Text-to-video generation has collapsed the distance between a written idea and a moving image. A short paragraph can become a shot with camera movement, weather, and a performance. The gap between a striking demo clip and a finished short film is still wide, though, and it is filled with craft: shot lists, prompt discipline, consistency tricks, sound design, and editing rhythm. Those are the things that separate a random folder of generated clips from something an audience will watch to the end.
This guide covers the full pipeline for a one-minute narrative short built mainly with text-to-video models. It stays deliberately tool-agnostic, so you can apply it whether you generate in Runway, Kling, Luma Dream Machine, Pika, Veo, or a local model on your own hardware. The workflow matters more than the brand name on the model.
What Text-to-Video Can and Cannot Do
The model is predicting, not filming
A video model has never seen your location, your actor, or your intent. It has learned statistical patterns linking words to pixels and pixels to motion. When you type "a woman walks through a rain-soaked market at dusk," the model is not building a scene from a script. It is sampling the most plausible-looking continuation of that idea, frame by frame, while trying to keep those frames coherent with each other.
That single fact explains most of the behavior you will see. It explains why lighting drifts mid-shot, why hands merge into objects, why a character's jacket changes color between cuts, and why the model sometimes reaches for the most generic version of a scene instead of the specific one you imagined. You are not directing a camera. You are steering a probability distribution, and steering requires specificity.
Limits worth designing around
Before you write a single prompt, accept four practical constraints.
- Shot length. Most models produce a few seconds of usable motion per generation. Longer outputs tend to accumulate drift. Design in short units and cut more.
- Temporal consistency. Details that appear once and disappear are cheap. Details that must persist across a shot — a scar, a specific jacket, a logo — need reinforcement through reference images or careful re-prompting.
- Physical causality. Complex interactions such as a door opening while someone carries two bags rarely resolve cleanly. Simplify the action, or split it across two shots.
- Text and faces at small scale. On-screen writing usually degrades, and small faces in wide shots lose identity. Keep reading material out of frame, and keep your subject reasonably large.
None of these are deal-breakers. They are production rules, the same way a low ceiling or a tight schedule is a production rule on a physical set. Good short films are built by people who write toward the constraints instead of fighting them.
Pre-Production: From Idea to Shot List
Write a one-page treatment
The fastest way to waste generation time is to start prompting before you know what the film is about. Write a one-page treatment first: who the character is, what they want, what stands in the way, and how it ends. Keep it to five sentences of story and one sentence describing the visual mood.
For a short film under two minutes, resist the urge to add subplots. One character, one want, one obstacle, one turn. A minute-long short with a clean turn feels finished. A minute-long short with three ideas feels like a trailer for a movie nobody made.
Turn beats into shots with duration budgets
Break the treatment into beats, then convert each beat into one to three shots. Give every shot an estimated duration before you generate anything. A typical one-minute short might look like this:
- Establishing wide, 4 seconds
- Character introduced in medium shot, 5 seconds
- Object detail insert, 3 seconds
- Confrontation in two shots, 10 seconds
- Reaction close-up, 4 seconds
- Movement montage, 12 seconds
- Climax, 8 seconds
- Resolution, 6 seconds
This shot list becomes your build order. It also tells you how much total screen time you need to generate, which is far more useful than thinking in terms of "how many clips."
Build a shot bible
A shot bible is a single document that locks down every recurring visual element so you are not reinventing them in each prompt. Include:
- Character sheets: age range, build, hair, wardrobe, distinguishing features, and 2–3 reference images if your tool accepts them.
- Location sheets: layout, time of day, weather, dominant colors, light sources.
- Style tokens: a fixed phrase describing your look, such as "35mm anamorphic, soft haze, muted teal and amber, shallow depth of field."
- Camera vocabulary: the specific terms you will reuse, such as slow dolly in, handheld follow, static wide.
Once these phrases are written down, you copy and paste them instead of improvising. Consistency in a generated film is mostly a documentation problem, not a model problem.
