Why AI Rewrites the Budget, Not the Craft
The cost of acquiring footage has collapsed. A shot that once required a permit, a crew of six, a fog machine, and a golden-hour window can now be produced in an afternoon at a desk. That is genuinely new. What has not changed is the part that decides whether anyone watches the result: judgment about what deserves to be on screen.
This is the trap most creators fall into. They treat a generative model as a vending machine — type a prompt, receive a clip, stitch the clips together, publish. The output is technically moving image and artistically inert. Meanwhile, a creator with a sharp shot list and a modest toolkit produces something that looks like it cost twenty times more.
The useful mental model: AI is a crew, not a director. It can operate a camera, light a set, build a background, cast a face, and score a scene. It cannot decide what the scene is about, which beat matters, or where the cut belongs. Those remain your job, and they are now the entire job.
This guide is structured as a working method rather than a tool list. Tools change every few months; the workflow below — pre-production, model selection, prompting, consistency control, sound, and finishing — survives each new release.
Pre-Production: The 90 Minutes That Saves Ten Hours
Generative work punishes improvisation because every prompt is a small production decision. If you do not know what the shot is supposed to accomplish, you will generate variants until you run out of patience and accept the least bad one. Ninety minutes of planning eliminates most of that churn.
From logline to beat sheet
Start with one sentence: who wants what, what blocks them, what it costs them. Then break that into four to six beats. A 60-second piece rarely needs more than five. Each beat gets a purpose, not a description — "establishes the problem" is a purpose; "woman looking at laptop" is not.
The shot list generation can actually follow
Build a table with columns that map directly onto prompt inputs:
- Shot number and duration. Keep individual shots between two and five seconds. Models hold short clips far better than long ones, and the edit gets more flexible.
- Framing. Wide, medium, close-up, extreme close-up, over-the-shoulder.
- Subject and action. Written in present tense, one action per shot. "She turns toward the window and exhales" is one action; "she turns, exhales, and picks up a phone" is three.
- Camera movement. Static, slow push in, truck left, crane up, handheld drift, whip pan.
- Light and mood. Time of day, key direction, color temperature.
- Continuity flags. Any shot that must match a face, wardrobe, prop, or location from another shot.
The continuity column is the one people skip, and it is the one that causes the most rework.
Look book and continuity bible
Collect five to eight reference images that define palette, contrast, lens character, and wardrobe. Then write a one-page continuity bible: the character's age, hair, clothing layers, key props, the two or three locations, and the lighting rule for each. When a generation drifts — and it will — this document tells you which detail broke.
Choosing a Model Per Shot Instead of One Model for Everything
Most creators pick a favorite model and use it for everything, then complain that dialogue shots look stiff and action shots look mushy. Different models have genuinely different strengths, and matching them to shot type is one of the fastest quality gains available.
What to match
- Motion complexity. Fast action, crowds, and sports need models with strong temporal stability; slow, dialogue-driven shots do not.
- Character performance. Subtle facial work and lip sync are their own specialty. Test a single line of dialogue in three candidates before committing.
- Duration and resolution. Some models cap at four or five seconds natively and degrade when extended. Others produce longer clips at lower fidelity.
- Image-to-video strength. If continuity matters, image-to-video from a locked reference frame almost always beats text-to-video.
- Iteration latency. A slower model with better adherence can be cheaper in total time than a fast model that takes eight attempts.
- Commercial licensing. Confirm the terms before you build a campaign around an output.
Tier your shots
Not every shot deserves the same effort. Sort your list into hero shots — the two or three that carry the story — and supporting shots: inserts, hands, environments, b-roll. Spend your best model and your longest iteration cycle on the heroes and generate the rest quickly. A viewer forgives a plain insert; they do not forgive a hero shot with warping hands.
Prompting Like a Cinematographer
Prompt writing is shot description. The people who get consistently good output are the ones who describe the frame the way a director of photography would on a call sheet.
The five-part prompt formula
Build every prompt from five components, in this order:
- Subject. Who or what, with one or two defining details.
- Action. A single present-tense verb phrase.
- Environment. Location, time of day, weather, depth cues.
- Camera. Framing, movement, lens, height.
- Light and style. Key direction, contrast, color, reference look.
Example: "A woman in her thirties in a charcoal wool coat, walking slowly toward a rain-streaked window, modern apartment at dusk, medium shot slowly pushing in at eye level, 35mm anamorphic, soft window key from the left, deep shadows, muted teal and amber palette."
Every clause does work. Remove the lens and the image flattens; remove the light direction and the model invents a flat frontal key.
Camera vocabulary that models respond to
Terms with consistent results include: static tripod shot, slow dolly in, dolly out, truck left and right, crane up, tilt down, orbit around subject, handheld documentary drift, whip pan, rack focus, shallow depth of field, wide-angle distortion, telephoto compression, low-angle, high-angle, over-the-shoulder. Vague words like "cinematic" are weaker than specific ones, though they can serve as a trailing style anchor.
Negative prompts and guardrails
Use negatives for the failures you have actually seen: extra fingers, warped hands, text artifacts, flickering, duplicate faces, jitter, oversaturated colors, watermark. Do not build a fifty-word negative list preemptively; each entry consumes attention that should go to the description.
Iterate one variable at a time
When a shot fails, change exactly one thing — the camera move, the light direction, the duration. Changing three variables teaches you nothing and usually produces a new, unrelated failure. Keep a log of what worked; a personal prompt library is worth more than any tutorial.
