Cinematic short-form is not a filter you apply at export. It is a stack of decisions made before a single frame exists: what the shot is for, where the camera stands, how the light falls, when the cut lands. Generative video tools have made raw footage almost free, but they have not made those decisions for you. The result is a flood of technically clean clips that feel weightless — pretty images with no point of view.
This guide is a workflow, not a toolkit tour. It covers the directing vocabulary worth internalizing, a short pre-production pass that prevents most rewrites, consistency techniques for faces and style, how to choose between generation modes, sound and edit passes, and a troubleshooting list for the failures that show up again and again. It stays deliberately tool-agnostic, because the principles outlive every model release.
Why direction is the real bottleneck
For a while, the hard part of AI video was fidelity. Faces melted, hands multiplied, motion smeared. That gap has narrowed fast. Modern text-to-video and image-to-video models routinely produce frames that hold up on a phone screen and often on a monitor. When the technology stops being the constraint, the constraint moves somewhere else: intention.
A viewer scrolling a feed gives you roughly one to two seconds to earn the next five. They are not evaluating render quality. They are evaluating whether something is happening, whether a question has been raised, whether the image is composed in a way that suggests a person made a choice. Two clips with identical visual fidelity can perform wildly differently purely because one was shot-planned and the other was prompted.
The practical consequence is that your highest-leverage work happens before generation. A twenty-minute planning pass — concept, beats, shot list, prompt brief — typically removes several rounds of regeneration later. That trade is one of the best available in this medium.
The cinematic grammar worth internalizing
You do not need a film degree, but you do need a small shared vocabulary. These three axes cover most of what makes a shot read as intentional.
Shot size and what it communicates
Shot size is emotional distance. A wide frames a person inside a world; a close-up frames a world inside a person. Shorts compress time, so moving between the two is how you signal escalation without dialogue.
- Extreme wide — scope, isolation, establishing geography. Use once at the top or at a reveal.
- Wide — context plus subject. The default opener for most shorts.
- Medium — conversation and gesture. Reliable, unremarkable, useful.
- Close-up — decision, emotion, detail. The workhorse of a payoff moment.
- Insert — a hand, a screen, an object. Cheap, high-information, superb for transitions.
A useful rule for a 30-second piece: one establishing shot, two to four medium or close shots carrying the story, one insert for texture, one final image that resolves the question. That is six shots, which is a manageable generation budget and a rhythm the viewer can follow.
Lighting as a narrative tool
Prompt-level lighting language does more heavy lifting than most creators expect. Instead of "beautiful lighting," describe a source and its behavior: "single window light from camera left, soft falloff, subject two stops brighter than the wall behind them." That phrasing gives a model something to solve.
Three patterns cover most needs:
- Motivated single source. One light with an in-world justification — a lamp, a window, a phone screen. Reads intimate and credible.
- Backlight plus fill. Subject rimmed from behind, gently filled from the front. Reads bold and separated from the background, ideal for silhouettes and night exteriors.
- Flat high-key. Even, bright, low-contrast. Reads commercial, instructional, and comedy-friendly.
Consistency matters more than sophistication. If shot one is motivated window light, shot six should not suddenly be neon noir unless the story justifies the shift.
Camera movement and when to use it
Movement should have a reason. The three safest choices are a slow push in (increasing tension or intimacy), a lateral tracking shot (revealing space or following action), and a static locked-off frame (letting performance or composition carry the moment). Handheld drift is a fourth option, but it is easy to overuse and hard to match between generated shots.
One practical tip: pick one dominant movement per short and let it recur. Repetition of a camera behavior reads as style. Random variation reads as noise.
The pre-production pass that takes twenty minutes
Step 1: Write the beat sheet
Before prompts, write four to six beats in plain sentences. A beat is a change, not an action. "She notices the door is open" is a beat. "She walks down a hallway" is filler.
A workable beat structure for a 30-second short:
- Hook — an unresolved image or line. No setup.
- Question — what is at stake, shown not explained.
- Complication — something goes wrong or gets strange.
- Turn — a reversal, a reveal, a choice.
- Payoff — the image that answers the hook.
Step 2: Convert beats into a shot list
Every beat gets one to three shots. Assign each shot a size, a lighting note, a movement note, and an estimated duration in seconds. Durations should sum to your target length minus the time you want for the title and end frame.
Step 3: Write the prompt brief
The brief is a compact translation of the shot list into model-readable language. Keep every entry in the same order so you can diff them quickly when something drifts:
| Field | Example |
|---|---|
| Subject | woman in her 30s, wet coat, dark bob |
| Action | slowly turning toward a doorway |
| Shot size | medium close-up |
| Camera | slow push in, 35mm feel |
| Lighting | single window source, camera left |
| Palette | desaturated blue with warm interior accent |
| Duration | 4 seconds |
Writing briefs this way means that when a shot fails, you know which field to change instead of rewriting the whole prompt and losing a good result.
Consistency: keeping faces, wardrobe, and grade stable
Continuity failure is the most common reason a planned sequence feels amateur. A character's jawline shifts, a jacket changes color, the grade jumps from teal to amber between cuts. Fix the anchors, not the symptoms.
Build a character reference sheet
Generate a small set of stills first: a neutral front-facing portrait, a three-quarter view, and a full-body frame in the target wardrobe. Pick the two best and treat them as the canonical references for every subsequent shot. Reference-driven generation modes that accept multiple images will hold identity far better than text alone.
