From a written prompt to a finished clip

The four-part prompt

Prompts that render well share a shape: a subject, what the subject does, how the camera behaves, and what the light is doing. "A red kite" is an image. "A red kite dipping and climbing above a windy beach, camera tracking from the sand, late afternoon backlight" is a shot. Each clause gives the model one thing to animate.

Order matters less than completeness. If you leave out the camera, the model will pick a static frame or an arbitrary drift. If you leave out the light, you will get flat midday illumination. Say what you want and the tiers will each attempt it within their limits.

• Subject: who or what is in frame, with one or two defining details
• Action: a single continuous motion that fits the clip length
• Camera: static, pan, tilt, dolly, tracking, handheld, aerial
• Light and look: time of day, direction, grain, color temperature

One motion per clip

The most common failure is asking for too much. A five-second clip can hold one camera move and one subject action. "She opens the door, walks to the window, pulls the curtain and turns around" is four beats; the model will compress or drop some of them. Split a sequence into separate prompts and render each as its own clip.

Clip length should match the action. A splash, a pour or a door opening fits in four seconds. A slow landscape reveal wants ten or more, which is where the Draft tier's twenty-second ceiling is useful for establishing shots.

Words the models understand

Camera vocabulary carries real weight. "Dolly in" produces forward travel, "tracking shot" follows a moving subject, "slow pan left" rotates the view, and "handheld" adds the small shake that sells realism. "Macro" tightens on texture; "aerial" or "drone" lifts the viewpoint.

Lighting words do the same work. "Golden hour", "overcast", "neon", "backlit", "rim light" and "soft window light" each push the render toward a recognisable look. Combining one camera term with one lighting term is usually enough; stacking five adjectives rarely helps.

Avoid negatives. Writing "no people" tends to produce people. Describe what should be in frame instead, such as "an empty platform".

Testing prompts cheaply

Write the prompt, render a two or three second Draft at 720p, and look only at the motion. If the camera does what you asked and the subject holds together, promote the exact same text to Standard or Cinematic. If it does not, edit one clause and try again. Draft iterations cost a credit per second, so five attempts on a short clip are cheaper than one blind Cinematic render.

Once a prompt works on Draft, compare mode is the fastest way to learn what the higher tier adds for that particular subject. Textured, detailed scenes gain the most; simple graphic scenes often look fine on Standard.