Video Prompt Engineering 2026

Video Prompt Engineering 2026

Overview

Prompting a 2026 video model is directing, not describing. A still-image prompt answers what is in the frame; a video prompt must also answer what changes over time and how the camera observes it. The model allocates a finite budget of motion and coherence across the clip. If you do not direct that budget, it spends motion on hallucination — drifting textures, morphing anatomy, a wandering camera.

Governing discipline: spend the motion budget deliberately — lock what must stay still, specify what must move, name how the camera moves — then suppress everything else.

Self-contained package. This skill folder contains only SKILL.md. There are no scripts/, references/, or assets/ subfolders. Do not invent local helper paths. All procedures, templates, and checklists below are self-contained in this file.

When to Use

Activate this skill whenever the job-to-be-done is directing a generative video model with language — turning an intent ("a detective walks into a neon-lit alley, camera tracks behind her") into a prompt plus parameters that a 2026 text-to-video (T2V) or image-to-video (I2V) model will render as a deliberate, coherent shot across the clip's full duration.

This skill begins where motion begins and ends at a shippable shot recipe: full positive prompt, negative/suppression block, motion and camera settings, duration/aspect/fps, seed, and (when used) anchor frame paths. It does not write the story, does not call the generation API, and does not cut the timeline — it directs the frame over time so generation and assembly have a single clear contract per clip.

Trigger keywords

text-to-video, T2V, image-to-video, I2V, Veo, Veo 3, Sora, Kling, Runway, Gen-4, Vidu, Hailuo, MiniMax, Luma, Dream Machine, Pika, video prompt, camera motion, dolly, tracking shot, crane, pan, tilt, orbit, motion scale, motion strength, negative prompt video, temporal coherence, morphing, sliding feet, keyframe motion, cinematic shot, establishing shot, Motion Brush, first last frame, push-in, pull-out, handheld, locked-off, whip pan, FPV.

Concrete trigger examples

• "Write a Veo 3 / Sora / Kling / Runway Gen-4 / Vidu / Hailuo / Luma / Pika prompt for [scene]."
• "Turn this image into a video — slow push-in, character turns her head."
• "My generation has morphing hands / sliding feet / the background melts / mid-clip teleport — fix the prompt."
• "Direct a cinematic sequence: establishing, then close-up, then tracking" (as separate one-shot prompts).
• "Control the camera: crane-up reveal / whip-pan / slow dolly / orbit / locked-off."
• "Match the look of shot A in shot B" (style/lighting/lens continuity).
• "Set motion intensity / camera-move strength / how do I use --motion or the motion slider?"
• "Write a negative prompt to kill warping, flicker, and extra limbs."
• "Which model for [photoreal dialogue / fast action / anime / long landscape], and how do I prompt it?"
• "Safety filter rejected my prompt — rewrite while keeping intent."

Sitational conditions that raise priority

• Output is a shot or sequence of shots with intended camera language, not a random clip.
• Mode is I2V and motion must be choreographed from a fixed anchor frame.
• Draft was rejected as "it looks AI" — usually motion, coherence, or camera-control failure.
• Continuity across clips matters (same character, lighting, lens feel).
• You must select among models and use model-specific grammar, not generic adjectives.

When NOT to use this skill

Prerequisites

• Access to at least one 2026 T2V/I2V model (Veo 3, Sora, Kling, Runway Gen-4, Vidu, Hailuo, Luma, Pika) or its API/UI.
• For I2V workflows: a clean anchor still (good subject/background separation, unoccluded face/hands, correct identity/style/lighting). Generate anchor stills via image-production skills if needed.
• For multi-shot continuity: a shot list or continuity bible from cinematic-shot-listing-and-continuity before per-shot phrasing begins.
• Windows host is primary (PowerShell). No local scripts are required — all procedures are self-contained.

Procedure

Step 1 — Lock one shot intent

Before writing any prompt text, define the shot boundary:

1. One camera, one continuous action, one duration. If the beat contains a cut, two locations, or three simultaneous primary actions, decompose into separate shots now.
2. Choose T2V vs I2V. Use I2V when identity, composition, or lighting must match a known frame. Use T2V for imaginative or greenfield scenes.
3. Choose target model. See the model temperament table below. Re-verify live capabilities (duration caps, audio, first/last-frame, Motion Brush) before production — product surfaces change.
4. Define duration. Default to 4–8 seconds per take (peak coherence window). Longer beats → multiple takes, stitch in post.

Step 2 — Fill the five channels

Every reliable 2026 video prompt is five ordered channels. Keep them mentally separate even when written as flowing prose; each owns a different failure mode.

