Most people treat an AI prompt like a search query: type a rough idea, hit generate, hope for the best. Then they blame the model when the result looks generic. The real problem is usually not the engine — it is the brief. A prompt is a production document compressed into a paragraph, and the quality of your output tracks the quality of that document almost linearly.
This guide walks through a repeatable prompting system for image and video generation: how to structure a prompt, which camera and lighting vocabulary actually changes results, how negative prompts work, how to keep characters and products consistent across shots, and how to iterate without burning hours on random rerolls. Everything here is model-agnostic, so you can apply it whether you are working in a diffusion image tool, a text-to-video engine, or an image-to-video pipeline.
Structure Beats Length Every Time
The most common failure mode is a long, poetic prompt with no hierarchy. A model cannot tell which of your twelve clauses matters most, so it averages them, and averaging produces mush.
A production-ready prompt has three properties:
- Ordered priority. The most important information appears first, because attention tends to decay across a long prompt.
- Separated concerns. Subject, action, environment, camera, light, and style live in distinct clauses instead of bleeding into each other.
- Concrete nouns and verbs. "A weathered fisherman" beats "a person who looks like they have lived a hard life." Models respond to nouns, not explanations.
A useful mental model is a shot card on a film set. A director does not say "make it cinematic." They say who is in frame, what they are doing, where the camera is, what the light is doing, and what lens is on the camera. Prompts work the same way.
Keep a reusable skeleton and fill in the blanks rather than writing freeform every time:
[subject + distinguishing detail], [action or state], [environment + time of day],
[camera position + movement], [lens + depth of field], [lighting + color],
[style or medium], [technical quality notes]
Even a mediocre idea plugged into this skeleton outperforms a brilliant idea written as a stream of consciousness.
The Six Building Blocks of a Working Prompt
1. Subject with one anchoring detail
Name the subject and give exactly one or two details that force a specific interpretation. "A ceramicist" is vague. "A ceramicist in her sixties with clay-dusted forearms and a faded denim apron" locks the model onto a person rather than a stock photo concept.
2. Action or emotional state
The subject should be doing something, even something small. "Standing" is fine; "standing with weight on one hip, looking off-frame as if interrupted" is better. Stills and video both benefit from implied motion because it gives the model a direction to animate toward.
3. Environment and time
Where and when. Include weather, surface textures, and background activity level. "Busy street" and "empty street at dawn" produce completely different films even with the same subject.
4. Camera and movement
This is the highest-leverage block for video and the most frequently skipped. Specify position, angle, and movement: low angle, eye level, over-the-shoulder, slow dolly in, handheld follow, static lock-off. If you say nothing, you get the model's default, which is usually a slow push in or a drifting pan.
5. Lighting and color
Light determines mood faster than any adjective about emotion. Name the source, the direction, and the quality: soft window light from camera left, hard practical neon behind the subject, overcast top light, warm bounce from a sandy floor.
6. Style and medium
Editorial photography, 35mm film still, cel-shaded animation, documentary handheld, product tabletop. Pick one and commit. Stacking three style references creates conflict and muddies the render.
Camera and Lens Vocabulary That Changes Results
Camera terms are not decoration. In most modern engines they act as real controls, influencing perspective, motion behavior, and depth rendering.
Shot size. Extreme wide, wide, medium wide, medium, medium close-up, close-up, extreme close-up. Shot size is the single fastest way to change the emotional register of a frame. A medium close-up reads as intimacy; an extreme wide reads as isolation.
Angle. Eye level feels neutral and documentary. Low angle confers power. High angle confers vulnerability or scale. Dutch angle introduces unease. Bird's-eye and worm's-eye are strong enough that they should only be used deliberately.
Lens and depth. An 85mm lens at a wide aperture gives compressed backgrounds and creamy separation. A 24mm lens gives environmental context and slight distortion at the edges. Naming a focal length and a shallow or deep depth of field guides the model's focus behavior far more reliably than "blurry background."
Movement. For video, choose one primary move per shot: slow dolly in, dolly out, truck left, crane up, orbit at 30 degrees, handheld follow, whip pan, static with subject motion only. Two simultaneous moves confuse motion synthesis and produce warping artifacts.
Frame rate and cadence. Terms like "24fps cinematic motion blur" or "crisp 60fps sports look" subtly shift how motion is rendered. They are soft controls, not switches, but they help.
