Start With the Vocabulary Problem, Not the Tool
Most creative teams do not actually have a model problem. They have a language problem. The same prompt pasted into three different generators can return a painterly portrait, a plastic-looking render, and a wide shot with the subject half out of frame. That is not proof that one engine is bad. It is proof that each engine parses words differently, weights them differently, and fills gaps differently. The teams that ship consistent work treat prompt language as a maintained asset: a shared vocabulary of keywords, phrases, and structures they refine over time instead of rewriting from zero for every job.
The shift matters because image and video generation has moved from generate something impressive to generate something directable. Current systems understand camera language, lighting setups, continuity constraints, aspect ratios, and shot timing. But they only respond to that language when you use it deliberately and consistently. Vague prompts do not fail loudly. They fail quietly, producing output that is 80 percent right and impossible to repair by tweaking one random word.
This guide lays out a practical keyword system for AI image and video production. It covers the layers of a directable prompt, the phrases that reliably change output, model-selection criteria, a full production workflow, quality control, and the failure modes that quietly consume the most production hours.
The Four Layers of a Directable Prompt
Every production-ready prompt can be decomposed into four layers. When output goes wrong, the fastest debugging method is to ask which layer is under-specified rather than to add more adjectives.
Layer 1: Subject and action
Who or what is on screen, and what are they doing in this exact moment. Strong subject phrasing is specific but not crowded: 'a middle-aged ceramicist shaping a bowl on a kick wheel, hands coated in wet clay' beats 'a woman making pottery.' Action verbs carry timing. Static nouns produce static frames, which is a common reason generated clips look like photographs with a slight drift.
Layer 2: Look and light
This layer controls style, palette, texture, and illumination. It is where most keyword lists live, and it is also where people over-stack. Five style words fight each other; two well-chosen ones reinforce each other. Useful phrasing here includes 'photorealistic cinematic lighting,' 'soft north-window light,' 'high-contrast rim light with warm practicals in the background,' or 'muted earth-tone palette with film grain.'
Layer 3: Camera and framing
Shot size, angle, lens feel, movement, and depth. 'Medium close-up, 50mm equivalent, shallow depth of field, slight handheld float' gives a generator more usable constraints than three paragraphs of story. If your output keeps drifting to wide shots, this layer is the missing one.
Layer 4: Continuity and constraints
Identity, wardrobe, props, time of day, environment, and what must not change between shots. Phrases like 'same jacket, same location, same time of day as previous shot' and 'character keyframe consistency' belong here. This is the layer that turns isolated clips into a sequence.
A useful habit: write four lines, one per layer, before you write a single sentence of prose prompt. It forces you to notice which layer you have been ignoring.
Image Keywords That Reliably Change Output
Image models respond most predictably to light, material, and lens language. Story language moves them far less. If you want a specific look, describe the physical conditions that create it rather than naming a mood.
- Lighting geometry: 'backlit with soft haze,' 'single hard key from camera left,' 'overcast diffused daylight,' 'warm practical lamps in frame,' 'golden-hour side light with long shadows.'
- Material and skin: 'matte skin texture with visible pores,' 'brushed aluminum with micro-scratches,' 'worn leather with cracked grain,' 'translucent frosted glass.' Material words fix the plastic look that plagues synthetic images.
- Lens and capture feel: '35mm, f/2.0,' 'telephoto compression,' 'anamorphic flare,' 'slight motion blur in the background,' 'documentary handheld framing.'
- Color treatment: 'teal shadows and amber highlights,' 'desaturated with a single red accent,' 'high-key pastel palette,' 'cross-processed greens.'
- Negative framing: what to exclude. 'No text, no watermark, no extra fingers, no mirrored reflections' is worth more than another style adjective in most workflows.
One practical rule: keep image prompts under roughly 60 to 80 meaningful words. Beyond that, models begin averaging your vocabulary into a mush of generic competence. If a shot is not right at 70 words, the answer is usually a reference image or a different model, not 30 more words.
