Cinematic Video

Write each shot like a director: what the camera sees, how it moves, through which lens, in which light, and what you hear. Make one clip per shot and keep a shared style line across shots.

Inputs to collect

• Shot list: one line per shot (subject + action + place). Ask for the order if there is more than one.
• Look: genre or film style, color palette, time of day, and era.
• Camera: shot size, angle, and one move per shot.
• Sound: ambience, effects, dialogue, or silence.
• Format: aspect ratio (16:9, 9:16, 21:9), length per shot, and resolution.
• Anchor image (optional): a still that fixes the look or the character. Use it with image-to-video.

Models

<!-- tokenlab:models:start --> <!-- Generated from the live catalog by scripts/sync_models.py. Do not edit; change preferredModels in catalog.json. -->

Operation	Model	Pick	Use
text-to-video	veo3.1	default	Cinematic shot with sound
text-to-video	seedance-2.0	alternative	Alternative; long clips, 21:9 (not live-tested)
text-to-video	kling-v3.0	alternative	Alternative (not live-tested)
text-to-video	hailuo-h3	alternative	Alternative; negative prompt (not live-tested)
text-to-video	wan-3.0	alternative	Alternative; long clips (not live-tested)
text-to-video	grok-imagine-video	alternative	Low-cost draft
image-to-video	veo3.1-fast	default	Still to shot with sound
image-to-video	veo3.1	alternative	Higher tier (not live-tested)
image-to-video	seedance-2.0	alternative	Alternative; long clips, 21:9 (not live-tested)
image-to-video	kling-v3.0	alternative	Alternative (not live-tested)
image-to-video	hailuo-h3	alternative	Alternative; negative prompt (not live-tested)
image-to-video	wan-3.0	alternative	Alternative; long clips (not live-tested)
image-to-video	grok-imagine-video	alternative	Low-cost draft

This list refreshes daily from the live catalog. Preferred models come first; otherwise models are ranked by their 24-hour success rate. Before a call, confirm the model with GET /v1/models/{model} and read its price from GET /v1/models/{model}/pricing.

<!-- tokenlab:models:end -->

How to choose

Exact fields and values come from each model's contract.

• Veo 3.1: the default text-to-shot model with sound. 4, 6, or 8 s at 720p or 1080p. Audio is always on.
• Veo 3.1 Fast: the default still-to-shot model; same durations and always-on audio, faster and lower cost. Veo 3.1 is the higher tier for a still.
• Seedance 2.0: an alternative for longer clips (4-15 s) and 21:9.
• Kling 3.0: an alternative, 3-15 s.
• Hailuo H3: an alternative that takes negative_prompt. 6 or 10 s at 768p.
• Wan 3.0: an alternative for long clips, 2-30 s at 480p-1080p.
• Grok Imagine video: a low-cost draft. 1-15 s at 480p or 720p.

All use POST /v1/videos/generations with an explicit operation. Send output_audio or negative_prompt only when the operation lists it.

Workflow

Run each call with tokenlab-run (contract → estimate → call → poll → result). Commands below use its helper, ../tokenlab-run/scripts/tokenlab_run.py, a path relative to this skill's folder. If you installed this skill without tokenlab-run, send the same requests with curl. --wait stops after 100 s; a longer task keeps running, so continue with the resume command it prints.

1. Plan the shots. Write one prompt per shot with the template in Technique. Keep one style line and reuse it word for word in every shot.

2. Read the contract: python ../tokenlab-run/scripts/tokenlab_run.py contract MODEL --operation text-to-video (or image-to-video). Use only listed fields and values.

3. Estimate the cost: python ../tokenlab-run/scripts/tokenlab_run.py price MODEL, then per-second price × duration for each shot. Veo audio is always on, so use its with-audio price. Use the Official price as the upper bound. Tell the user the total before you start.

4. Make a cheap test of the hardest shot first (short duration, lowest resolution, or a low-cost model). Fix the prompt, then make the final shots.

5. Build each body:

   {"model": "MODEL_ID", "operation": "text-to-video", "prompt": "Slow dolly-in, eye level, 35mm lens. A lighthouse keeper in a yellow raincoat climbs wet stone stairs at dusk; waves crash below. Cold blue light with a warm lamp glow from the window. Film noir mood, light rain. Sound: wind, surf, his boots on wet stone.", "duration": 8, "resolution": "720p", "aspect_ratio": "16:9"}
   

   Take MODEL_ID from the Models list above for the shot's operation. Set duration and resolution to values its contract lists.

6. Call and wait: python ../tokenlab-run/scripts/tokenlab_run.py run /v1/videos/generations --json @shot1.json --wait. Save each task ID. Poll every 5-10 s.

7. If a wait times out, resume with python ../tokenlab-run/scripts/tokenlab_run.py poll TASK_ID. Never create a duplicate task.

8. Return each shot's video URL, model, task ID, and estimate, in shot order.

Technique

• Prompt template (the parts Google lists for Veo prompts): camera move and position → composition → subject and action → place → lens and focus → light and ambiance → style → sound. Leave out any part you do not need.
• One camera move per shot. "Slow dolly-in" or "pan left", not both. Make a second shot for the next move.
• Name the lens and focus. "35mm, deep focus" for context; "85mm, shallow focus" for faces; "macro lens" for textures; "wide-angle lens" for space.
• Light sets the mood. Name its direction and color: "low warm sun from behind", "cold blue night with neon spill", "single overhead practical".
• Sound on Veo. Veo always makes audio. Describe effects and ambience explicitly. Put dialogue in quotes: She whispers, "We're late."
• Consistency across shots. Reuse the same character description and style line. For a fixed character or set, make a still first (see tokenlab-character-consistency) and use image-to-video from it.
• Exclusions. Use negative_prompt only on models that list it (for example Hailuo H3). Elsewhere, describe the result you want instead.
• Shot vocabulary: see references/shot-language.md.

Quality bar

• Each shot shows the requested shot size, camera move, and light.
• The camera move is smooth and single; no unasked cuts inside a shot.
• Characters, wardrobe, and the style line stay the same across shots.
• Sound matches the action (Veo) or the shot is silent when asked (models with output_audio).
• The user saw the total estimate first, and the answer lists every shot's model and task ID.

Related skills

• tokenlab-run: the call, poll, and result mechanics.
• tokenlab-product-ad-video: product-focused shots.
• tokenlab-animate-image: animate a single still.
• tokenlab-keyframe-transition: fix the first and last frame of a shot.
• tokenlab-extend-video: make a shot longer.
• tokenlab-prompting: prompt craft per model family.