How to get usable shots out of AI video tools

Every AI video tool demos beautifully and fails specifically. The failures are not random: they cluster around hands, text, crowds, fast motion, long camera moves and anything that has to stay the same from one shot to the next. Getting usable shots is mostly a matter of planning around those clusters instead of hoping the model has finally learned to draw fingers.

These are the habits that survive contact with a deadline. For the underlying mechanics, read how AI video models work; for prompt craft in general, the prompting best practices go deeper.

Write the shot, not the story

The generation unit is a shot, not a scene. If you write a paragraph of plot, the model will try to cram it into a few seconds and produce a muddle. Write one camera set-up instead: who or what is in frame, what they do, where they are, how it is lit, what the lens is doing.

Luma's own documentation for Ray2 gives a useful ordering for this: main subject, action, details, scene, style, camera, reinforcer. Runway's multi-shot guidance says the same thing more bluntly — keep each shot's prompt focused on a single scene, and if you want a sequence, provide a shot list rather than one sprawling description.

That is also why a storyboard helps even when nothing is drawn: a shot list of three to five set-ups, each with a single action, is a better input than a treatment.

Feed the model a picture

Text is the weakest form of control. If a character, product or location has to look right, give the model a reference image or a first frame. Most serious tools support this now, and the vendors' own notes are consistent about what makes a good reference:

• Resolution and quality. Use the sharpest source you have. Compressed, soft or noisy references produce softer output.
• Isolation. One clear subject, free of clutter and competing elements.
• Lighting. Even, neutral light reproduces most faithfully; heavy shadows and colour casts carry into the result.
• A small cohesive set rather than one image, when you are defining a style.

One warning the documentation states plainly: you are responsible for the rights to any likeness you submit. Using a reference of a real person without permission is a legal problem before it is an artistic one.

Say what you do not want

Generations drift toward the average of what the model has seen, which is why unwanted faces appear in crowds and stray objects at the edges of frame. Negative prompts are the cheapest fix. Google's own Veo API example passes a negative prompt to suppress unwanted sounds alongside the positive prompt; the same logic works visually.

Keep the camera modest

Big moves — crane up and through a window, whip pan, orbit around a falling object — are where temporal consistency collapses. A slow push, a small handheld drift or a static frame holds together far more reliably, because generous movement gives the model more frames to get wrong.

Camera language does work: Google's published Veo examples describe shot sizes, moves and lens behaviour in sequence, and the results follow. The trick is to ask for one coherent move, not three in the same clip.

Test cheap, then commit

Generate at the lowest cost and shortest duration that can answer the question "is this the right shot?" Long, expensive renders are for shots you have already proven. Veo 3's API was priced per second of output, with cheaper fast tiers for exactly this loop; Luma's Ray2 produces short base clips that extend, so you can judge a shot before paying for length.

Pin your model version when a client project depends on a look. Runway's API documentation lets you pin a dated workflow version instead of tracking the latest, which is the difference between a reproducible pipeline and a moving target.

Build continuity deliberately

Consistency across shots is not a prompt problem; it is a pipeline problem. Three techniques do most of the work:

1. Reference and keyframe every returning element. Same image in, same look out.
2. Reuse the exact style language across a sequence. Runway explicitly recommends repeating the same product or style language across shots for visual consistency. Renaming a style between shots invites a new interpretation.
3. Block the sequence so that changes are motivated. A costume change, a new location, a night scene — anything that explains a visible difference also hides the model's inconsistencies.

Judge takes like an editor

A take that looks good in motion often dies on the cut. Screen each candidate for the things that end a shot: a hand that melts, a mouth that drifts out of sync, text that turns to glyphs, an extra who walks through a wall. Watch at normal speed once, then scrub the first and last frames — those are your in and out points.

Keep a shortlist, not a favourite: take three with a slightly different camera is often what the edit needs.

Plan sound separately

Native audio — dialogue, effects and music generated with the picture in a single pass — changes the workflow. Veo 3 was Google's first model to do this, and multi-shot recipes now generate audio by default. It is useful for scratch tracks and lip-sync tests.

But treat generated audio as a first pass. Dialogue intelligibility, room tone and music rights still need human judgement, and a shot can be visually perfect and unusable because the line is garbled.

Disclose, and keep records

Two things are converging. Platforms are labelling more aggressively: YouTube applies prominent labels to photorealistic AI content and now automatically labels content its systems detect as significantly AI-generated. Regulation is arriving too: the EU AI Act's Article 50 requires providers to mark synthetic outputs in a machine-readable format and requires deployers to disclose deepfakes when they use them.

The practical habit is simple. Preserve provenance — Content Credentials, the C2PA standard, attach a signed record of origin and edits to a file — and keep a project log of prompts, references, model versions and seeds. When a client or a festival asks how a shot was made, that log is the answer.

When a shot keeps failing, change the shot

The most common mistake is rewriting the prompt for the tenth time. If a shot fails repeatedly, the shot is asking for something the model cannot do reliably. Split it in two. Move the camera. Cut before the difficult action happens. Put the object out of focus. Constraint is not a compromise here; it is the craft.

The short version

• Write one shot per prompt, with a clear subject, action and camera.
• Reference anything that has to stay consistent, and check the rights.
• Use a short negative prompt for known artefacts.
• Iterate cheaply, then render the winner at full length.
• Judge takes on the frames you will cut to, not the ones that look best paused.
• Plan sound as its own layer, and disclose the tools you used.