Video Keywords: Motion, Camera, and Time
Video adds a dimension that breaks most image-trained prompt habits: change over time. A video prompt has to describe not just a frame but a trajectory. There are four keyword families worth building separately.
Motion verbs
'Walks,' 'turns,' 'reaches,' 'pours,' 'opens,' 'steps backward,' 'glances up.' Prefer one primary action per clip. Models struggle to render two simultaneous independent actions without one of them collapsing into morphing artifacts.
Camera movement
'Slow push in,' 'lateral tracking shot,' 'orbit around subject,' 'static tripod,' 'handheld follow,' 'crane up revealing the skyline.' Naming the movement explicitly is the single fastest way to make generated footage feel intentional rather than drifting.
Timing and pacing
'Slow motion, 120fps feel,' 'real-time pace,' 'quick three-beat montage,' 'long hold with minimal movement.' Timing words help you plan how a clip will cut against music or narration, and they reduce the risk of a clip that is technically fine but editorially unusable.
Physics and continuity
'Fabric settles naturally,' 'liquid pours with realistic viscosity,' 'hair moves with wind from the left,' 'consistent lighting direction across the cut.' Physics keywords reduce the rubbery, gravity-free motion that viewers notice immediately even when they cannot name it.
A workable structure for a video prompt: subject and action, then camera, then light and look, then motion and physics constraints, then what must stay consistent with the previous shot. Four to six short clauses, not one flowing paragraph.
Building a Reusable Template Library
Ad hoc prompting does not scale. A template library does. The goal is not to remove creativity but to remove repeated decision-making.
The base template
Four lines, matching the four layers: subject and action, light and look, camera and framing, continuity notes. Fill it in for every shot. After twenty shots you will start noticing which phrasings you reuse, and those become your house style.
Named presets
Create short, memorable preset names for recurring looks: 'interview-soft,' 'product-hero-hard,' 'street-doc-handheld,' 'night-neon-wet.' Each preset expands into a fixed block of keywords. This keeps look consistent across a series and lets a teammate generate matching footage without a briefing call.
Negative keyword lists
Keep a shared list of exclusions per project. Brand-sensitive projects need 'no logos,' 'no text overlays,' 'no recognizable faces.' Character projects need artifact exclusions. Store the list once and reference it everywhere instead of retyping it.
Prompt-to-shot mapping
Number every shot and pair it with a prompt file, a seed or reference image, and the intended duration. When a client asks for a revision on shot 12, you open shot 12's file rather than re-deriving the look from memory. This single habit prevents the most common cause of visual drift in long edits.
Choosing the Right Model for the Shot
Model choice is a decision matrix, not loyalty. Evaluate each option against five criteria.
- Realism ceiling. How does it handle skin, hands, and small text? Test with your own subject matter rather than demo reels.
- Motion quality. Does movement stay coherent for the full clip length, or does the subject degrade after a few seconds?
- Controllability. Can you supply reference images, depth maps, or pose guidance? Controllability matters more than raw beauty when you are matching existing footage.
- Consistency across shots. Can the same character appear three times in a sequence without drifting in face, wardrobe, or hairline?
- Throughput. How long does a usable take take, and how many attempts does it need on average? A model with stunning output that needs twelve attempts is slower than a modest model that lands in two.
A practical approach is to assign roles: one engine as your photoreal workhorse, one for stylized or animated looks, one for quick storyboard drafts where speed beats finish. Document the role of each in your template library so nobody relitigates the choice mid-project.
The End-to-End Production Workflow
A repeatable pipeline beats inspiration. Here is a structure that works for short-form video, product spots, and narrative sequences.
Step 1: Script to shot list
Break the script into shots with a duration estimate for each. One idea per shot. If a shot needs two ideas, split it.
Step 2: Style frame first
Generate a single hero image that defines look, palette, and lighting before generating anything with motion. Approving a style frame is far cheaper than approving motion that has to be redone.
Step 3: Blocking pass
Generate the entire sequence at low fidelity or low resolution. The goal is rhythm and coverage, not beauty. Watch it end to end and fix story problems here, not later.
