Start With Story, Not With Settings
Every conversation about AI filmmaking eventually collapses into a list of settings: resolution, duration, motion strength, seed value. Those parameters matter, but debating them before you know what your film is about is like renting a lens before you have a script. The creators making genuinely watchable short films with generative video treat the prompt as a director's instruction to a crew, not as a search query for a stock clip.
The practical consequence is a shift in where you spend your time. Instead of cycling through dozens of vague prompts hoping something looks good, you write a short screenplay, break it into shots, and only then translate each shot into model-ready language. That sequence — story, then shot, then prompt — is the whole discipline. Skip a step and you get beautiful footage that never becomes a film.
A good test before you open any tool: can you describe your short in two sentences, including what changes between the first frame and the last? If yes, you are ready to generate. If you can only describe a mood or an aesthetic, you are about to make a demo reel, not a story.
This guide covers the prompt hierarchy, camera vocabulary, continuity systems, sound planning, a full worked example, and the mistakes that most often derail AI short films.
The Prompt Hierarchy: Four Layers That Do the Heavy Lifting
Most prompt failures come from cramming everything into one flat sentence. Models respond better to a layered description, ordered from what matters most to what matters least. A reliable order is: subject and action, then camera, then light and lens, then continuity anchors.
Layer 1: Subject and Action
Lead with who or what is on screen and what they are doing right now. Be concrete about age range, clothing, posture, and the specific physical action. "A woman walks through a city" gives the model almost nothing to work with. "A woman in her late thirties in a soaked wool coat limps toward the camera, favouring her right leg" gives it a body, a wardrobe, and a movement signature you can repeat in the next shot.
Keep each shot to one dominant action. Two simultaneous actions — she runs while a car explodes while a dog barks — usually produce mush, because the model has to allocate attention across competing instructions.
Layer 2: Camera and Framing
Describe the shot size, the angle, and the movement. A useful shorthand: wide establishing, medium two-shot, close-up, extreme close-up, over-the-shoulder, low angle, high angle. Then add motion if you want it: slow push-in, lateral tracking, handheld drift, static lock-off, crane rise.
Motion instructions are where prompts most often contradict themselves. "Static shot with slow orbiting camera" cannot be resolved. Pick one dominant move per shot and let the next shot carry the variety.
Layer 3: Light, Lens, and Palette
This layer creates the look. Specify the source and quality of light — hard midday sun, soft window light, sodium-vapour street lamps, flickering practicals — plus a lens feel such as shallow depth of field, wide-angle distortion, or long-lens compression. Add a restrained palette: two or three colours, not eight. Rain and haze give depth cues that models render convincingly.
Layer 4: Continuity Anchors
Finish with the details that must survive into the next shot: wardrobe, props, hair, time of day, weather, and grade. These anchors become your carry-over template. Copy them verbatim into subsequent prompts rather than paraphrasing, because small wording changes produce visible drift.
Build a Shot List Before You Generate a Single Frame
A shot list converts narrative intent into production units. Write it as a table with five columns: shot number, story function, shot size, duration, and prompt seed. The story-function column is the one people skip, and it is the one that saves you.
If a shot's function is "establish that the town is dying," you will instinctively give it enough time and the right framing. If the function is "show her decision," you will reach for a close-up rather than a drone shot, no matter how nice the drone shot would look.
| Shot | Function | Size | Duration | Prompt seed |
|---|---|---|---|---|
| 1 | Establish place and mood | Extreme wide | 6s | Coastal village at dusk, empty harbour, static |
| 2 | Introduce protagonist | Medium | 4s | Woman late 30s, wool coat, walking, tracking |
| 3 | Plant the object | Close-up | 3s | Hands lifting brass key from drawer |
A ninety-second short typically needs twelve to twenty shots. Anything fewer and the rhythm drags; anything more and each shot becomes too brief to register. Budget your total runtime first, then divide it into shots of three to eight seconds, which is also the sweet spot for most generation tools.
