AI Filmmaking Prompting: What Actually Works

Prompting for film is not a writing contest. It is shot design done in words, and most of the craft is in deciding what belongs in the prompt at all. Two failure modes account for most disappointing generations: the prompt that describes a feeling and leaves the camera unmentioned, and the prompt that describes the whole scene, including everything already visible in the reference image, until the model has no room left to move. A working prompt sits between those two. It names a subject, an action, a camera and a light, and then stops.

Prompt one shot at a time

Every generation should correspond to one shot you intend to cut. Runway's Gen-4 video guide describes five- and ten-second clips and suggests planning each clip as one scene, not a compressed sequence. This is not a limitation to work around; it is a discipline that improves the film. If a shot needs three separate actions, it is probably three shots.

Write the prompt in the edit's vocabulary. If you cannot say what the shot does for the scene, no amount of prompt detail will fix it.

Use four slots, then stop

A reliable video prompt has four slots, roughly in this order:

1. Subject and action. Who is in frame and what they physically do. Write it as a single visible action, not an intention.
2. Camera. Size and movement: a medium shot, a slow push in, a locked-off wide, a handheld track. These terms are worth learning properly, and Runway's camera reference pairs each one with an example output.
3. Light and atmosphere. Direction, quality and colour: low sun from frame left, a single practical lamp, overcast diffuse light, haze. Lighting words change an image more than style words do.
4. Style and optics, briefly. Film stock, grain, lens character, palette. One or two descriptors are usually enough; a paragraph of adjectives tends to compete with the shot.

The image prompt has a slightly different job. Runway's Gen-4 image guide lists subject, scene, composition, lighting, colour, style, focus, angle, text and mood as usable elements, and notes that fuller sentences buy more control while short prompts buy more variation. Neither is wrong. What matters is that you know which one you are choosing.

Describe the motion, not the photograph

When a generation starts from an input image, restating the image in words tends to reduce motion: the model receives the same information twice and becomes cautious. Runway's guidance is direct about this — use the text prompt to describe what should move, and refer to existing subjects in general terms such as "the subject" or "she". Save the detailed description for elements you want to add or change.

This is also why the best motion prompts are modest. "She turns her head toward the window, then looks down" will usually beat a sentence that attempts to choreograph six seconds of performance. You are directing one gesture, not a scene.

Say what should happen, not what shouldn't

Negative phrasing is a trap in most current models. Runway's guides state that negative prompts are not supported for Gen-4 image or video, and that describing what should not appear can produce the opposite result. "No camera movement" is a bad instruction; "locked camera, the subject remains still" is a good one. "A man with no hair" becomes "a bald man".

The practical rule: every sentence in the prompt should describe something visible in the frame. If a sentence only describes an absence, rewrite it as a presence.

Separate identity from style

Consistency comes from references and repetition, not from longer descriptions. Current reference systems can hold a character across lighting and locations, and Runway's References guide recommends using a clean, evenly lit, neutral-expression image as the base — and allows up to three reference images per generation. Use one reference for the person, one for the place, one for the grade, and keep them the same across the film.

Keep a written character sheet as well: age range, build, hair, wardrobe, two or three distinguishing features. When a take drifts, compare it against the sheet before blaming the model. Drift is usually a prompt that quietly changed.

Keep a log that makes takes repeatable

Every approved take should be reconstructable. Record the model and version, the settings, the reference images, the exact prompt, and one line on what changed from the previous attempt. This turns prompting from a slot machine into a process with a memory.

Change one variable at a time. If you alter the camera, the light and the wardrobe together, you learn nothing from the result except whether you liked it. The same logic applies to batches: generate variations that differ in one dimension so the comparison is honest.

Prompt for the edit

A prompt can produce a beautiful shot that is useless in the cut. Guard against this by prompting for what the edit needs: an entrance, an exit, a look, a reveal, a reaction. Movement that starts and ends within the clip is easier to cut than movement that is still accelerating when the clip stops. Static shots are achievable but need to be requested plainly and reinforced, because models are built to move.

If a shot keeps failing, do not keep re-rolling the same prompt. Change one of three things: the prompt (make the action more physical), the reference (make the identity stronger), or the shot (cut it, or split it in two). The third option is the one most filmmakers forget, and it is usually the fastest. When the prompting is working, the rest of the workflow is covered in the step-by-step first film guide; when it is not, the AI video generation tips cover the common fixes. Either way, keep the prompt short enough that you could say it out loud to an actor and have them understand the shot.