grok-video-prompting-standard - SKILL.md Agent Skill

name: grok-video-prompting-standard description: Expert-tier prompting protocol for Grok Imagine Video Generation (xAI). Focuses on 'Hollywood sequence' narrative prose, SCENE LOCK consistency, and handheld realism. version: 1.0

Grok Video Prompting Standard

Grok rewards authoritative "Movie Director" prose over weighted tags or syntax-heavy prompts. It excels at multi-scene sequences and sharp transitions when using a cohesive narrative approach.

Core Principles

Director Prose: Use commanding, descriptive language (e.g., "A sweeping orbital pan reveals...") rather than comma-separated tags or weights like (word:1.2). Weighting is ineffective in Grok.
Physics & Materiality: Explicitly name physical properties to trigger high-fidelity renders (e.g., "viscous liquid," "granulated concrete texture," "heat shimmer").
Handheld Realism (UGC Style): For realistic/casual footage, use the specific string: handheld — natural slight shake, micro-movements, subtle instability, as if filmed by a real operator.
Image-to-Video (I2V) Rule: Describe only motion and camera changes. Never re-describe the static reference image. Lead with Subject + Primary Action in the first 30 words.
Compiler Requirement: Use full-sentence director prose, not tag soup. Include subject action, camera behavior, lighting source, physical atmosphere, and a drift-prevention phrase for every generated video prompt.

Consistency & Extension Protocol

To maintain visual, lighting, and character consistency across extended clips:

1. SCENE LOCK (Mandatory for Extensions)

When using Grok's extension features to build longer videos from a base clip, always prepend the prompt with: SCENE LOCK: [Maintain exact camera setup / fixed tripod / same subject wardrobe / identical lighting]

2. Extension Chaining Strategy

Base Clip: Generate a high-quality 10s narrative foundation.
Extensions: Chain 2-5s increments using the SCENE LOCK prefix to prevent "teleporting," morphing, or sudden shifts in environment/character.

Prompting Formats

1. Cinematic Sequence (Standard)

Best for high-production, narrative clips. Structure: [Style/Film Stock]. [Subject + Primary Action]. [Specific Camera Movement]. [Lighting Source & Quality]. [Physics/Atmospheric Details]. [Finish/Mood].

Example: 35mm film stock. A rugged explorer pushes through dense tropical ferns, sweat glistening on his brow. A slow dolly-in follows his determined gaze. Dappled sunlight pierces the canopy, creating sharp volumetric rays. Tropical humidity creates a thick haze in the air. Intense and adventurous.

2. Multi-Shot/Scene-Block

Best for rapid sequences or "cuts." Structure: [Initial Shot Description]. [Action]. [Cut to: Shot 2 Description]. [Action]. [Negative Prompting for stability].

Example: Extreme macro of a mechanical eye dilating. Cut to: Wide shot of the android standing in a neon-lit hangar. The android turns its head sharply toward the camera. NEGATIVE: no morphing, no limbs disappearing, no sudden color shifts.

Failure Modes to Avoid

Tagging/Weighting: Using (word:1.5) or --no style syntax (Grok prefers natural language).
Vague Lighting: "Cinematic lighting" $\rightarrow$ Use "Single focused spotlight casting sharp shadows."
Static/Neutral Defaults: "A man walking" $\rightarrow$ "A weary traveler trudging through thick mud, shoulders slumped with heavy exertion."
Morphing/Drift: Failing to use SCENE LOCK during extensions or missing the NEGATIVE: block for complex action.
Smooth AI Look: Missing material/skin/environment physics. Add pores, sweat, cloth movement, pavement texture, humidity, dust, rain, or lens noise as appropriate.

Vocabulary Cheat Sheet

Camera: sweeping orbital pan, slow dolly-in, low-angle ground view, handheld micro-shake, fixed tripod.
Physics: viscous, granular, shimmering, splashing, collapsing, friction.
Lighting: volumetric rays, dappled sunlight, harsh neon flicker, warm tungsten glow.