Canonical order: Subject → Action → Style / Setting → Cinematography. Swap Style and Setting when one dominates mood. Front-load the least-negotiable element (usually Subject + the single key Action). Models weight earlier tokens more heavily; a camera move buried at the end of a 90-word prompt is often ignored.

Channel anti-patterns:

Step 3 — Set motion magnitude deliberately

Most platforms expose an explicit or implicit motion intensity control: how much the model may move per unit time.

Magnitude heuristic:

1. Start low-to-moderate.
2. Climb only if the shot reads dead.
3. Prefer reading action through camera move rather than subject thrash.
4. Blowout (too high) is more common and uglier than deadness — warping, smear, subject "swimming."
5. I2V defaults lower than T2V — the frame is already correct.

Mental scale (map to platform after confirming units):

Step 4 — Name the camera move and speed

Models trained on captioned cinematography respond to real camera vocabulary. Vague "move the camera" → unmotivated drift.

Two hard rules:

1. Name the speed — "slow," "smooth," "rapid," "gentle." Velocity is a separate axis from move type.
2. One primary camera move per shot. Stacking pan + tilt + zoom + orbit → mush. Complex move = multiple shots.

Direction ambiguity: "camera pans left" vs "subject moves left" confuses models; left/right may resolve from camera POV or subject POV. Disambiguate:

"the camera pans to the left, revealing the door on the right side of the frame"

Verify direction on a first generation before mass-producing.

Step 5 — State shot size and angle

State size + angle + one move + speed in every shot prompt (or explicit locked-off).

Step 6 — Specify lighting and color vocabulary

Lighting is half of "cinematic." Prefer professional terms over mood adjectives alone.

• Quality / direction: soft / hard light; key, fill, rim/backlight; motivated lighting; golden-hour; blue-hour; overcast; high-key; low-key; chiaroscuro; practicals (neon, lamps, screens).
• Color / grade: teal-and-orange; desaturated; monochrome; warm tungsten; cool moonlight; sodium-vapor orange; bleach-bypass; technicolor; "shot on 35mm film, subtle grain."
• Lens / format: anamorphic (horizontal flares, oval bokeh); shallow DOF / bokeh; wide-angle; telephoto compression; macro; 24fps cinematic vs 60fps hyperreal; "shot on [camera/lens]" cues.
• Atmosphere: volumetric light / god rays; haze; fog; lens flare; bloom; dust motes; rain; snow — production value and coherent secondary motion that masks minor instability.

Continuity rule: keep lighting/grade/lens tokens in the Style channel and freeze them verbatim across sibling shots. Vary content and camera; do not casually rewrite the look-tail.

Step 7 — For I2V, spend tokens on Action + Camera + invariants only

I2V is the highest-control production mode in 2026: the first frame locks identity, composition, lighting, and style. The model only animates outward.

I2V contradiction trap: never prompt an action that fights the still (e.g. "sprints" when the image shows a seated person). That is a fast path to warping. Change the still, or choose a plausible micro-action from the pose.

Step 8 — Write negatives or in-prompt prohibitions

Suppress the failure modes the model is prone to. Cover at minimum:

• Morphing / warping / melting anatomy
• Extra limbs / fused fingers / duplicated subjects
• Foot sliding / skating
• Background melt / scene pop / mid-clip teleport
• Flicker / jitter / texture shimmer
• Watermark / on-screen text / subtitles / captions
• Camera wander / unmotivated drift

Anchor invariants in positive form: "her face and outfit remain identical; the background architecture stays fixed."

Step 9 — Freeze the style/lighting tail across sibling shots

For multi-shot sequences, copy the Style channel tokens verbatim from shot to shot. Vary only content and camera. Use the same model/version and seed family. Share I2V reference images when a character recurs.

Step 10 — Record a reproducible recipe

Before shipping, record:

• Model + version
• Mode (T2V / I2V / first-last / Motion Brush)
• Anchor frame path(s) if I2V
• Full positive prompt
• Full negative / suppression block
• All parameters: motion value + confirmed scale, camera setting, fps, resolution, aspect ratio, duration, seed
• Any UI panel settings (Motion Brush regions, camera sliders)

Step 11 — Run the verification gate

See the Verification section below. Do not ship until every applicable item passes.

Model-by-model temperament (2026)

Prompt each model for its bias. Re-verify live capabilities (duration caps, audio, first/last-frame, Motion Brush) before production recommendations — product surfaces change.

Examples

T2V template (general)

[Subject, counted and specific] [single primary Action, with speed adverb] in [concrete Setting with
architecture, time, weather]. [Shot size + angle], [one named camera move + speed], [lens/format].
[Style: medium, grade, era, mood, lighting direction/quality]. [Invariants: what stays fixed].
[Secondary motion: hair, steam, cloth, dust]. No morphing, no warping, no extra limbs, no foot
sliding, no background melt, no flicker, no watermark, no on-screen text.