Camera Language That Video Models Actually Understand
Models have learned visual vocabulary from film captions, so cinematic terms perform better than invented ones. The following phrasing is generally reliable:
- Static lock-off — no camera movement, subject moves within frame.
- Slow push-in — camera advances toward subject, tension builds.
- Lateral tracking — camera moves sideways, often with a walking subject.
- Handheld — subtle instability, documentary energy.
- Crane or drone rise — expands the world, good for act breaks.
- Rack focus — shifts attention between foreground and background.
- Whip pan — fast rotation, useful as a transition generator.
Two rules make these work. First, name the direction of movement (left to right, toward camera, away) because models frequently guess wrong. Second, describe the subject's speed separately from the camera's speed. "She walks slowly while the camera tracks her at the same pace" is far more controllable than "tracking shot of a walking woman."
For coverage, generate each scene from at least two sizes — a medium and a close — even if you plan to use only one. This gives you cut points when a performance or a hand gesture does not land.
Character and Location Continuity Across Shots
Continuity is the single largest technical obstacle in AI short films. Faces drift, coats change colour, and a kitchen counter rearranges itself between cuts. The fix is a character bible and a location bible, both written once and reused as literal strings in every prompt.
A character bible entry might read: "MIRA — woman, late thirties, dark hair tied back loosely, olive wool coat with a missing second button, small scar above left eyebrow, tired eyes." That string goes into every shot she appears in, unchanged. It is long, and that is fine; consistency beats brevity.
For locations, lock the geography: "harbour office, peeling green paint, single window facing east, brass lamp on the desk, stack of tide charts on the left." Mentioning the same two or three props in every shot of that location helps the model treat it as the same room.
When you have strong reference images, use image-to-video rather than text-to-video for character shots. Feeding the same still into multiple generations is the most reliable continuity tool available, and it also shortens your prompts because the model no longer has to invent the face.
Sound, Dialogue, and Rhythm in a Generation-First Edit
Most video generation still treats audio as a separate problem. Plan for that from the start instead of discovering it in the edit.
- Dialogue-heavy scenes are the hardest to generate convincingly. Keep spoken lines short — under eight words — and expect to record them separately with a voice performance. Design shots that show a face mid-speech rather than demanding perfect lip-sync on long monologues.
- Ambience carries continuity. A harbour scene with gull calls and rope creak reads as one place even if the visuals wobble. Lay ambience before you fuss over music.
- Sound bridges hide cuts. Start the next scene's audio a half-second before the picture cut and the audience will forgive a small visual discontinuity.
- Silence is a tool. Cutting all sound for one beat before a reveal is cheap and enormously effective.
Pacing is decided in the edit, but it is enabled by your shot list. If every shot is six seconds, the film will feel metronomic. Alternate three-second inserts with eight-second wide shots and the rhythm starts to breathe.
A Worked Example: A Ninety-Second Short in Twelve Shots
Logline: a night-shift lighthouse keeper receives a radio call from a boat that sank forty years ago. The film ends when she answers.
- Extreme wide, 6s. Lighthouse on a black cliff, rain, single beam rotating, static camera. Palette: deep blue, sodium orange.
- Medium, 4s. Keeper, fifties, oilskin jacket, climbing the spiral stair, handheld following her back.
- Close-up, 3s. Her hand turning the brass radio dial, static, shallow depth of field.
- Insert, 2s. Tide log open on the desk, pages fluttering, ceiling lamp swinging.
- Medium, 5s. She sits, pours coffee, steam rising, slow push-in.
- Close-up, 3s. Static hiss from the radio, her eyes lifting, rack focus from dial to face.
- Over-the-shoulder, 4s. Radio on the shelf, faint voice, she leans in.
- Close-up, 3s. Her mouth forming a word, no audio needed — hold the cut.
- Wide interior, 5s. Room tilts in the lamp light, she stands abruptly, chair scraping.
- Extreme wide exterior, 6s. The beam sweeps the sea; far off, one small light answers.
- Close-up, 4s. Her hand on the microphone key, trembling, static lock-off.
