For visual creators, the best workflow is the one that preserves intent while making each decision easy to inspect and revise. This edition emphasizes those practical controls. A reference video is a sequence of production choices, not one giant prompt; preserving those choices shot by shot is what makes the result reusable.

Disclosure: I work with PixMind.

A reference video is rarely one prompt. It is a sequence of shots, and each shot has its own subject, action, camera movement, lighting, sound, and transition. This guide shows how to turn that sequence into structured, reusable AI video prompts without flattening the whole clip into a vague summary.

You will learn a practical video prompt reverse-engineering workflow, a timestamped output schema, six scene breakdowns, and a checklist for converting the result into Veo, Runway, Seedance, or another target model.

Want the first draft automatically? Upload a clip to PixMind Video to Prompt, then use this guide to inspect and refine the shot list.

Part 1: Core Concepts & Parameter Cheat Sheet

What Is Video Prompt Reverse-Engineering?

Video prompt reverse-engineering means analyzing a clip shot by shot and converting its visible and audible choices into structured production language. Instead of guessing the original creator's prompt, you document what is actually present: subject, action, environment, framing, camera motion, lighting, color, sound, and transitions.

The result is not a forensic copy of the source. It is an editable creative specification you can reuse with a different subject, product, location, or target video model.

The Five-Layer Prompt Structure

Every strong video prompt is built from five stacked layers:

Layer What It Captures Example Descriptors
| Subject | Who or what is in the frame | "a woman in a white linen dress"
| Action / Motion | What's moving and how | "walking slowly through tall grass"
| Environment | Location, time of day, weather | "golden-hour meadow, soft wind"
| Camera | Shot size, movement, lens character | "wide tracking shot, slight lens flare"
| Style / Mood | Aesthetic direction, color grade, tone | "cinematic, warm tones, film grain"

Core Parameter Cheat Sheet

Parameter Starter Default Advanced Options
| Shot size | Medium shot | Extreme close-up / aerial
| Motion speed | Normal speed | Slow motion / time-lapse
| Lighting | Natural daylight | Golden hour / neon backlight
| Color grade | Neutral | Teal-orange / desaturated
| Camera movement | Static | Dolly / handheld shake
| Duration hint | 5–8 seconds | 15–30 seconds
| Audio hint | None | Ambient sound / dialogue
| Aspect ratio | 16:9 | 9:16 (vertical) / 1:1

Core principle: Start with the details that materially change the shot, then add one control at a time. A short, internally consistent prompt is more useful than a long prompt containing competing camera, lighting, or action instructions.

A Shot-by-Shot Video Prompt Output Schema

For multi-shot clips, create one record per shot instead of one paragraph for the whole video:

Field What to Record Example
| Timecode | Start and end of the shot | 00:04–00:07
| Subject | Visible person, object, or product | Runner in a red windbreaker
| Action | Subject and environmental motion | Runner turns; rain blows left to right
| Camera | Shot size, angle, and movement | Low-angle medium shot, handheld tracking
| Lighting and color | Source, direction, contrast, palette | Cool overcast key, muted blue shadows
| Audio | Dialogue, ambience, effects, music | Footsteps, rain, low bass pulse
| Transition | How the next shot begins | Whip-pan cut on movement

This schema directly addresses scene-by-scene extraction queries such as “AI prompt from clip to each scene.” It also makes errors easy to spot: if a tool labels a locked shot as a dolly, you can correct one field without rewriting the entire prompt.

Part 2: Scene Walkthrough — Cinematic Nature Landscape

Goal

You find a travel documentary clip: a mist-covered mountain range at dawn, a slow aerial pull-back, no people. You want to recreate that atmosphere.

✅ Reference Visual Example

(Illustrative example — not an actual PixMind model output.)

Recommended Prompt Template

Aerial drone shot slowly pulling back from a mist-covered mountain ridge at dawn.
Pale blue and soft orange light filters through low clouds. Pine trees visible below.
No people. Cinematic color grade, anamorphic lens flare, 4K quality.
Mood: serene, vast, slightly melancholic.
Duration: ~8 seconds.

Step-by-Step

1. Mute the audio and watch the clip three times. Focus only on what the camera is doing — ignore the subject content for now.
2. Identify the dominant direction of movement. In this case: upward + backward (pull-back).
3. Name the light source and its quality. Dawn = low angle, diffused, warm-to-cool gradient.
4. Compress the style into 2–3 adjectives. "Cinematic, serene, vast."
5. Paste your notes into PixMind's Video-to-Prompt tool to auto-extract a draft, then refine it manually.

⚠️ Common Mistake

Don't write the entire scene description as one continuous run-on sentence. Models like Veo 3 perform noticeably better when subject, motion, and style are separated with line breaks or commas — burying everything in a single paragraph degrades output quality.