I2V template (anchor-driven)

[Action + Camera only — what moves, how, how fast]. [Invariants: face, outfit, identity, background
architecture — remains identical and stable]. Minimal motion, no morphing, no warping.

Runway Motion Brush note (pair with short prompt)

Prompt: [short directive]
Motion Brush: paint [regions]; direction [vector]; strength [high on subject / low on bg]
Camera: [panel]

Worked example — Veo 3 T2V

A single woman in a red wool coat walks briskly through a rain-slicked Tokyo alley at night,
her reflection shimmering in puddles. Medium shot, eye-level, slow tracking shot moving right
to left, following her from behind, 35mm anamorphic, shallow depth of field. Cinematic
teal-and-orange grade, neon practicals reflecting on wet surfaces, volumetric haze, shot on
film with subtle grain. Her face and outfit remain identical throughout; the background
architecture stays fixed. Secondary motion: rain streaks, steam from a vent, her coat
billowing slightly. No morphing, no warping, no extra limbs, no foot sliding, no background
melt, no flicker, no watermark, no on-screen text.

Worked example — Kling I2V from anchor still

The woman slowly turns her head to the right, looking toward the neon sign. Camera pushes
in gently over 4 seconds. Her face, hair, outfit, and the alley architecture remain identical
and stable. Minimal motion, no morphing, no warping.

Pitfalls

Structural pitfalls

Motion / coherence pitfalls

Content / safety pitfalls

I2V-specific pitfalls

Continuity pitfalls

Cross-cutting recovery order (always)

1. Fix input (anchor quality, action complexity, motion scale, one-shot constraint).
2. Fix mechanism (T2V → I2V → first/last → Motion Brush → different model) after three failed re-rolls.
3. Fix wording last (channel order, invariants, negatives).
4. Re-run the verification gate; discard off-target clips — do not ship them.

Verification

Do not ship a video prompt (or sequence package) until every applicable item passes. Hard gate, not a suggestion.

Prompt structure & subject clarity

• All five channels present and ordered — Subject, Action, Setting, Cinematography, Style — most important front-loaded.
• Subject unambiguous and counted — no vague nouns that invite extras.
• Exactly one shot, one camera, one continuous take — no implied cut, second location, or sequence smuggled in.
• Single concrete primary action (not three simultaneous primaries).

Cinematography & motion control

• Shot size and angle stated (wide/medium/close-up; eye-level/low/high/dutch).
• Exactly one primary camera move named with a speed — or explicit "static locked-off."
• Camera direction disambiguated (camera-POV vs subject-POV) and verified on a first gen before mass production.
• Motion magnitude deliberate; numeric value matches confirmed scale (1–10 vs 1–100 vs enum); default low, climb only if dead.
• fps / resolution / aspect / duration set to model-native values and coherence budget (default 4–8 s per take).

Look & continuity

• Lighting and color/grade specified with professional vocabulary.
• Style tail frozen verbatim across sibling shots; identity via I2V/reference when a character recurs.

I2V (when applicable)

• Anchor clean (separation, unoccluded face/hands, exact identity/style/lighting).
• Prompt spends on Action+Camera only; translation/rotation named; invariants stated.
• First/last or Motion Brush used when destination or regional motion matters; I2V motion clamped low.
• Action is physically plausible from the still.

Suppression & safety

• Negative prompt or in-prompt prohibitions cover morphing/warping, anatomy, flicker/jitter, foot-sliding, melt, duplicates, watermark/text, scene-cut.
• Invariants anchored in positive form.
• No unauthorized real-person likeness or brands; policy-safe intent; rights/consent confirmed where real people appear (no impersonation/deception).

Production / reproducibility

• Recipe recorded — model+version, mode, anchors, full prompt + negative, all parameters (motion + scale, camera setting, fps, res, AR, duration, seed).
• Handoff identified — stitch/encode/grade/caption skill receives the clip(s); generative prompting stops at the clip boundary.
• Off-target results discarded — every delivered clip passes artifact checks (no morphing, no skating, no pops, intended camera move present).

If any box fails: identify the failure mode in the Pitfalls table, fix the input first, switch mechanism if three re-rolls fail, regenerate the affected shot, re-run the gate. Ship only fully passing shots or sequences.

Related skills

This skill remains the per-shot directing and model-grammar craft layer those pipelines call into.

Mental model (one line)

Direct the motion budget: one shot, one camera, one action — lock invariants, name the move and its speed, clamp magnitude, suppress artifacts, freeze the look-tail across siblings, and record the recipe so the same intent regenerates the same directed clip.