Lock the style with one anchor frame
Choose one image that represents the look you want — grade, grain, contrast, lens character — and attach it as a style reference wherever the tool supports it. This single habit does more for visual cohesion than any prompt adjective.
Control drift with a fixed vocabulary
Write your lighting, lens, and palette descriptors once, then reuse them verbatim across all shots. Paraphrasing introduces variance that models interpret literally. A shared phrase list is your continuity department.
Choosing a generation mode
Text-to-video
Best when the shot is atmospheric, simple in action, and does not contain a returning character. Landscapes, textures, product hero shots, abstract transitions. It is the fastest and least controllable mode, which makes it ideal for exploratory passes and b-roll.
Image-to-video
Best when composition is already decided. Generate or photograph a still, then animate it. You retain control of framing, and the model focuses on motion. This is the mode most creators should default to for story shots, because a strong still is easy to evaluate and cheap to iterate before you spend generation time on motion.
Reference-driven and multi-image generation
Best when identity or style must persist. Feeding several reference images lets the model triangulate a face, an outfit, or a look. It is slower and fussier, but for a recurring character across six shots it is the only mode that reliably survives a cut.
A practical hybrid: use reference-driven generation for any shot containing your protagonist, image-to-video for supporting shots, and text-to-video for inserts and textures. This split keeps spend and iteration time pointed at the shots the audience actually studies.
Sound: the half of cinema that people skip
Silent shorts feel like demos. Add three layers and they feel like scenes.
- Ambience — a continuous bed that implies place. Room tone, street hum, rain, wind. Even at low level it removes the "generated" feeling.
- Foley and accents — footsteps, cloth, a door latch, a glass set down. These sync to visible action and create the illusion of physical weight.
- Music — one idea, not a playlist. If the piece has a turn, the music should change at the turn, not run continuously underneath.
Mix conservatively: dialogue and key accents forward, ambience low but present, music ducking under any spoken line. On phones, low frequencies vanish, so make sure your critical information sits in the midrange.
Edit, finish, and deliver
Cut for rhythm before you cut for beauty. Watch your assembly with sound and ask where you get bored — that is where a shot is two seconds too long. Shorts rarely suffer from being too fast; they frequently suffer from a slow first two seconds.
Finishing checklist:
- Hook in the first second. Start inside the moment, not before it.
- Cut on motion rather than at rest. Motion hides the seam.
- Match the grade across shots with a shared adjustment layer or lookup.
- Captions inside the safe area and legible at 50% brightness, since many viewers watch muted.
- Vertical, square, and horizontal masters exported from the same timeline, not re-cropped by hand.
- End frame that resolves the opening image, reinforcing the loop.
Troubleshooting the failures you will hit
Identity drift across shots. Shorten the shot, increase reference strength, and reduce how much the camera moves. Fast movement plus a low reference weight is the classic recipe for a changing face.
Flicker and texture shimmer. Usually a sign of an overly literal prompt describing fine detail — fabric weave, hair strands, foliage. Simplify the description and let the model fill in.
Morphing hands and props. Frame them out. An insert shot of an object held off-screen is cleaner than a compromised close-up of a hand.
Uncanny motion. Reduce action complexity. One clear verb per shot outperforms three chained verbs, and slower motion reads as more professional than fast motion.
Muddy or dimensionless light. Specify direction and ratio. "Soft light" alone gives you flat light; "soft key from camera left, subject brighter than background" gives you shape.
Grade jumps between cuts. Attach the same style reference or grade every clip in a sequence, and check shots side by side rather than one at a time.
A quality-control checklist and iteration cadence
Run this before publishing, every time:
- Does the first second create a question?
- Can a viewer describe the premise in one sentence without sound?
- Does every shot have a stated purpose in the shot list?
- Is the character recognizably the same person throughout?
- Does the light direction stay consistent between adjacent shots?
- Is there an ambience bed under the whole piece?
- Are captions legible and clear of the platform UI?
- Does the last frame close the loop opened by the first?
On cadence: batch your work. Generate all stills in one session, then all motion, then edit, then sound. Context switching is the hidden cost in this workflow, and grouping tasks by type typically cuts total production time noticeably compared with finishing one shot end to end before starting the next.
FAQ
Do I need a shot list for a 15-second clip?
For a single-shot piece, no. For anything with a cut, yes — even a three-line list. It takes two minutes and prevents the most common failure, which is a sequence with no shape.
How many shots should a 30-second short have?
Between four and eight. Fewer than four and the piece feels static; more than eight and no shot gets enough time to register.
Is image-to-video always better than text-to-video?
Not always. Text-to-video is faster and better for atmosphere, textures, and exploratory work. Use image-to-video whenever composition or continuity matters.
How do I stop a character looking different in every shot?
Fix two reference stills, use a mode that accepts multiple reference images, keep the camera movement modest, and repeat the same descriptive vocabulary in every prompt.
Should I generate a whole sequence at once?
Generate the establishing shot and one close-up first, watch them back to back, and only then produce the rest. Early continuity checks save entire regeneration passes.
What if the model keeps producing something good but off-brief?
Save it. A usable off-brief shot often becomes the seed for the next piece. Keep a folder of these; it is effectively a free shot library.
How long should each shot be?
Two to five seconds for most shorts. Under two and the viewer cannot read the frame; over five and the piece starts to drag unless the shot contains genuine change.
The through-line is simple: decide what each shot is for, lock your anchors, generate in the right mode for the job, and treat sound and the edit as part of the film rather than cleanup. Models will keep improving. Taste, structure, and continuity remain the parts you supply.
