--- title: "How to Write AI Video Prompts: The Complete Guide" description: "Learn how to write effective AI video prompts using a 6-part framework for better shots, consistent visuals, camera control, and campaign-ready results." canonical: "https://lumalabs.ai/news/write-ai-video-prompts" source: "https://lumalabs.ai/news/write-ai-video-prompts.md" --- # How to Write AI Video Prompts: The Complete Guide _By Luma team ยท August 13, 2026_ The difference between a generic clip and a campaign-ready video comes down to how you write the prompt. Structured prompts give AI video models clearer direction on subjects, motion, framing, lighting, and style, reducing the amount of creative interpretation left to the model. That can make iterations more predictable and help teams reach usable footage with fewer unnecessary regenerations. This guide breaks down the framework creative directors use to write prompts that deliver campaign-ready footage from the first brief through final approval. ## **Key Takeaways** - **The 6-part prompt framework** (Subject + Action + Setting + Camera + Lighting + Style) produces consistent results across all [AI video](https://lumalabs.ai/create/ai-video-generator) platforms - **Front-load the first 20-30 words** with your most important details. AI models weight these heavily - **Use only ONE camera movement** per clip to avoid unpredictable results - **Never ask AI to render readable text or logos**. Composite these in post-production - **A 5-10-1 iteration workflow** can keep experimentation costs down: explore several inexpensive variations, refine the strongest direction, then reserve premium generation settings for the final render - **Adding specific style references** like "shot on 35mm film" or "Blade Runner 2049 color grading" eliminates the generic AI look - **Write prompts like camera notes: **shot type, subject, action, technical specs [Try Luma Now](https://lumalabs.ai/ https://auth.lumalabs.ai/sign-up) ## **Understanding the Basics of AI Video Prompts** A prompt is a set of instructions that tells the AI what to generate. Unlike traditional production where you direct talent and position cameras, prompt writing requires translating your creative vision into language the model understands. **Strong prompts share common elements:** - **Specific subject descriptions** rather than generic categories - **Active verbs** describing movement and action - **Technical camera language** the model recognizes - **Lighting descriptions** that create atmosphere - **Style anchors** that establish visual identity Weak prompts leave decisions to the AI. Strong prompts make decisions for it. The basic shift: you're not describing what you want to see. You're describing what should be in the frame, how the camera captures it, and what feeling the light creates. ### **What Separates Pro Prompts from Basic Attempts** - **Amateur prompt: **"A woman walking in a garden." - **Professional prompt: **"Medium tracking shot of a woman in a flowing red dress walking through a sunlit Victorian garden, camera following from the side, 35mm lens, golden hour lighting, shallow depth of field, gentle camera movement, muted color grading." The amateur prompt gives the AI too many decisions. Shot type? The AI guesses. Camera movement? Random. Lighting? Whatever the model defaults to. Style? Generic. The professional prompt specifies every creative choice that matters. The AI executes rather than improvises. ## **The 6-Part Prompt Framework** This structure works across every major [text to video](https://lumalabs.ai/create/ai-video-generator-from-text) platform. Master it once, apply it everywhere: ### **1. Subject + Description** Who or what is the focus? Include specific details like clothing, colors, materials, and distinctive features. ### **2. Action/Movement** What happens in the frame? Use active verbs: walking, turning, rising, falling, flowing. ### **3. Setting/Scene** Where does this take place? Include time of day, weather, environment details. ### **4. Camera Movement** How does the camera behave? Options include: static, slow push-in, pull-back, dolly, tracking, orbit, crane. ### **5. Lighting** What's the light source and quality? Golden hour, soft window light, dramatic side lighting, neon glow. ### **6. Style/Aesthetic** What's the visual identity? Film stock references, color grading, aspect ratio, genre conventions. ## **Putting the Framework to Work** Here's a starter prompt for a product video: "Slow push-in on a ceramic coffee mug on a wooden table, steam rising from hot coffee, morning sunlight streaming through window from the left, warm golden lighting, shallow depth of field, soft focus background, cozy morning atmosphere, shot on 35mm film." **Each element serves a purpose:** - **Subject**: ceramic coffee mug on wooden table - **Action**: steam rising - **Setting**: morning, implied kitchen - **Camera**: slow push-in - **Lighting**: morning sunlight from left, warm golden - **Style**: 35mm film, shallow depth of field Your first usable take arrives faster than setting up a product shot. ## **Advanced Prompt Techniques for Better AI Video** ### **Learning Camera Language** AI models recognize professional camera vocabulary. Learn these terms and your prompts produce predictable results: #### **Shot Types:** - Extreme close-up (ECU): fills frame with detail - Close-up (CU): face or single object - Medium shot: waist up or product in context - Wide shot: full figure or scene establishment - Extreme wide: landscape or architectural scale #### **Camera Movements:** - Static: locked off, no movement - Push-in: camera moves toward subject - Pull-back: camera retreats from subject - Dolly: camera moves parallel to subject - Tracking: camera follows moving subject - Orbit: camera circles around subject - Crane: vertical camera movement **Key Rule:** Use only ONE camera movement per generation. "Slow push-in then pan left" confuses the model. Split complex movements into separate shots and cut them together in post. ### **Directing Motion and Transitions** For multi-shot sequences, think in terms of controlling what happens between frames. Many platforms let you specify start and end frames, giving you control over what happens between. **Describe transitions naturally:** - "Camera begins tight on hands, slowly reveals full figure" - "Opens on empty room, woman enters from frame right" - "Starts in shadow, light gradually illuminates subject" These descriptions guide the AI through the shot progression rather than leaving the movement to chance. ### **Getting Rid of the Generic AI Look** Generic AI footage has a telltale quality: too smooth, too clean, no character. Combat this with specific style anchors: #### **Film Stock References:** - "Shot on Kodak Vision3 500T" - "16mm documentary grain" - "IMAX large format" #### **Color Grading References:** - "Blade Runner 2049 teal and orange" - "Wes Anderson pastel palette" - "70s Kodachrome warmth" #### **Lighting Recipes:** - "Soft window light from camera-left, warm tone" - "Harsh overhead fluorescent" - "Practical neon signage as key light" Stack 2-3 references for distinctive looks: combining composition from one film, atmosphere from another, and camera movement from a third creates original work that doesn't feel generic. ## **Getting Consistency Across Campaign Assets** A single hero video rarely ends a project. The brief calls for social cutdowns, localized versions, product swaps, and last-minute copy changes. Consistency across dozens of assets separates professional campaigns from disjointed collections. ### **Keeping Your Brand Look Consistent** Build prompt templates that lock in brand elements: #### **Brand Template Example:** "[SHOT TYPE] of [SUBJECT] in [SETTING], [BRAND CAMERA STYLE: handheld with gentle movement], [BRAND LIGHTING: warm natural light], [BRAND COLOR: muted earth tones with selective color pops], shot on 35mm film." Swap the variables. Keep the constants. Every asset looks like it belongs to the same campaign. ### **Using Reference Images** [Image-to-video](https://lumalabs.ai/create/ai-video-generator-from-image) workflows let you start from approved stills. Upload your hero product shot, and the prompt describes what happens to it: "Product slowly rotates, revealing all angles, soft studio lighting, minimal camera movement, premium commercial aesthetic." The AI inherits the lighting, color, and composition from your reference. The prompt directs the motion. This approach transforms approved creative into working footage without rebuilding every element. The client-approved product shot becomes every regional campaign. Change the background. Keep the product exactly as approved. ### **Character and Subject Consistency** Multi-shot narratives require the same character across different scenes. Most platforms offer reference features that maintain appearance: - Upload 3-4 images of the same subject from different angles - Include character description in every prompt - Generate all shots in one session when possible When perfect consistency matters more than generation speed, consider generating shots that feature hands, faces, or distinctive features in tight compositions where variation shows less. ## **The Prompt Library Approach** Professional teams don't write prompts from scratch for every generation. They build libraries organized by: - **Client/Brand**: Templates with brand-specific style constants - **Content Type**: Product demos, lifestyle, testimonials, B-roll - **Platform**: 16:9 for YouTube, 9:16 for TikTok/Reels, 1:1 for Instagram - **Shot Category**: Establishing, product, transition, reaction Build 50-100 tested prompts over your first month. When the next brief arrives, combine existing templates rather than inventing new language. ### **Smart Iteration Workflow** A smart iteration workflow helps you find the best creative direction without wasting generations: #### **Step 1: Quick Exploration** Generate several initial attempts at lower quality settings to explore directions. #### **Step 2: Refinement** Take the most promising direction and generate variations with refined prompts. #### **Step 3: Final Render** Use your optimized prompt for a final [high-quality generation](https://lumalabs.ai/create/ai-video-generator-from-text). This process mirrors traditional creative development: rough concepts, selected refinement, polished final, at compressed timelines. ## **Using AI Video Prompts for Product Launches** ### **Making Product Demo Videos** E-commerce teams face a familiar challenge: hundreds of SKUs need video, and the production budget doesn't cover everything. #### **The prompt template for product demos:** "Slow 360-degree orbit around [PRODUCT NAME] on [SURFACE], soft studio lighting from above, shallow depth of field, premium commercial aesthetic, product details in focus, clean background, 5 seconds." Starting