Turn a written brief into a finished AI video: shot lists, layered prompts, consistency anchors, sound design, and a review loop that saves passes.
A written brief already contains every decision a finished video needs, except the pictures. The distance between the two is rarely talent. It is a repeatable sequence of decisions: what the piece is for, which single frame carries the meaning, what the camera is allowed to do, and how a take gets approved or rejected. Teams that ship AI video consistently are not luckier with prompts. They convert prose into structure before a single frame exists, and they review against criteria they wrote down in advance.
The workflow below assumes you have a script, a client note, or at minimum a paragraph describing what you want, and no appetite for generating fifty clips in the hope that one lands. It is deliberately unglamorous, and it is the difference between a folder of promising clips and a video you can publish.
Decide the shape of the piece before you touch a generator
Most disappointing text-to-video output is a planning problem wearing a technical disguise. A paragraph of prose contains dozens of possible films, and a model asked to render a busy market at dawn with a tense atmosphere will choose something — just rarely the something you pictured. Three decisions made in ten minutes prevent most of that waste.
Duration and destination. A 15-second vertical ad and a 90-second brand film need different shot densities. Six shots in fifteen seconds reads as energetic. Six shots in ninety seconds reads as unfinished. Write the target runtime and aspect ratio on the first line of your document, because vertical and horizontal frames are different images, not crops of each other.
The one frame people must remember. If the video only lands a single image, which image is it? Build the shot list around it and give it the most generation attempts.
What the viewer must understand with the sound off. Silent scrolling is the default viewing condition, so action has to carry meaning. If a story beat only works with narration, it is a script problem, not a shot problem.
Write these three answers down. They become your acceptance criteria later, at the exact moment you are arguing with yourself about whether take four or take nine is the keeper — and they are the only reliable tiebreaker.
Turn the script into a shot list a model can act on
A shot list is the contract between the writer and the generator. Each line describes one camera setup, one dominant action, and one visible consequence. If a line has two actions, it is two lines.
The four things every shot line needs
Weak: she realizes the letter is from her father. Strong: medium close-up, woman in her thirties at a kitchen table, hand trembling as she lifts one sheet of paper, eyes tracking across the text, soft morning light from a window behind her.
The strong version names subject, framing, action, and light. Those four elements are the minimum a model needs to make a decision that matches yours. Add a fifth only when it changes the image: motion direction, camera movement, or a color accent.
One motion anchor per shot
Every shot needs a single named movement. Too little and the clip looks like a photograph with a slight drift; too much and the model invents new subjects mid-take. Say it explicitly: only the curtain moves; the camera dollies slowly right while the subject stays seated; steam rises and everything else is static. That one sentence prevents more failed takes than any style keyword you can stack into a prompt.
A practical test: read your shot line aloud and ask what moves. If you cannot answer in three words, the model cannot either. Common anchors are a person walking, a hand reaching, a vehicle crossing frame, fabric in wind, or a locked-off camera with environmental movement only.
Lock the visual system in one reusable block
Consistency across shots is the hardest part of AI video, and it is far cheaper to solve upstream than downstream. Pick a small set of repeatable parameters and paste them verbatim into every prompt in the project.
A workable baseline: one lens character (wide 24mm or portrait 85mm), one light direction (soft side light from camera left), one palette of three colors, one texture direction (clean digital or visible grain), and one grade direction (warm highlights, cool shadows). That block is your visual system. When a shot looks wrong but you cannot articulate why, nine times out of ten it deviates from the block.
Keep the block short enough to remember. A six-line system that everyone actually uses beats a forty-line style guide nobody opens.
Build a reference frame before you build motion
Generate a still first. Stills are fast to iterate, easy to compare side by side, and cheap to reject. Once the frame is right, drive the motion step from that image — motion models are much more obedient when they animate an existing composition rather than inventing one. That is the practical reason to treat the AI image generator and the video step as one continuous action: frame first, motion second.
