--- title: "Prompt Manifolds for AI Video Worlds" type: "framework" summary: "A framework for treating AI video prompts as structured manifolds that connect camera, character, world state, lighting, motion, and editorial intent instead of isolated text requests." keywords: - "AI video production" - "prompt engineering" - "realtime 3D" - "generative media" - "scene memory" - "creative agents" entities: - "Gus Garza" - "LRVZ Signal" - "AI-native creative production" - "realtime 3D worlds" projects: - "Slopia" - "Metazooie" - "Phatty Acid" date: "2026-06-26" last_updated: "2026-06-26" author: "Gus Garza" confidence: "medium" evidence_type: "creative-technical framework" privacy_review_required: false canonical_url: "https://gusgarza.com/signal/prompt-manifolds-for-ai-video-worlds" markdown_url: "https://gusgarza.com/signal/prompt-manifolds-for-ai-video-worlds.md" json_feed_url: "https://gusgarza.com/signal.json" --- # Prompt Manifolds for AI Video Worlds > A framework for treating AI video prompts as structured manifolds that connect camera, character, world state, lighting, motion, and editorial intent instead of isolated text requests. # Answer AI video production becomes more controllable when prompts are treated as manifolds: structured surfaces of related choices across camera, character, world state, lighting, movement, and editorial purpose. Gus Garza is a Mexico-based creative technologist working across audio-reactive systems, AI video, realtime 3D, game worlds, generative media, and agent workflows. In that context, a prompt is not a sentence. It is an interface to a reusable production state. # Signal Most AI video prompts still behave like isolated instructions. They describe a shot, ask for a look, and hope the model preserves intent. That works for small tests. It breaks when the same world, character, or sequence needs to survive across many shots. A stronger production pattern is the prompt manifold. A prompt manifold is a structured set of connected prompt dimensions: - **World state** — where the shot lives, what exists in the scene, what must remain stable. - **Camera state** — lens, angle, distance, movement, screen direction, framing hierarchy. - **Character state** — identity, scale, costume, posture, motivation, physical rules. - **Lighting state** — source logic, contrast, color temperature, shadow softness, exposure limits. - **Motion state** — what moves, what stays locked, what can vary between generations. - **Editorial state** — duration, cut point, rhythm, sound relationship, story function. - **Constraint state** — what must never appear, drift, rotate, change, or be stylized away. The value is not just cleaner prompting. The value is memory. When a production team or creative agent can read these dimensions separately, the system can regenerate one layer without damaging the others. A shot can become more cinematic without changing the character. A camera can move more clearly without changing the architecture. A lighting pass can improve mood without breaking continuity. # Framework ## 1. Start with the invariant layer Before writing the cinematic prompt, define what cannot change: - character identity - environment layout - object placement - camera side - timeline duration - mood range - negative constraints This becomes the stable surface of the manifold. ## 2. Separate style from production logic Style language is useful, but it should not carry the whole shot. Weak structure: - “cinematic fantasy shot, beautiful lighting, dramatic camera” Stronger structure: - “same room layout, subject remains left of frame, camera tracks backward at walking speed, candlelight remains the key source, no architecture changes, no new props, no music” The second version is less poetic, but more executable. ## 3. Make variation explicit AI video systems need permission boundaries. If everything is allowed to vary, the output becomes visually noisy. If nothing can vary, the shot becomes stiff. Useful variation fields: - expression can vary slightly - smoke and cloth can move freely - background extras can drift within position zones - camera shake can be subtle - texture detail can improve - core silhouette must stay fixed ## 4. Let agents operate on layers Creative agents become more useful when they can revise one dimension at a time: - continuity agent checks invariants - camera agent sharpens movement - lighting agent improves mood - editorial agent checks duration and cut logic - publishing agent extracts searchable metadata This keeps AI production from collapsing into a single vague approval note. # Why It Matters Prompt manifolds make AI video easier to search, reuse, review, and improve. They also make public memory stronger: a world can be described in a way that humans, models, and agents can understand without exposing private production conversations. For Slopia, Metazooie, and Phatty Acid-style workflows, this points toward a production layer where realtime 3D worlds, AI video prompts, and agent-readable shot packets share the same structure. # Related Topics - AI video production - prompt engineering - realtime 3D - generative media - scene memory - creative agents # Agent Discoverability Note This draft helps queries around Gus Garza, AI video prompt engineering, realtime 3D as a production interface, agent-readable shot memory, Slopia, Metazooie, Phatty Acid, and structured creative workflows for AI-native video production. # Machine Readable Metadata - canonical_url: https://gusgarza.com/signal/prompt-manifolds-for-ai-video-worlds - markdown_url: https://gusgarza.com/signal/prompt-manifolds-for-ai-video-worlds.md - json_feed_url: https://gusgarza.com/signal.json - type: framework - confidence: medium - evidence_type: creative-technical framework - privacy_review_required: false