| Output | Details |
|---|---|
video.mp4 |
1080p, about 10 minutes, H.264, loudness-normalized |
| Voice | ElevenLabs narration with word-level timing, so visuals land on the right word |
| Visuals | Three engines picked per scene: web (HTML/GSAP via Playwright), Blender (3D hero moments) and Manim (code, charts, diagrams) |
| Subtitles | subtitles.srt built from the alignment |
| Thumbnails | Three variants (A/B/C) ready for YouTube's Test & compare |
package.md |
3 title options, description with chapters and sources, tags, pinned comment |
| Shorts | 2-4 vertical clips cut from the finished video (/aurora shorts) |
/aurora interview # once per channel: persona, voice, palette, tone
/aurora create "why sourdough starters fail"
/aurora create auto # theme from trends + your preferences + what worked before
/aurora feedback "the voice felt fast"
/aurora batch # a week of videos: approve 7 themes, review drafts later
/aurora analise ~/Downloads/studio-export.zip
/aurora shorts <slug>
/aurora publish <slug>
Full reference in the docs.
curl -fsSL https://raw.githubusercontent.com/igormcsouza/aurora/main/install.sh | bashThe installer asks where to put Aurora: Claude Code, Codex, both, or other (you name the folders). Skip the question with ... | bash -s -- --target claude|codex|both|other. It clones Aurora to ~/.aurora and symlinks the skill and the four agents into your host's folders (~/.claude/ for Claude Code, ~/.agents/ for Codex); run it again to update. Start a new Claude Code session, cd into your channel project and run /aurora interview. You will need ffmpeg, Node.js 20+, Python 3 and an ELEVENLABS_API_KEY; Aurora's preflight checks everything and tells you exactly what is missing. See Install.
skills/aurora/ SKILL.md (the program), state.py, tiers.py, organize.py, slop_check.py,
preflight.sh, kit/ (web, Blender and Manim render helpers)
agents/ video-researcher, video-scriptwriter, video-animator, video-producer
docs/ the mkdocs site
The skill is the program; the agents are its functions. Channel-specific facts never live here, they live in your project's preferences.json, so one install serves any number of channels in any language or niche.
Aurora is early and opinionated, and there is plenty to do: a YouTube API upload step, a configurable TTS engine, sturdier Blender heroes, more web templates, more engines. Read CONTRIBUTING.md, pick something from the known issues, and open a PR. Improving a sentence in SKILL.md counts as a code change here.
MIT © 2026 Igor Souza