scene-expander

scene-expander

Expand a one-line idea or script line into the Cinematic Scene Blueprint — 12 labeled blocks consumed by parse-blueprint.sh. Picks hollywood-director for live-action; hands off animated beats to pixar-storyteller. Output is provider-agnostic — provider tuning is motion-choreographer.

When to use

• A script line, idea, or beat needs a full blueprint before any adapter runs.
• A /video:from-script run is parsing ## Scene N headings and needs each scene expanded.

Do NOT use when:

• The brief is already an 11-block cinematic prompt → call video-director directly to refine it.
• The brief is a static graphic → canvas-design.
• Provider-specific token tuning is the next step → motion-choreographer.

Procedure

Step 0: Inspect

1. Read the input line. Classify as live-action / photoreal or animated / stylized.
2. Live-action → load hollywood-director voice. Animated → hand off to pixar-storyteller (its procedure carries the acting / beat-decomposition lens). Hybrid (live-action with VFX) → hollywood-director; record VFX intent in ENVIRONMENT.
3. Check for an existing character.json lock under agents/reference/ai-video/<project>/characters/.

Step 1: Emit the 12 blocks

One label per line. Order is mandatory.

1. STYLE — stylistic anchor (live-action: film stock + decade, e.g. "Kodak 5219, 2015"; animated: specific film + year).
2. SUBJECT — character read; verbatim identity tokens when a lock exists.
3. ENVIRONMENT — location, time-of-day, weather, era; what the world does in response to the subject.
4. ACTION — anticipation / action / reaction with beat counts (0.5s / 1.2s / 0.8s). No adjective paragraphs.
5. CAMERA — position, height, distance, move as an INTENT class (static hold, push-in, pull-back, lateral track, handheld drift, orbit). Intent vocabulary, never one provider's move grammar — the per-provider encoding happens in motion-choreographer. Off-axis when on-axis is default.
6. LENS — focal length in mm and aperture intent. "Cinematic" alone fails. mm + aperture are cinematography units, not provider tokens — they stay.
7. LIGHTING — key / fill / back / practical named; "golden hour" requires a sun angle.
8. MOOD — one emotional read.
9. DIALOGUE — optional. If present: speaker: "line", one per line. Marks audio: native capability requirement.
10. AMBIENT SOUND — optional. Layer list (wind, traffic, crowd, ocean). Marks audio: native or routes to ffmpeg mux.
11. DURATION — seconds (integer or one decimal). Free value; per-provider clip-length clamping is the encoder's concern, not the blueprint's.
12. NEGATIVE — clichés to reject, load-bearing order top-first. Always names: centered framing, symmetric composition, generic "cinematic", soap-opera contrast.

Step 2: Self-review

1. Live-action / animated classification consistent across blocks?
2. LENS present with mm (live-action) OR stylistic anchor with film+year (animated)?
3. LIGHTING with a direction?
4. ACTION names beat counts, not adjectives?
5. DIALOGUE / AMBIENT SOUND present → the run requires audio: native adapter OR ffmpeg-mux fallback declared.
6. NEGATIVE ≥ 4 entries, load-bearing top?
7. SUBJECT verbatim from character.json when a lock exists?

Any "no" → revise that block.

Step 3: Validate

1. Pipe output through scripts/ai-video/lib/parse-blueprint.sh — exits 0 and emits valid adapter-contract JSON.
2. No provider tokens (no aspect / model / duration flags).

Output format

1. scenes/<id>/prompt.txt — 12 labeled blocks, ready for parse-blueprint.sh.
2. scenes/<id>/blueprint.json — parser output, adapter-stdin ready.
3. scenes/<id>/review.md — one-paragraph rationale per non-obvious choice.

Gotcha

• The model wants to skip optional DIALOGUE / AMBIENT SOUND blocks silently — if they could plausibly belong, emit them; the parser treats missing blocks as null, not as an error.
• Live-action without LENS mm fails the parser's strict mode.
• Animated without a film+year anchor in STYLE drifts on every run.
• ACTION written as "the character does X dramatically" fails — adapters need verbs with beat counts.
• DIALOGUE forces audio: native requirement — flag this to the orchestrator so it picks an audio-native adapter (capability.audio = native).

Do NOT

• Do NOT emit provider tokens — that is motion-choreographer.
• Do NOT shape any block toward one provider's prompt grammar — the blueprint is the decoupling layer; the anti-leak schema test scans it for provider tokens.
• Do NOT skip the blueprint parser validation step.
• Do NOT paraphrase identity tokens when a lock exists.
• Do NOT mix live-action LENS prescriptions with animated STYLE anchors in the same scene — pick one mode.

Policies

Paths, enforcement model, and the full set: the media policy preamble.

The 12-block Cinematic Scene Blueprint is the policy choke point — every downstream skill (motion-choreographer, video-director) inherits whatever the blueprint encodes. Before emitting:

• likeness — when the SUBJECT block names or visually identifies a real person.
• public-figures — when the SUBJECT block is a recognised public figure.
• brand-impersonation — when STYLE / ENVIRONMENT references a recognised brand's visual identity.
• style — when STYLE anchors to a named living artist or studio as the primary signature.
• disclosure — every distributed blueprint output carries the AI-generation disclosure downstream.

Refuse-and-surface at the blueprint layer; do not push policy questions down to the adapter.