Part 3: Scene Walkthrough — Product Advertisement

Goal

A 10-second beauty ad: a product on a marble surface, slow push-in, soft studio lighting, pastel background. You want to distill a reusable product video template.

✅ Reference Visual Example

(Illustrative example — not an actual PixMind model output.)

Recommended Prompt Template

Close-up slow zoom-in on a [product name] bottle placed on white marble surface.
Soft diffused studio lighting from upper-left. Pastel pink background, out of focus.
Subtle water droplets on the product. No hands, no people.
Elegant, minimal, luxury aesthetic. Smooth camera movement, no shake.
5-second clip, 16:9.

Step-by-Step

1. Pause on three key frames in the reference ad: the opening, the midpoint, and the end.
2. Observe what changes between frames. Here, only the camera distance changes (push-in) — the background and lighting remain constant throughout.
3. Identify brand-specific elements and swap them out. Replace the product name and color palette with your own.
4. Test with PixMind's Seedance 2 AI video generator — its product ad scene handling is specifically optimized for this kind of controlled studio aesthetic.

⚠️ Common Mistake

Avoid vague luxury descriptors like "high-end" or "premium." Instead, describe the visual evidence of luxury: marble, soft shadows, minimal composition, slow movement. Models respond to concrete visual signals, not abstract quality labels.

Part 4: Scene Walkthrough — Urban Street Scene with People

Goal

A street photography-style video: a busy intersection at night, handheld camera, neon lights reflecting off wet pavement, pedestrians moving quickly through the frame.

✅ Reference Visual Example

(Illustrative example — not an actual PixMind model output.)

Recommended Prompt Template

Handheld medium shot of a busy city intersection at night, wet pavement reflecting
neon signs in red, blue, and yellow. Crowds of people walking quickly in multiple
directions. Slight motion blur on pedestrians. Shallow depth of field.
Urban, gritty, high-contrast. Tokyo or New York aesthetic.
Camera: slight sway, no stabilization. 6–8 seconds.

Step-by-Step

1. List every light source in the scene. Neon signs, headlights, shop windows — each one is a descriptor.
2. Describe crowd density. "Sparse," "moderate," or "dense" gives the model a population signal.
3. Capture the camera's personality. Handheld means intentional imperfection — explicitly write "no stabilization" or "slight camera sway."
4. Use PixMind's Text-to-Prompt tool to generate a base draft from your scene notes, then layer in camera descriptors manually.

⚠️ Common Mistake

Naming a specific real-world location (e.g., "Shibuya Crossing") helps establish an aesthetic reference, but it can't replace visual description. Models may ignore the place name and render a generic street. Always describe what you see, not just where it is.

Part 5: Scene Walkthrough — Emotional Close-Up Portrait

Goal

A documentary-style close-up: an elderly person's face, natural window light, slow push-in, no dialogue, contemplative atmosphere.

✅ Reference Visual Example

(Illustrative example — not an actual PixMind model output.)

Recommended Prompt Template

Slow push-in close-up of an elderly person's face, mid-60s, neutral expression,
thoughtful and calm. Natural soft light from a window on the left side.
Slight skin texture visible. Background: blurred warm interior.
Documentary style, desaturated color grade, no music cue.
Camera: very slow dolly-in, ultra-stable. 8 seconds.

Step-by-Step

1. Identify the emotional anchor. What single word captures the mood? Here it's "contemplative." Write it explicitly.
2. Describe the direction and quality of light. "Window light from the left" gives the model far more to work with than "natural light."
3. Note what is deliberately absent. No smile, no action, no dialogue — negative descriptors are just as effective as positive ones.
4. Test on Veo — based on announced specs, Veo 3/3.1 handles realistic human close-ups with strong lighting fidelity particularly well.

⚠️ Common Mistake

Never specify a real person's face or likeness in a prompt. Describe demographic and emotional characteristics instead. This keeps your prompt within model usage guidelines and produces more consistent results across multiple generations.

Part 6: Scene Walkthrough — Action / Sports Footage

Goal

A surfing video: aerial perspective, athlete riding a massive wave, slow motion, spray sparkling in sunlight, high energy.

✅ Reference Visual Example

(Illustrative example — not an actual PixMind model output.)

Recommended Prompt Template

Aerial shot looking down at a surfer riding a large breaking wave, slow motion.
White water spray exploding upward, backlit by bright midday sun — sparkle effect.
Ocean: deep blue-green. Surfer: small relative to the wave.
High energy, dynamic composition. Camera: hovering drone angle, slight tilt.
Slow motion at 50% speed. 6 seconds, 16:9.

Step-by-Step

1. Establish scale relationships. "Surfer appears small relative to the wave" gives the model a compositional anchor point.
2. Describe how the light interacts with the scene. Backlit spray creates a sparkle/halo effect — call it out explicitly.
3. Specify slow motion with a number. "50% speed" or "120fps playback" is clearer than just writing "slow motion."
4. Check the Seedance 2 prompt page for motion-specific prompt structures optimized for high-action scenes.