Prompt Engineering for Motion, Not Just Beauty
The most common beginner mistake is writing prompts that describe a beautiful still image. Models are good at beauty. Films are made of motion, and motion needs to be described explicitly.
The five-part shot prompt
A reliable shot prompt has five components, roughly in this order:
- Subject and action — who is doing what, with an active verb. "A teenage girl sprints" beats "a girl running."
- Setting and time — location, weather, and hour. Specific light is more controllable than mood adjectives.
- Camera — shot size, angle, and movement. "Low-angle medium shot, slow push in."
- Lighting and palette — source direction and color. "Backlit by setting sun, warm rim light, cool shadows."
- Style and texture — film stock, grain, lens character, rendering feel.
Written out, a finished prompt might read: "Low-angle medium shot, slow push in on a teenage girl sprinting across a wet rooftop at dawn, backlit by a rising sun, warm rim light against cool blue shadows, shallow depth of field, 35mm anamorphic with light grain." That prompt gives the model a subject, an action, a camera instruction, a lighting plan, and a texture. It has far less room to guess.
Negative prompts are a fixed list, not an improvisation
Keep a saved negative prompt block and reuse it. Typical entries include: extra limbs, distorted hands, warped faces, text artifacts, watermark, jump cut, flickering, oversaturated, plastic skin, blurry background, duplicated subject. Add to the list only when a specific failure repeats.
Version control your prompts
Save every prompt with a short label and note what changed between attempts. When you finally get the shot you want, you will need to reproduce that exact phrasing for the sequel, the reshoot, or the alternate take. Treat prompts like source code. Change one variable at a time when debugging.
Picking the Right Mode for Each Shot
Text-to-video, image-to-video, and video-to-video
These three modes solve different problems, and a good short film usually uses all three.
- Text-to-video is the fastest way to explore. It is ideal for establishing shots, atmosphere, and any moment where precise character identity matters less than energy.
- Image-to-video starts from a still you already approve. Use it whenever a character must look the same across multiple shots, or when a composition must match exactly.
- Video-to-video restyles or extends existing footage. It is the tool of choice for turning a cheap live-action plate into something stylized, or for extending a shot that ended too early.
Decision criteria that actually help
Ask three questions per shot. Does identity need to persist? If yes, start from an image. Does composition need to be exact? If yes, start from an image or a storyboard frame. Is the motion simple and readable? If yes, text-to-video will usually handle it in fewer attempts.
A practical default for narrative work: generate a keyframe image first for every shot featuring your main character, then animate it with image-to-video. Reserve pure text-to-video for landscapes, inserts, and transitional material.
Consistency: Characters, Wardrobe, and Places
Character consistency is the single hardest part of AI short films, and it is solved through redundancy rather than any single trick.
Start with a character sheet image generated in a still model. Approve it. Then use that image as the starting frame for every shot the character appears in, and repeat the same descriptive phrases in each prompt — do not paraphrase them. "Cropped denim jacket, silver hoop earrings, dark curly hair tied back" should appear verbatim in every prompt. Paraphrasing invites drift.
Lock your seeds where your tool exposes them. A fixed seed combined with a fixed prompt gives you reproducibility, which is what makes iteration possible.
Locations need the same treatment. Generate one master wide shot of each location and treat it as canon. Every subsequent shot in that location inherits its palette, its architecture, and its weather. If a scene takes place at night, say so in every prompt for that scene, even if the previous shot already established it.
Finally, plan wardrobe changes deliberately. If you cannot hold a costume consistent across six shots, give the character a reason to change clothes. A cut between two outfits reads as intentional; a spontaneous jacket color shift reads as a mistake.
Sound, Editing, and Finishing
Generated video arrives silent, and silence is what makes AI footage feel artificial. Sound design is where a mediocre sequence becomes convincing.
Build your audio in four layers:
- Ambience — room tone, wind, city hum. This layer does the heavy lifting for realism and should run under almost every shot.