A compact example for a video shot:
Medium close-up, eye level, slow dolly in on a chef plating a dessert,
steam rising, 50mm lens, shallow depth of field, warm tungsten practicals
from camera right, deep shadows, documentary style
Notice that nothing in that prompt says "beautiful." Beauty is implied by the specifics.
Lighting, Color, and Mood Without Empty Adjectives
Words like "epic," "stunning," and "masterpiece" do almost nothing. They are not selectable features in the model's training distribution. Replace them with physical descriptions of light.
Direction. Front light flattens. Side light sculpts. Backlight separates and creates rim glow. Top light is harsh and interrogative. Underlight is unnatural and unsettling.
Quality. Hard light produces crisp shadows and high contrast. Soft light wraps and flatters. Diffused overcast is even but flat. Mixed hard and soft creates the depth that most people actually mean when they say "cinematic."
Color temperature. Golden hour warmth, blue hour coolness, sodium-vapor orange streetlights, fluorescent green-cyan cast, candle flicker at low kelvin. Contrast between a warm subject and a cool background is one of the most reliable ways to make a frame feel professionally lit.
Practical sources. Name lights that exist inside the scene: a desk lamp, a phone screen lighting a face, taillights, a refrigerator glow, firelight. Practicals give motivation to the lighting and make the image read as photographed rather than generated.
Palette. Instead of "moody," name two or three colors: teal shadows, amber highlights, desaturated greens. Constraining the palette is one of the strongest consistency tools across a series of shots.
Negative Prompts and Guardrails That Actually Work
Negative prompts list what you do not want. They are most effective when they target recurring failure patterns in the specific engine you are using, not when they are a giant block of copied text.
Use a short, maintained list. Something like:
extra fingers, fused limbs, warped hands, text artifacts, watermark,
duplicated subject, smeared background, over-sharpened halos,
plastic skin, oversaturated neon, jittery motion, morphing faces
Then add one or two project-specific entries. If you are generating a period piece, add "modern signage, contemporary clothing, visible power lines." If you are generating a product shot, add "extra logos, distorted label text, reflections that reveal a camera rig."
Two rules keep negatives useful:
- Never negate something you also want. If you want a dark scene, do not write "dark" in the negatives. Contradictions push the model toward unpredictable compromises.
- Prune monthly. Every engine update changes which artifacts are common. A negative list that was essential a year ago may now be obsolete or actively harmful.
For text rendering, be honest about limits. If the shot requires legible typography, generate the plate without text and composite the type afterward. Fighting a model into perfect typography is slower than doing it properly in a design tool.
Adapting One Idea Across Different Model Types
The same creative intent needs different phrasing depending on the engine family. There are three broad categories to plan for.
Diffusion image models. These reward dense descriptive detail and strong style references. They respond well to artist-adjacent descriptors, materials, and texture words. Prompt order matters, and many support weighted terms or regional attention, which lets you push emphasis onto specific nouns.
Text-to-video engines. These reward motion clarity and brevity. Long, novelistic prompts often produce mush because the model has to allocate capacity across time as well as space. Write one clean sentence of subject and action, then a second sentence of camera and light. Keep total length shorter than you would for a still.
Image-to-video and video-to-video pipelines. Here the prompt is a modifier, not a creator. Your job is to describe the motion you want added, not to redescribe the frame. "Slow parallax to the right, hair moving in the breeze, background crowd continues walking, no change to subject identity" is the correct register.
A practical habit: keep a single project file with three variants of every key shot — a still version, a text-to-video version, and an image-to-video motion instruction. When you switch engines mid-project, you already have the translation ready.
Some current engines — including Flux-class image models, and video systems such as Runway, Sora, Kling, Veo, Pika, Luma, MiniMax, and Vidu — differ in how literally they interpret camera language and how much they tolerate prompt length. Rather than memorizing rumors about each, run a two-minute calibration test: generate the same six-word prompt, a twenty-word prompt, and a sixty-word prompt on any new engine. You will learn its appetite for detail immediately.
Consistency Across Shots: Characters, Products, and Multi-Image Fusion
A single beautiful frame is easy. A sequence where the same person appears in six shots is where most projects fall apart.
Lock a reference block. Write a fixed paragraph describing your character or product — face shape, hair, wardrobe, key accessories, signature color. Paste it verbatim into every prompt. Never paraphrase it. Small wording changes produce visible drift.
Fix the palette across the sequence. Decide the two or three dominant colors before you generate anything, then repeat them in every prompt. Consistent color grading hides minor identity inconsistencies from viewers.
