A six-layer framework for prompting AI video models

Why vague prompts fail

The same prompt can produce a masterpiece on one model and a blurry mess on another, and the most common cause is a prompt that is too vague. "A woman walking in a garden" gives a generic result. "Medium tracking shot of a woman in a flowing red dress walking through a sunlit Victorian garden, 35mm lens, golden hour lighting, shallow depth of field, gentle camera movement following her from the side" gives something cinematic.

The reason is simple: these models are trained on professional film and video data, so they respond to cinematography language, not everyday descriptions. Think like a director giving instructions to a camera operator. The more precise your direction, the less you iterate, and the less you spend.

Throughout this guide the work is done inside Venice AI, which puts the leading open-source and proprietary video models in one interface.

The six-layer framework

The framework works across every model. Each layer adds a kind of direction the model understands.

1. Subject and action. Who or what is the focus, what are they doing, and what emotional energy do you want? Example: a woman in a flowing red dress walking through a sunlit Victorian garden with a contemplative expression.
2. Shot type and framing. Wide (full environment), medium (waist up, balances subject and setting), or close-up (intimate detail). Then the angle: eye level (neutral), low angle (powerful, dramatic), or high angle (vulnerable).
3. Camera movement. Static, tracking (stays with the subject), panning (rotates horizontally to reveal more), or dolly push in/out (moves closer or farther for intensity). Specify slow, deliberate movement like "slow dolly forward" or "gentle tracking shot" for a cinematic feel.
4. Lighting and atmosphere. Golden hour (warm, soft, romantic), blue hour (mysterious twilight), or studio (precise, controlled, professional). Also call out light quality (soft or hard, diffused or direct), environmental effects (rain, fog, particles), and contrast.
5. Technical specs. Tell the model which lens and film stock to simulate: 35mm wide angle, 50mm natural, 85mm portrait, or macro. Add depth of field (shallow, deep, bokeh, rack focus) and film aesthetics like "35mm film grain," "anamorphic lens flares," or "Kodak Portra color palette."
6. Duration and pacing. Rhythm and speed. Many interfaces, including Venice, set clip length as a separate parameter rather than in the prompt, but you can still specify speed (slow motion, 120fps, time lapse) and transitions (smooth fade, hard cut).

You do not need every layer in strict order, but lead with shot type and subject/action. The more layers you include, the more granular your control.

Building a complete prompt

A reliable general structure is:

[shot type] of [subject] [doing action] in/at [setting], [camera movement], [lens/camera], [lighting], [atmosphere], [technical details]

Here is the framework applied to a single shot:

Perfect medium shot framing of a female athlete in vibrant athletic wear performing a dynamic leap dance move between buildings in a city's downtown, smooth dolly camera movement visible through motion blur trails, dramatic golden hour lighting with volumetric rays streaming through urban architecture, shot with 85mm lens creating beautiful bokeh depth of field, composed with dynamic pacing and rhythm, includes subtle translucent technical annotations, ultra high-quality 8K resolution professional cinematography, award-winning composition. Premium film production meets technical excellence.

Every clause maps to a layer, and the closing line sets an overarching style. This will always beat "woman jumping on the street."

Kling 2.5 and the duration trap

Kling 2.5 is strong at athletic movement, character animation, and motion fluidity. A good structure for it: detailed visual description, camera movement, professional cinematography, specific style reference, lighting conditions, and a quality tag, which closely follows the six layers.

The bigger guide here is duration. If a shot only needs five seconds, do not give it ten. The model fills extra time with whatever it thinks fits the prompt, which is how a basketball dunk turns into impossible, inhuman motion. Notice, though, how far the models have come: even when the movement is unrealistic, limbs stay clear and nothing morphs, unlike earlier generations. The fix is a tighter prompt and a few rounds of iteration.

Sora 2 for multi-shot scenes

Sora 2 builds multi-shot sequences from a single prompt, holding spatial relationships and consistency across angles. One prompt can produce an establishing shot, action, detail, and reaction in a single sequence. Like the others, it responds best to professional camera language.