from product photography, teams generate demos that show the product in motion. Pages with video see higher conversion than static images alone. ### **Creating Campaign Variants** One hero video rarely covers the full media plan. The brief calls for: - 60-second hero cut - 30-second broadcast version - 15-second pre-roll - 6-second bumper - Vertical social versions - Localized versions for regional markets With prompt libraries established, each variant starts from the approved creative rather than from scratch. The same product shot [reframes for vertical](https://lumalabs.ai/reframe). The same scene regenerates with different background elements for regional relevance. The headline changes. Everything else stays put. ## **Fixing Common Prompt Problems** ### **Distorted Hands and Faces** Hands remain the most common failure point. **Solutions:** - Frame shots to minimize hand visibility when hands aren't essential - Add to your prompt: "natural hand positioning, anatomically correct fingers" - Generate 3-5 variations and select the best result - Use tighter compositions where hand detail matters less - When hands must be prominent, use image-to-video starting from a photograph with clearly visible hands ### **Controlling Camera Movement** When the camera does something unexpected: - Verify you're using only ONE camera move per prompt - Add "static camera" to lock position when movement causes problems - Be more specific: "gentle, steady push-in over 5 seconds" rather than just "push-in" - For complex movements, generate separate shots and cut together ### **Avoiding the Generic Look** Videos that look "too AI" lack the imperfections and character of real footage: - Add specific style references: "Shot on 35mm Kodak film, slight grain" - Include detailed lighting descriptions rather than generic "good lighting" - Reference specific films for color and composition - Add environmental details: "dust particles visible in light beam" ### **Why Text and Logos Fail** AI models cannot reliably render readable text. This is a basic limitation, not a prompt problem. **The solution:** Never include text generation in your prompts. Plan for text overlay in post-production. This takes 30 seconds in any video editor and produces perfect results every time. ## **How Luma helps:** A campaign brief becomes product videos, social cutdowns, and localized ads without rebuilding each version. - [Ray 3.2](https://lumalabs.ai/ray) turns that brief into production-ready footage with multi-frame control, so the director decides what happens frame by frame rather than hoping the AI interprets the prompt correctly, and the export lands in Premiere as an EXR file ready for color grading - With [Layers](https://lumalabs.ai/news/introducing-layers), the product changes while the lighting, background, and approved composition stay put. Regional versions launch with the same prompt library that built the hero campaign, producing Spanish, French, and Japanese variants without starting over - When [Ray 3.2's controls](https://lumalabs.ai/learning-center/articles/ray-3-2-prompting-outputs-and-controls) match your creative vocabulary, the brief becomes the campaign becomes the delivery, without regeneration cycles The prompt gets you most of the way there. [Luma](https://lumalabs.ai/) helps you finish without rebuilding from zero every time the client has notes. [Try Luma Now](https://auth.lumalabs.ai/sign-up) ## **Frequently Asked Questions** ### **What is an AI video prompt and why does structure matter?** An AI video prompt is the text instruction that tells a generative model what to create. Structure matters because vague prompts force the AI to make creative decisions (shot type, lighting, camera movement) that should belong to the director. The 6-part framework (Subject + Action + Setting + Camera + Lighting + Style) ensures you maintain creative control over every element that affects the final look. ### **How long should an AI video prompt be?** Good prompts typically run 100-150 words. Shorter prompts lack necessary detail; longer prompts dilute the important information. Front-load your most critical details in the first 20-30 words, as most models weight early prompt content more heavily. ### **Can AI video prompts include specific brand elements?** Yes, but with limits. Brand colors, general aesthetic direction, and approved style references work well. Specific logos and readable text do not. AI models cannot reliably render these. Plan to composite logos and text in post-production. For character consistency, upload reference images of talent rather than describing them in text. ### **How do I make AI-generated video look less artificial?** Add specific style anchors that introduce character and imperfection. Reference actual film stocks ("shot on Kodak Vision3 500T"), include detailed lighting descriptions ("soft window light from camera-left with warm tone"), and cite specific films for color grading reference. The goal is giving the AI a specific visual target rather than a generic aesthetic. ### **What's a good way to iterate on AI video prompts?** Generate several quick variations to explore directions, then take the most promising result and create variations with refined prompts, and finally use your optimized prompt for a final high-quality generation. This mirrors traditional creative development (rough concepts, selected refinement, polished final) at compressed timelines.