Quickstart¶
Run everything from your channel project root. Every step has a worked example with the output to expect in Examples.
1. Tell Aurora about the channel¶
A short conversation (usually 8-15 exchanges): audience, niche, language, the narrator persona and her story, tone and humor, voice, palette, mascot, topics you love or avoid, how involved you want to be. Aurora then writes preferences.json and adds a ## Video preferences section to your CLAUDE.md / AGENTS.md.
Optional, once per machine:
It runs one trivial task on each free/cheap model tier to prove they work here (a few cents).
2. Make a video¶
Aurora runs preflight, researches, plans, writes the script, records the voice, and renders a low-quality draft with the scene number burned into the top-right corner.
3. Review the draft¶
You get the draft path, a summary, and what is verified versus unverified (audio is always unverified by the assistant: only your ear judges it). Answer by scene:
Only the named scenes are redone. General notes ("voice too fast") are stored as feedback for all next videos.
4. Final and package¶
Once you approve, Aurora renders the final (no badge), mixes loudness, and builds package/: video.mp4, three thumbnails, subtitles.srt and package.md with titles, description, tags and pinned comment. The anti-slop gate runs first and a FAIL blocks packaging.
5. Upload, then publish¶
Upload in YouTube Studio using package.md, then:
This moves the video to published_videos/, keeps what rebuilds it and trashes the rest (reversibly). Later, feed results back:
A week at a time¶
You approve 7 themes once; Aurora builds seven drafts; you review them whenever you like with /aurora review ("drafts 1, 3 approved; draft 2 has an issue at scene 5").