Use image references where available. Multi-image or reference-image conditioning is dramatically more reliable than text descriptions for identity. Feed a clean, well-lit reference, ideally neutral expression and simple background. Generate one, then use it as the anchor for the rest.
Control wardrobe changes deliberately. If the character must change clothes between scenes, change only that clause and keep everything else identical. This isolates the variable and keeps the face stable.
For products, keep angle discipline. Front, three-quarter, side, and detail macro. Repeating the same set of angles across a campaign produces a coherent library instead of a random assortment.
Check continuity in a contact sheet. Assemble all shots in a grid before judging. Drift that is invisible in isolation becomes obvious side by side.
A Five-Pass Iteration Workflow
Random rerolling is the most expensive habit in generative work. Replace it with structured passes.
Pass 1 — Composition only. Ignore style, color, and polish. Generate several variations at low resolution and pick the framing that tells the story. Do not fall in love with a beautiful frame that has the wrong composition.
Pass 2 — Lock the subject. Once framing is right, refine the subject and action clause until identity and pose are correct. Change nothing else.
Pass 3 — Light and grade. Add your lighting, palette, and contrast language. This is the pass that usually makes the biggest visual jump.
Pass 4 — Motion. For video, add camera movement and subject motion. Keep it to one move. If motion warps, reduce amplitude rather than rerolling repeatedly.
Pass 5 — Polish and fix. Upscale, repair hands or small artifacts with an editing or inpainting pass, and clean audio or add sound design.
Save the exact prompt and settings for every shot you keep. A project without a prompt log cannot be revised, reshot, or extended, and clients always come back with one more variant.
Common Mistakes and How to Fix Them
Stacking style references. "Blade Runner meets Studio Ghibli meets Wes Anderson" produces an unresolved average. Pick one primary reference and one supporting adjective at most.
Describing the result instead of the scene. "A viral-looking video" tells the model nothing. Describe what is on screen and let the audience decide if it is viral.
Overloading one shot. Cramming three actions into a five-second clip guarantees that none of them render cleanly. Cut it into three shots.
Ignoring the first three seconds. In video generation, the opening frames set the motion trajectory. If the first frame is ambiguous, the whole clip drifts.
Fighting the model's strengths. If an engine is excellent at natural landscapes and weak at hands, do not build your project around close-ups of hands. Route the work to the tool that handles it and composite the rest.
Never documenting failures. A prompt that failed is data. Note what broke, and your personal negative-prompt list becomes genuinely valuable within a few weeks.
Frequently Asked Questions
How long should a video prompt be?
Long enough to specify subject, action, camera, and light — usually one to three sentences. Beyond that, you are usually adding style adjectives that dilute the motion instructions. If your prompt is longer than a paragraph, cut the weakest half and compare results.
Do negative prompts work in every engine?
No. Some engines have a dedicated negative field, some accept inline negations, and some ignore them almost entirely. Test with a single obvious negative like "watermark" and see whether behavior changes. If it does not, move that cleanup into post-production instead of fighting the model.
Why does my character change between shots?
Ninety percent of the time it is because the description was reworded. Use an identical reference block, identical palette clauses, and an image reference if the engine supports it. Then check a contact sheet before generating more.
Is prompt weighting worth learning?
Yes, if your engine supports it. Weighting is far more precise than repeating words or adding emphasis adverbs. But use it sparingly — a single weighted term usually does more than five.
How do I get realistic camera motion instead of a floating, dreamy drift?
Name a physical camera behavior: tripod lock-off with subject motion only, handheld follow, slow dolly on a track. Add a motion-blur cue. Most drifting comes from leaving movement unspecified and letting the model default to a slow push.
Should I generate at high resolution on the first pass?
No. Iterate at low resolution and only render finals at full quality. You will cut rendering time dramatically and avoid polishing frames you are going to discard.
How do I keep a series visually unified?
Three things: a fixed palette of two or three colors, a consistent lighting direction, and a single style descriptor repeated verbatim. Consistency comes from restriction, not from adding more description.
What is the fastest way to improve my prompts today?
Add a camera clause and a lighting clause to every prompt you write. Those two additions change output more than any other single edit, and they take about ten seconds to write.
Prompting is a craft with a short learning curve and a long mastery tail. The practitioners who get consistently good results are not using secret words — they are writing organized briefs, testing systematically, and keeping records. Start with the six building blocks, calibrate each new engine with a length test, and build a prompt log for your current project. Within a week, your hit rate will be high enough that rerolling stops feeling like a strategy and starts feeling like what it is: a symptom of an unclear brief.
