The pipeline

State for each video lives in draft_videos/<slug>/state.json, managed only through state.py. Steps: preflight, topic, plan, script, voice, draft, review, final, package. Every step is resumable; /aurora resume continues from the first pending one and never redoes finished work.

flowchart LR
    P[preflight] --> T[theme] --> PL[plan] --> S[script] --> V[voice] --> D[draft render]
    D --> R{{your review}}
    R -->|scene fixes| D
    R -->|approved| F[final] --> K[package] --> U[you upload] --> PUB[publish]

Step	Who	Output
0 Preflight	main chat	./video-kit ready, run_prefix printed
1 Theme	video-researcher	3 ranked options; Aurora picks one
2 Plan	video-scriptwriter	plan.json with scenes, claims, sources.md
3 Script	video-scriptwriter	narration.md in the persona's voice (about 1500 words)
4 Voice	video-producer	scenes/NN.mp3, cues.json (word timing), durations.json
5 Draft	video-animator + video-producer	low-quality draft, scene badges, contact sheet, qa.md
6 Review	you	scene-numbered feedback
7 Final	video-animator + video-producer	1080p, CRF 17, about -16 LUFS
8 Package	video-producer	package/ + anti-slop gate

Facts

Every external claim in plan.json has at least one source_urls. Sources are read in full; what cannot be validated is cut or attributed ("according to X"); paywalled means attributed, never verified. Opinions and local code runs are kept in a separate list. slop_check.py fails a plan with unsourced claims.

Voice

ElevenLabs eleven_v4 with the voice_id from preferences, requested per sentence with neighboring context and a fixed seed to prevent accent drift in long scenes. Digit strings and acronyms are spelled out in the TTS input only. Alignment timestamps drive subtitles and visual beats, so a fact appears when it is spoken.

Package rules (enforced)

• No < or > anywhere in titles, description or tags (YouTube rejects them; this includes ->).
• Description at most 5000 characters and 5000 UTF-8 bytes; titles at most 100 characters; about 15 tags, at most 500 characters total.
• First two description lines are the hook; chapters from real scene starts; sources only from sources.md; an AI-use note in the channel's language; at most 3 hashtags.

Anti-slop gate

python3 ~/.claude/skills/aurora/slop_check.py <slug> checks template variety, overlap with earlier videos, unsourced claims, the metadata limits above and engine balance (warns above 70% of one engine, a single engine, three identical scenes in a row, fewer than 2 evidence scenes, no hero). A FAIL blocks packaging. It cannot judge originality: you do. Disclosure of synthetic content is recommended and left to you.