⚠️ Common Mistake

High-action scenes are where models produce the most artifacts — distorted limbs, incorrect water physics. Adding constraints like "physically realistic water motion" or "no distortion" measurably reduces the likelihood of these issues.

Part 7: Scene Walkthrough — Brand Story / Narrative Short

Goal

A 15-second brand film: a craftsperson's hands shaping clay on a pottery wheel, warm workshop lighting, close-up detail, slow cuts between shots.

✅ Reference Visual Example

(Illustrative example — not an actual PixMind model output.)

Recommended Prompt Template

Close-up of weathered hands shaping wet clay on a pottery wheel, slow deliberate
movement. Warm tungsten workshop light, dust particles visible in the air.
Background: blurred wooden shelves with ceramic pieces.
Tactile, artisanal, warm color grade. Camera: slow macro push-in on hands.
Ambient sound: soft spinning wheel, no music. 10–12 seconds.

Step-by-Step

1. Lead with texture descriptors. "Weathered hands," "wet clay," "wooden shelves" — tactile language activates richer visual rendering.
2. Add ambient particle details. "Dust particles visible in the air" adds significant depth and realism without increasing compositional complexity.
3. Include an audio hint if the model supports it. Both Veo 3 and Seedance 2.5 support native audio generation — "ambient sound: soft spinning wheel" is a valid and effective prompt element.
4. Reference the Seedance 2 product ad use cases for brand narrative prompt structures validated against commercial output.

⚠️ Common Mistake

Multi-shot narrative prompts (Shot A → Shot B → Shot C) tend to confuse single-clip models. If you need cuts between shots, generate each clip separately and assemble them in post — don't try to describe an editing sequence inside a single prompt.

Part 8: Universal Prompt Framework & Pre-Submit Checklist

Universal Video Prompt Framework

A structure that works across every scene type:

[Shot size] + [Subject] + [Action/Motion],
[Environment] + [Lighting],
[Camera movement] + [Camera character],
[Style/Color grade] + [Mood],
[Duration] + [Aspect ratio].
Optional: [Audio hint].

Filled-in example:

Wide tracking shot of a woman in a red coat walking through a snowy forest,
late afternoon light filtering through bare trees, soft blue shadows on snow.
Camera: slow lateral track, smooth and stable.
Cinematic, desaturated cool tones, quiet and melancholic.
8 seconds, 16:9. No dialogue.

Prompt Length Reference

Prompt Length Best For Risk
| Under 30 words | Quick tests, style exploration | Too vague, inconsistent output
| 40–80 words | Most production use cases | Sweet spot
| 80–120 words | Complex multi-element scenes | Possible element conflicts
| 120+ words | Rarely appropriate | High contradiction risk

Pre-Submit Checklist

Run through this before submitting any video prompt:

• [ ] No conflicting motion instructions — e.g., "handheld" and "perfectly stable" cannot coexist in the same prompt.
• [ ] Light source is named — "natural light" is weaker than "golden-hour sidelight from the right."
• [ ] Camera movement is specified — static, dolly, pan, tilt, handheld, aerial: pick one and state it clearly.
• [ ] Subject scale relationships are defined — especially important in landscapes, action scenes, and crowd shots.
• [ ] Style adjectives are visual, not abstract — "warm tones, film grain" beats "cinematic masterpiece."
• [ ] Duration is included — models use duration to calibrate motion pacing and transitions.
• [ ] No real person's likeness is referenced — describe physical and emotional characteristics, not a specific identity.
• [ ] Audio hint added for models that support it — don't leave the audio field blank in Veo 3 or Seedance 2.5 prompts.

Quick Reference: Recommended PixMind Tools by Step

Step Task Recommended Tool
| 1 | Upload footage, get a prompt draft | Video-to-Prompt
| 2 | Refine scene notes into a polished prompt | Text-to-Prompt
| 3 | Generate video from your final prompt | Veo or Seedance 2
| 4 | Go deeper on prompt strategy | YouTube Video-to-Prompt Guide

Part 9: Putting It All Together

Text-to-prompt is fundamentally a translation skill — converting visual information into the specific vocabulary that AI models are trained to understand. The more precisely you describe what you see (rather than what you feel), the more consistently the model can reproduce it.

Start with the five-layer framework, use the universal template as scaffolding, and run the checklist before every submission. Over time, you'll build a personal library of tested prompt templates that can be adapted for any new project.

The fastest way to accelerate that process: use PixMind's Video-to-Prompt tool to auto-extract a draft from reference footage, then apply the techniques in this guide to refine it manually. The combination of machine extraction and human refinement consistently outperforms either approach on its own.

Save the strongest result together with its prompt structure and reference choices. That small archive becomes far more useful than a gallery with no record of how the work was made.

Originally published by the PixMind Editorial Team https://www.pixmind.io/posts/text-to-prompt-guide