- Medium, 8s. She presses the key and speaks. Slow push-in, then cut to black on the sound of static.
Notice how each prompt carries only the information that shot needs, plus the continuity anchors. Shot 11 is a static close-up because the trembling hand is the story; adding a camera move would dilute it. Shot 12 is longer because the film's meaning happens there and the audience needs room to feel it.
Common Mistakes That Ruin AI Short Films
Overloading a single prompt. Twelve clauses produce a lottery draw, not a shot. If a prompt has more than about sixty words, split it into two shots or cut the least important layer.
Changing too many variables at once. When a shot fails, change one thing — camera, or lighting, or action — and regenerate. Otherwise you learn nothing about what the model responded to.
Ignoring aspect ratio and delivery. Decide now whether you are delivering vertical or widescreen. Cropping a widescreen-generated film into vertical later destroys compositions and cuts off faces.
Expecting one generation to be the final shot. Treat every output as a take. Ten to twenty attempts per finished shot is normal, and the attempts you reject teach you the model's limits.
Neglecting the edit plan. Generation is half the work. If you have no timeline, no scratch music, and no rough cut structure, you will keep generating new shots instead of finishing the film.
Asking for on-screen text. Logos, signage, and subtitles remain unreliable. Add them in post.
Letting the tool choose the story. Generators are excellent at texture and terrible at intention. You supply the intention.
Review, Iterate, and Finish
Watch your assembled cut once at normal speed without pausing. Note the exact moment your attention drops — that timestamp is your editing problem, not your prompting problem. Only then go shot by shot and rate each on three criteria: story clarity, motion artifacts, and continuity with neighbours.
Keep a prompt log: shot number, full prompt text, seed, model, and a one-line verdict. After a few sessions you will see patterns — which phrasings produce stable faces, which camera verbs cause warping — and your hit rate will climb sharply.
When a shot is ninety percent right, resist the urge to regenerate. Fix the remaining ten percent in the edit with a trim, a speed change, or a cutaway. Finishing is a skill separate from generating.
FAQ: Storytelling Prompts for AI Short Films
How long should each generated clip be?
Three to eight seconds covers most needs. Generate slightly longer than you plan to use so you have handles for trimming and transitions.
Do I need a full screenplay?
No, but you need a logline, a beat sheet of five to eight beats, and a shot list. That is roughly one page of writing and it will save you hours.
Why do my characters change between shots?
Almost always because the descriptive wording changed. Freeze a character bible string and paste it verbatim into every prompt, or use image-to-video with a consistent reference frame.
Can these tools handle dialogue?
Short lines sometimes work; longer exchanges rarely do. Write dialogue that can be delivered in a shot where the mouth is partly obscured or turned away, then record the voice separately.
Should I use text-to-video or image-to-video?
Use text-to-video for establishing shots and environments, image-to-video for anything with recurring characters or precise composition. Most projects mix both.
How many takes should I expect per shot?
Plan for ten or more. Reviewing takes quickly and rejecting early is a bigger productivity gain than writing longer prompts.
How do I keep a consistent look across the whole film?
Lock a three-colour palette, one lens feel, and one lighting logic, then repeat those phrases in every prompt. Grade the final cut in one pass so small variations disappear.
What if I cannot get a shot to work at all?
Redesign it. Generation is not a rendering farm; it is a collaborator with strong preferences. If a low-angle tracking shot of a running child will not stabilise, change to a static wide and let the child run through frame.
The Repeatable Workflow in One Page
Write two sentences of story. Break them into five to eight beats. Turn each beat into one to three shots. For each shot, write a four-layer prompt: subject and action, camera and framing, light and lens, continuity anchors. Generate in batches of four to six takes. Log what you kept and why. Assemble in a timeline with scratch sound. Cut once for pace before you fix anything visual. Then regenerate only the shots that the cut proves you need.
That loop is unglamorous, and it is the difference between an impressive folder of clips and a short film someone will actually watch to the end. The tools will keep improving; the discipline of thinking like a director — intention first, execution second — is what stays valuable.