Step 4: Hero pass
Regenerate only the shots that carry the message at full quality, using the approved style frame as reference. Keep the seed stable where a shot works.
Step 5: Repair pass
Fix hands, edges, reflections, and text. Prefer targeted fixes over full regeneration, because regeneration resets everything you liked about the take.
Step 6: Assembly
Cut against music and narration. AI clips often run slightly differently than planned, so build in a beat of flexibility at the head and tail of each clip.
Step 7: Finish
Color unify, add grain or texture, and check audio sync. A unifying grade hides minor inconsistencies between shots better than any prompt trick.
Quality Control: Common Failures and Fixes
Most defects fall into recognizable categories. Fixing them is faster when you know the likely cause.
- Morphing faces and shifting identity. Add explicit continuity keywords, supply a reference image, and reduce the number of actions per clip.
- Rubbery motion. Add physics keywords and simplify to a single movement. Long, complex trajectories degrade on almost every model.
- Plastic skin and surfaces. Add material and texture language, remove over-smoothing style words like 'perfect' or 'flawless,' and reduce stylization.
- Wide-shot drift. Specify shot size and lens explicitly. Generators default to wide framing when framing is unstated.
- Lighting inconsistency across a sequence. Lock a lighting phrase into your preset and repeat it verbatim in every prompt for that scene.
- Garbled text and signage. Avoid text in generation entirely. Add typography in post.
- Over-stacked style words. Cut style keywords to two. If the look does not survive that reduction, the problem is the reference, not the vocabulary.
One meta-rule: change one variable at a time. When a take improves after you changed five keywords, you have learned nothing reusable. When it improves after you changed one, you have added a permanent tool to your library.
Team Handoff, Naming, and Versioning
Keyword systems break down at handoff. Solve it with boring conventions.
- File naming: project, scene, shot number, version. Nothing else.
- Prompt naming: the preset name plus any shot-specific overrides, so a reader can see at a glance what deviates from the house style.
- Change logs: one line per revision describing what changed and why. This becomes your institutional memory.
- Preset reviews: once a month, promote the phrasings that worked repeatedly and retire the ones that never did. Vocabulary rots when nobody prunes it.
This is unglamorous work, but it is the difference between a team that re-solves the same problem every week and a team that compounds its skill.
Frequently Asked Questions
How long should an AI image or video prompt be?
For images, roughly 40 to 80 meaningful words. For video, four to six short clauses covering subject, camera, light, and motion. Longer prompts dilute emphasis rather than adding control.
Do keyword order and phrasing really matter?
Yes, though the effect varies by model. Most engines weight earlier clauses more heavily, so put subject and action first, then camera, then style. Consistent ordering also makes your own debugging easier.
How do I keep a character consistent across many shots?
Combine three things: a locked reference image, explicit continuity keywords repeated verbatim in every prompt for that scene, and minimal action complexity per clip. If identity still drifts, reduce shot length and add a repair pass rather than regenerating everything.
Should I use the same model for every shot?
Usually not. Assign engines to roles based on realism, motion quality, controllability, consistency, and throughput. Document those roles so the choice is not re-debated every production.
What is the fastest way to improve my output quality?
Stop writing prose prompts. Write four-line structured prompts, keep a preset library, and change one variable per test. Most quality gains come from discipline, not from discovering a new tool.
How do I handle scenes with dialogue or text on screen?
Generate the visuals without text and add all typography, subtitles, and signage in post. Text generation remains the least reliable part of the pipeline and the most expensive to repair.
Where does upscaling and final polish fit?
The final stage, after selection. Upscale approved shots only, then apply a unifying grade across the sequence. Upscaling early just makes bad takes more expensive to store and review.
The Takeaway
AI image and video generation rewards people who build systems. A maintained keyword vocabulary, four-layer prompts, named presets, clear model roles, and a disciplined repair pass will outperform any single tool upgrade. Start small: pick one project, write structured prompts for every shot, and keep a log of what changed the output. Within a few productions you will have a library that is genuinely yours, and the next model release will be an upgrade to your workflow rather than a reason to rebuild it.