Keep a small library of approved looks
After two or three projects you will notice you keep re-deriving the same three looks. Save them as presets — base prompt, palette, lens, grade — and setup time on the next project drops to minutes. Reusable starting points also make it possible for a second person to produce shots that match yours without a meeting. A template library is worth more than a longer prompt document, because it encodes decisions instead of describing them.
Write prompts as layered specifications
A prompt is a technical document with room for style. Structure it in layers so that when something goes wrong, you know which layer to edit instead of rewriting the whole thing.
The five-layer skeleton
- Subject. Who or what, with two or three identifying details that persist across shots.
- Framing and lens. Shot size, angle, focal-length feel, depth of field.
- Action and motion anchor. The one movement, plus an explicit statement about what stays still.
- Light and palette. Direction, quality, time of day, dominant colors.
- Texture and finish. Grain, cleanliness, grade, and a visual tradition such as documentary handheld or editorial product photography.
The order matters less than completeness, but keeping the order fixed across a project helps you spot omissions at a glance. When a clip drifts, layer three is missing more often than any other.
When output drifts, subtract before you add
The instinct when a take goes wrong is to add instructions. Usually the fix is subtraction. Remove adjectives that fight each other — serene plus chaotic plus energetic leaves the model no dominant note. Remove camera movement when the subject action is already complex. Remove secondary characters. Cut the prompt to two sentences, confirm the core reads correctly, then rebuild one layer at a time and note which addition changed what.
This is why logging matters. Twenty variants of the same shot with no record of what changed is not iteration, it is a slot machine with extra steps. Change one variable per attempt and write one line about the result. Five logged attempts teach you more than thirty unlogged ones.
Solve consistency for characters, rooms, and props
Faces and rooms are where AI video breaks, so plan for them instead of hoping.
Anchor every recurring character with a fixed description. Not a young man, but a man in his late twenties, short dark hair, thin scar above the left eyebrow, grey wool coat. The odd detail is the point: generic descriptions drift toward the model's average face, while a specific detail gives it something to hold onto.
Reuse one reference frame per character. Generate a clean portrait, then drive every appearance from that image rather than from text alone. The same logic applies to locations: one approved wide shot of the room becomes the reference for every angle inside it.
Use the cut as a consistency tool. A cut hides small differences that a continuous take exposes. If two shots of the same character never share a frame, minor variation reads as normal cinematic variation — the same principle editors have relied on for a century, related to how viewers build continuity from adjacent images.
Inventory your props. A red mug that changes shape between shots is more noticeable than a face that changes slightly, because the eye tracks objects by silhouette. Describe props once, in the same words, every time they appear.
Generate for the edit, not the highlight reel
Beautiful isolated clips often refuse to assemble. Prevent that by generating with the timeline in mind.
Match motion direction across cuts
If shot one ends with the subject moving right, shot two should not begin with movement left unless you want a jolt. Note the dominant motion direction on each shot line and check adjacency before you spend generation attempts on a shot that will fight its neighbours.
Generate handles, not exact durations
Ask for slightly longer clips than the timeline needs. Two extra seconds on each end give the editor room to find the natural in-point, trim the model's weakest moment, and hide a transition. Exact-duration generation sounds efficient and usually costs more time overall, because a clip that is one frame short forces a regeneration.
Transitions you can actually control
Match cuts, whip pans, occlusions — a hand, a passing vehicle, a pillar — are achievable and forgiving. Complex morphs between unrelated scenes are where AI video looks cheapest. When you need a hard scene change, cut on a sound or on motion instead of asking the model to morph a forest into a kitchen.
A useful planning habit: sketch the timeline as a strip of paper with clip lengths and transition marks before generating anything. If the strip looks boring, no amount of render quality will save it.
Treat sound and on-screen text as separate workstreams
Audio is not a postscript. Three tracks carry almost every AI video: music, ambience, and voice. Decide early whether narration exists, because a script written for voiceover has shorter sentences and clearer nouns than one written for captions.
For narration, generate voice first and cut picture to it, not the reverse. Timing dialogue to pre-rendered footage is always a compromise, and it shows as long, dead gaps.