Establishing wide shot of a bustling Parisian cafe at golden hour, slowly pushing into a medium shot of a woman reading at a corner table, then cutting to a close-up of an espresso cup, rack focus to her contemplative expression as she looks up toward the window. 35mm lens, warm cinematic lighting, shallow depth of field, seamless transitions, maintaining spatial consistency.

Run on Sora 2 Pro, this followed the prompt closely, including the rack focus from the espresso cup to the woman. Reach for Sora 2 Pro when you want multi-shot storytelling rather than a single clip.

Wan 2.5 for open-source lip sync

Alibaba's Wan 2.5 is open source, which on Venice means it is anonymized and noticeably cheaper. A 10-second 1080p clip runs about 165 credits, roughly half of Sora 2 Pro or less. It is also uncensored.

Its standout strength is lip sync. Give a character dialogue and it will speak the line, which is still hit or miss on other models.

Medium shot of a visionary figure in Renaissance Venice standing on a grand marble balcony overlooking the Grand Canal at golden hour, neoclassical futuristic aesthetic, ornate Venetian architecture, rich period costume, 4K 60fps. The character says "Don't sleep on Venice AI" with a commanding, visionary tone, emphasis on "don't" and "Venice AI." Perfect lip sync with mouth movements, theatrical hand gesture on "don't sleep," proud sweeping gesture toward the Venice skyline on "Venice AI," eyes gleaming with conviction.

Consider Wan 2.5 for multilingual work, music videos with singing, and any heavy character-narrative content.

Google Veo 3 and JSON control

Google Veo 3 and 3.1 let you prompt with JSON, which gives more precise, consistent control. Each key-value pair states one piece of direction unambiguously, which is especially useful when generating prompts programmatically or via an API.

The same scene as plain text versus what the JSON encodes:

A cyberpunk street market at night with neon signs, slow dolly forward at eye level, 35mm lens at 0.5x speed, neon lights, volumetric fog, high contrast, teal and orange color grading, cinematic feel.

When that data sits in JSON fields, the model reads the movement, angle, lens, and speed as discrete instructions instead of guessing. In the generated result the eye-level framing, slow dolly, fog, and high contrast all held. Reach for Veo 3 on commercial and big-vision projects where you have an exact plan, since duration is set in Venice rather than in the prompt.

Three techniques to iterate smarter

The 5-10-1 rule keeps costs down. Make five variations on a cheaper model like Kling or Wan, pick the best, run ten more iterations to refine the prompt, then do one final render on a premium model such as Veo 3.1 full quality (around 352 credits). Finding your shot this way might cost about 1,000 credits instead of thousands spent re-rendering on premium the whole time.

Negative prompting tells the model what to exclude: "blurry, low quality, distorted faces, warped hands, text artifacts, watermark." Veo 3 has a dedicated negative prompt field, especially via JSON. Kling needs it folded into the main prompt with words like "avoid" or "without." Sora has no negative prompt, so state the positive instead: say "very focused and crisp," "high quality," "very clear faces."

Style reference stacking combines film references for a unique look, for example "aesthetic combining Blade Runner 2049 color grading plus Se7en atmosphere plus Heat camera movement, anamorphic lens, cinematic bokeh." Stack two to three references; more than that dilutes the result.

You can also have a text model do the heavy lifting. Ask Venice Large what visual details define your chosen films, then feed its breakdown (palette, framing, camera behavior) into your video prompt. In the final demo, a Veo 3.1 full-quality render adhered closely to a detective-in-the-rain prompt built this way.

Wrapping up

You now have a repeatable system: the six-layer framework for structuring prompts, a sense of which Venice model fits athletic action, multi-shot scenes, lip sync, or JSON precision, and three techniques for iterating cheaply and shaping style. The gap between amateur and professional AI video is technique, not talent. Open Venice, pick a model, and start building shots.

Key takeaways

• Write prompts in cinematography language across all six layers: subject/action, shot type, camera movement, lighting, technical specs, and pacing.
• Match the model to the job: Kling for athletic motion, Sora 2 for multi-shot scenes, Wan 2.5 for cheap open-source lip sync, Veo 3 for JSON-precise control.
• Give a shot only the duration it needs, or the model invents motion to fill the time.
• Use the 5-10-1 rule, negative prompting, and style stacking to cut credit costs and sharpen results.