- Foley — footsteps, cloth movement, object handling. Even rough Foley tightens the connection between image and reality.
- Effects — doors, impacts, whooshes. Use sparingly; too many effects make a scene feel like a trailer.
- Music — a single sustained bed or one motif is enough for a short. Music tells the audience how to feel; if your images already do that work, keep the score quiet.
Dialogue is the hardest element to generate well. Two safer alternatives: write the film as a visual piece with no dialogue, or record voice-over yourself and cut the images to your own performance.
In the edit, cut faster than feels comfortable during action and hold longer than feels comfortable on reaction shots. Generated motion often has small imperfections in the middle of a clip. Trimming to the first and last usable frames hides most of them. Add a subtle grade across the whole timeline — one consistent look does more for coherence than any individual perfect shot.
Walkthrough: Eight Shots in One Minute
Here is how the pieces fit together in practice for a one-minute short about a courier delivering a package in a rainstorm.
- Write the treatment. A courier must deliver a package before a building closes; the storm delays her; she runs the last block.
- Build the shot bible. Wardrobe: yellow rain jacket, black backpack. Location: neon-lit downtown street, night, heavy rain. Style: anamorphic, wet reflections, cyan and magenta.
- Generate the character sheet. Approve one image of the courier in the approved wardrobe.
- Keyframe every shot with the character. Use image-to-video with the approved sheet as the base frame.
- Generate atmosphere shots with text-to-video. Rain on asphalt, a flickering sign, a puddle reflecting neon. These are cheap and forgiving.
- Assemble a rough cut. Place all clips on the timeline, trim to the usable range, and check whether the story reads without sound.
- Add sound. Rain ambience across everything, footsteps and splashes under movement, one low synth pad rising into the final shot.
- Grade and export. One unifying look, consistent grain, and a short fade to black.
Total generation time is usually one focused afternoon once the shot bible exists. The preparation is what makes it fast.
Mistakes That Kill AI Shorts
- Chasing a perfect single shot. Do not spend hours on one clip. Get 80 percent and cut away.
- Changing multiple prompt variables at once. You will not know what fixed the problem, and you cannot reproduce it.
- Ignoring eyeline and screen direction. If a character looks left in one shot, they should enter from the right in the next. Editors fix this with mirrored frames, but planning avoids it.
- Too many wide shots. Wide shots hide faces and identity. Medium and close shots are where generated performances actually land.
- No sound pass. Silent AI footage looks like a test render.
- Overloading a single clip with story. One shot should carry one idea.
- Forgetting an ending. A short film needs a final image that resolves the turn, not just a clip that stops.
FAQ
How long should an AI-generated short film be?
One to three minutes is the sweet spot. Long enough to tell a story with a turn, short enough that consistency never collapses.
Do I need to know how to edit?
You need basic timeline skills: trimming, ordering, audio layering, and a simple color pass. That is a weekend of learning, and it is more valuable than any prompt library.
Which is better, text-to-video or image-to-video?
For exploration and atmosphere, text-to-video. For character-driven narrative shots, image-to-video starting from an approved keyframe. Most finished shorts use both.
How do I keep a character looking the same?
One approved character sheet, identical descriptive phrasing in every prompt, the sheet used as the starting frame, and fixed seeds where available. Redundancy beats any single trick.
Why do my shots look great alone but wrong together?
Almost always a palette, grain, or lens mismatch. Apply one consistent grade and one consistent film texture across the entire timeline.
Can I generate dialogue?
You can, but results are inconsistent. Visual storytelling or self-recorded voice-over is faster and more controllable for short films.
How many attempts should a shot take?
Three to six is normal for a clean result. If you are past ten, the prompt is probably too complex — simplify the action or split it into two shots.
What is the fastest way to improve?
Finish something. A completed 60-second short teaches more about prompt structure, consistency, and pacing than a dozen abandoned experiments.
