Most AI video disappointments are not model failures they are communication failures. Image models forgive a vague prompt because a still frame is a single moment; video models must animate your intent across time, and every ambiguous word becomes an unpredictable decision made twenty-four times per second. The creators whose reels stop the scroll have learned one thing: writing for motion is a different dialect of prompting. This guide teaches that dialect.

Why video prompts are a different language

An image prompt answers the question "what do I see?" A video prompt must also answer "what changes, and how fast?" When you prompt "a city at night," an image model renders a plausible skyline and stops. A video model must decide whether the camera moves, whether lights flicker, whether rain falls, whether the scene is a establishing shot or a handheld close-up and it will guess. Often wrong.

The fix is to treat every video prompt as a micro-screenplay with three non-negotiable elements: a subject doing a single clear action, observed by a described camera. If your prompt contains all three, most models will give you something usable. Add style, lighting, and mood as a second sentence, and you enter professional territory.

A fisherman hauls a net onto a weathered dock at dawn. Slow lateral tracking shot, camera skimming just above the waterline. Documentary style, cold blue morning light, sea mist.

Notice what is absent: no genre mashups, no ten adjectives, no "beautiful stunning masterpiece." Video models weight prompt tokens against motion coherence, and fluff dilutes the physics.

Framing: choose the shot before the model does

Framing vocabulary is the cheapest upgrade to output quality because it removes the model's first guess. Learn these six and you cover 90% of storytelling needs:

  • Extreme wide / establishing place the scene. Use for openings and transitions.
  • Wide shot subject full-body with environment. Establishes spatial relationships.
  • Medium shot subject waist-up. The default for dialogue and narration.
  • Close-up face or object detail. Where emotion lives.
  • Extreme close-up eyes, hands, texture. Use sparingly; it signals intensity.
  • Over-the-shoulder implies conversation or observation, a cinematic shorthand for relationship.

Name the shot explicitly in the first clause: "Medium shot of a baker..." beats "a baker in his bakery..." every time. Then stay consistent across clips in the same sequence cutting from an extreme close-up to an establishing shot reads as intentional; drifting framing reads as accident.

Camera motion: verbs the model understands

Camera directives are the heart of video prompting. Use one primary move per clip; two or more simultaneously ("dolly-in while crane up") produces wobble in most current models. The reliable vocabulary:

  • Static / locked-off no camera movement. Contrary to instinct, this is often the hardest to get right and the most elegant when it works: motion comes only from the subject.
  • Slow dolly-in camera moves toward the subject. Signals growing attention or tension.
  • Dolly-out retreat to reveal context. Perfect for endings.
  • Lateral tracking camera moves parallel to the subject. Essential for walk-and-talk energy.
  • Pan left/right camera rotates from a fixed point. Use to follow action or survey a scene.
  • Tilt up/down vertical reveal. A tilt up to a tall building is the classic power move.
  • Orbit / arc camera circles the subject. Expensive in artifacts; keep the subject centered and the arc under 90 degrees.
  • Handheld subtle organic shake. Signals documentary realism; combine with "slight" to avoid nausea.
  • Aerial drone shot high, sweeping movement. State the direction: "aerial drone shot pulling back over the coastline."

Pair the move with its speed adverb slow, gentle, steady are almost always correct. Fast motion multiplies artifacts exponentially, and most social clips benefit from deliberate slowness anyway.

Lighting directives that survive motion

Lighting is where video prompts most often fall apart, because a model that nails "golden hour" in a still may let that light drift incoherently across frames. The trick is to anchor lighting to a source rather than an abstract quality: "lantern light flickering across her face" gives the model a stable generator of illumination; "moody lighting" gives it nothing to hold onto.

Reliable lighting constructions:

  • Source-anchored: "neon signs reflecting in wet pavement," "candlelight trembling on the ceiling," "headlights sweeping across a wall."
  • Time-anchored: "golden hour backlight," "blue hour over the harbor," "harsh midday sun" time-of-day is stable physics.
  • Directional: "rim lighting from behind," "single hard light from the left" direction survives frames better than quality words.

One clause about light per prompt is enough. Two competing lighting descriptions ("warm sunset and cold fluorescent") rarely blend they flicker.

Negative prompts: the video edition

Negative prompts in video serve two masters: the usual image artifacts (extra fingers, warped faces, watermarks) and the motion-specific failure modes. Keep the list short eight to twelve tokens and prioritize:

morphing, melting, flickering, warping faces, extra limbs, distorted hands, text, watermark, logo, oversaturated, blurry
  • morphing / melting the signature video failure where features drift between frames.
  • flickering exposure or color pulsing across the clip.
  • warping faces identity instability; worth burning a negative token on for any human subject.
  • text, watermark, logo models love sneaking pseudo-text into cinematic frames.

A repeatable formula

Put it together and you get a four-line structure that works across nearly every video model Krilx exposes:

  1. Shot + subject + action (one verb): "Medium shot, a violinist raises her bow."
  2. Camera: "Slow dolly-in, locked focus on her hands."
  3. Style + lighting: "Concert documentary style, single spotlight, dust in the beam."
  4. Negative: the eight-to-twelve token list above.

Generate three variations, keep the seed of the best, and iterate on one clause at a time. Within five runs you will have a clip that looks directed rather than generated because it was.

The camera does not just show the story; it tells it. Prompt the camera, and the model follows your direction.

Ready to practice? Open Krilx's Create screen, pick Text to Video, and run this article's fisherman example verbatim then change one clause per run and watch the grammar of motion reveal itself.