On-screen text belongs in editing, not in the frame. Models still garble typography, and one misspelled word undoes an otherwise professional piece. Keep generated footage text-free, then composite titles and captions in your editor where you can control kerning, safe margins, and legibility.
Ambience does more for perceived realism than resolution. Room tone, distant traffic, cloth movement, a fridge hum in a kitchen scene — these tell the viewer's ear that the image is real even when the eye is unsure. A silent cinematic clip reads as synthetic almost instantly.
Finally, decide your captions format early. Burned-in captions force a specific framing with headroom; if you plan to add them later, leave that space in the composition from the first frame.
Review in passes, not by feel
Reviewing by instinct produces endless iteration, because there is always something to dislike. Review against a checklist, in three separate passes, and score each pass independently. Mixing them means a story problem gets fixed by regenerating a shot that was already correct.
The three-pass check
Pass one, story. Does the sequence communicate the idea with the sound off? If not, no grade will rescue it.
Pass two, continuity. Same character, same light direction, same palette, no unexplained jumps in motion or screen direction.
Pass three, craft. Framing, timing, audio balance, text legibility, export settings, file naming.
Regenerate or repair: a decision rule
| Problem | Action |
|---|---|
| Wrong action, wrong subject, wrong framing | Regenerate |
| Late start, dead half-second, small color mismatch | Repair in the edit |
| Text garbled in frame | Remove from generation, composite in editing |
| Motion direction conflicts with neighbour | Regenerate, or reverse the clip if the subject allows |
Regenerating a cosmetic flaw costs a full cycle. Trimming it costs thirty seconds. Teams that internalise this rule typically cut iteration counts nearly in half, and the savings compound across a series.
Mistakes and a realistic production rhythm
The six mistakes that quietly burn a week: prompt sprawl with no change log; generating before the script is locked, so every revision invalidates finished shots; chasing photorealism when a stylised look would hide artifacts better; ignoring aspect ratio until export; no naming convention, so half the late-stage time goes to finding the right file; and overloading a single clip with two ideas instead of one.
A five-shot, thirty-second piece usually runs in three blocks. Block one: script, shot list, visual system, reference frames — no motion generation yet. Block two: motion for every shot, two takes each, reviewed against the checklist. Block three: assembly, sound, text, grade, export.
The temptation is to jump to block two on day one because it feels like progress. It is the fastest way to produce clips that never assemble. The teams that finish fastest spend the longest in block one, and they spend it cheaply — stills and text cost minutes, motion takes cost the real effort.
FAQ
How many shots should a one-minute AI video have? Eight to fifteen for a paced narrative, fewer for atmospheric work. Let the script's beats decide density rather than a formula.
Why do my characters change between clips? Because text descriptions alone drift toward an average face. Anchor each character with a reference image and reuse it for every appearance.
Should I generate stills before animating? Almost always. Stills are faster to judge, cheaper to discard, and give the motion step a fixed composition to work from.
How do I stop the camera from moving when I want it locked off? State it as a constraint: static locked-off camera, no movement. Silence is interpreted as permission.
Can I fix a bad clip in the edit instead of regenerating? If the problem is timing, colour, length, or audio, yes. If the subject or the action is wrong, no.
Which matters more, prompt length or prompt clarity? Clarity. A four-line prompt with one motion anchor beats a paragraph of competing adjectives every time.
How do I keep a series visually consistent across weeks? Save the visual system, the reference frames, and two approved looks per project. Reopen the preset before writing a single new prompt.
Start with one shot you can defend
Pick the frame you would put on a poster, write its shot line, generate the still, then animate it with one motion anchor. That single loop — frame, motion, review — is the whole method in miniature, and it will teach you more in an afternoon than a week of scattered experiments.
Keep your shot list open beside the AI video generator, pull proven prompt patterns from the Orelon prompt library, and treat every generation as a test of one specific decision rather than a lottery ticket. Cinematic ideas in motion start with a plan you can actually see — and a plan you can see is one you can fix.