Consistency: The Hardest Problem in AI Video
Continuity is where AI video announces itself. A character whose jacket changes color between shots, or a room whose windows move, breaks the illusion faster than any artifact.
Reference-first workflow
Generate the character once, in a neutral pose and light, and save that frame. Then drive every subsequent shot with image-to-video or a reference mechanism rather than text alone. Multi-image reference features, where a model accepts several frames as anchors, are the strongest tool available for keeping a face stable across angles.
Character sheets and wardrobe locks
Create a sheet with three views: front, three-quarter, profile. Write the wardrobe in fixed, specific language — "charcoal wool coat over cream turtleneck, silver watch on left wrist" — and reuse that exact string in every prompt. Never paraphrase it. Small wording changes produce small visual changes, which accumulate into an inconsistent character.
A continuity checklist between shots
Before you assemble, compare adjacent shots for: hair length and style, clothing layers and colors, prop position, screen direction of movement, light direction, and background landmarks. If the character walks left to right in one shot and right to left in the next, either mirror one shot in post or insert a neutral cutaway.
Sound Design: The Cheapest Way to Look Expensive
Audiences forgive imperfect images far more readily than imperfect audio. A clean mix with believable ambience will make a mediocre shot read as intentional; a hissing, unbalanced track makes good footage feel amateur.
Voice and lip sync
Generate scratch voiceover early — before picture lock — so you can cut to the rhythm of speech instead of stretching speech to fit the picture. When lip sync matters, keep mouths small in frame, avoid extreme close-ups of dialogue, and prefer profile or three-quarter angles where sync errors are less visible.
Ambience, foley, and music
Build three layers: a continuous ambience bed (room tone, street, wind), spot foley for actions that matter (a cup set down, a door), and music. Keep music under dialogue at roughly -18 to -22 dB relative to speech. Silence is a tool too — a half-second of ambience before a reveal does more than a swell.
Basic mixing targets
For web delivery, aim for integrated loudness around -14 LUFS with true peaks below -1 dBTP. Speech should sit clearly above the bed. High-pass rumble below 80 Hz on dialogue tracks; it clears mud without making voices thin.
Post-Production: Where AI Video Gets Exposed
Edit for rhythm, not for clips
AI clips tend to begin and end with dead frames — a beat of nothing before the action starts, a beat after it finishes. Trim aggressively. Cut on motion: when a hand rises or a head turns, that is where the eye expects the cut. Vary shot length deliberately; six consecutive three-second shots feel like a slideshow.
Cleanup
Scan each clip at reduced speed for morphing, dissolving limbs, flickering textures, and edge warble. Fixes: mask and track the artifact with a clean plate from another frame, shorten the shot so the failure lands outside the edit, or regenerate the tail only. Blurring a defect rarely works; it reads as a soft spot in an otherwise sharp frame.
Finishing
Three passes make AI footage feel like camera footage. First, unify: apply a single grade and a subtle LUT across all shots so they share a palette. Second, add texture: a light film grain and slight halation reduce the plastic smoothness that signals generation. Third, resolve: upscale to delivery resolution, sharpen carefully, and check for the ringing that over-sharpening creates. Export at your platform's spec and verify on a phone screen — most viewers will see it there.
A One-Day Workflow for a 60-Second Spot
Hour 1 — Plan. Write the logline, five beats, shot list with continuity flags, and look book.
Hour 2 — Anchor images. Generate the character sheet and two key environments. Lock them. Everything downstream depends on these.
Hours 3-4 — Generate. Produce hero shots first at higher effort, then the supporting shots in batches. Log every prompt that worked.
Hour 5 — Sound. Record or generate voiceover, lay ambience and music, rough mix.
Hour 6 — Edit. Cut to the voiceover rhythm, trim dead frames, insert cutaways where continuity breaks.
Hour 7 — Finish. Cleanup pass, unify grade, grain, upscale, final mix, export, phone check.
This schedule assumes the plan is done and the anchors are stable. If you skip the first hour, expect the day to run to fourteen hours.
Eight Mistakes That Make AI Video Look Cheap
- Long unbroken shots. Fix: split into two to four second pieces and cut.
- Text prompts for recurring characters. Fix: reference images plus a fixed wardrobe string.
- Default lighting. Fix: state key direction and contrast in every prompt.
- No screen direction logic. Fix: map movement on paper before generating.
- Over-detailed prompts. Fix: five components, nothing more.
- Raw generated audio. Fix: rebuild the track in layers.
- Ignoring the phone screen. Fix: check every version on a small display.
- No timecode discipline. Fix: lock durations before editing, not during.
FAQ
Do I need a powerful computer? Not for generation if you use hosted tools, but editing and upscaling benefit from a machine with a decent GPU and fast storage. Proxy workflows keep older hardware usable.
How many generations does a good shot take? Expect five to fifteen for a hero shot and two to four for supporting shots once your prompt library is mature. Early on, double those numbers.
Can I mix AI footage with real footage? Yes, and it usually improves the result. Real inserts give the eye an anchor. Match grain, black levels, and lens character in the grade.
What resolution should I output? Match the platform. 1080p vertical for short-form, 4K horizontal for brand and broadcast work. Upscale late, after the edit is locked.
How do I keep a series consistent across episodes? Freeze the look book, character sheets, wardrobe strings, and grade settings as a reusable project template. Treat it like a series bible and version it.
Is any of this a substitute for storytelling? No. It removes the cost of production, which means the only remaining differentiator is whether the story works. That raises the bar rather than lowering it.
