Start With the Ad Format, Not the Model
Most expensive AI video mistakes begin in the same place: someone opens a generation tool, types a poetic prompt, and only later asks which channel the video is actually for.
Advertising and promo work is format-driven. The format dictates duration, aspect ratio, hook timing, shot count, caption placement, and how much story you can carry. Decide the format first and the production plan almost writes itself.
| Format | Duration | Aspect | Shots | Primary constraint |
|---|---|---|---|---|
| Bumper / pre-roll | 6s | 16:9 or 1:1 | 2-3 | One idea, zero setup |
| Social vertical | 9-16s | 9:16 | 4-6 | Readable with sound off |
| Standard spot | 30s | 16:9 | 8-12 | Narrative arc |
| Product hero loop | 3-8s | Any | 1-2 | Product fidelity |
| Explainer | 45-90s | 16:9 | 10-20 | Information density |
| UGC-style testimonial | 15-30s | 9:16 | 3-5 | Perceived authenticity |
Two consequences follow immediately. First, generative shots should never carry the whole film. A typical 30-second spot needs a real product plate, a logo end card, legal lines, and often a live-action spokesperson. Second, shot count is a budget decision. Twelve generated shots means twelve review cycles, twelve chances of warped geometry, and twelve clips to color-match. If the format only supports four shots, resist the urge to generate twelve.
Write the format on a single line at the top of the project folder: 30s spot, 16:9, sound-on, three product beats, one human moment, logo end card. That line is the contract. Everything below either serves it or gets cut.
Build a Creative Brief a Generation Pipeline Can Actually Use
Traditional creative briefs are written for humans who infer. Generation pipelines do not infer, so the brief has to become unusually explicit.
A workable brief contains seven blocks:
- Single promise. One sentence a viewer could repeat after one viewing.
- Audience and context. Where they watch, what they already know, what they scroll past.
- Tone, translated. Not "premium" but "slow camera, wide negative space, muted palette, no handheld."
- Mandatory elements. Logo lockup, legal disclaimer, product name spelling, offer end date.
- Reference set. Five to eight frames, clips, or stills with a note about what to borrow from each.
- Forbidden elements. Competitor colors, clichéd imagery, gestures, anything legally risky.
- Deliverable list. Every aspect ratio, duration, and language variant required on delivery.
The translation step is where most briefs break. Adjectives are not instructions. Turn each one into observable visual attributes before anyone opens a tool.
From adjectives to observable attributes
| Brief word | Observable translation |
|---|---|
| Premium | Slow dolly, shallow depth of field, restrained palette, no lens flares |
| Energetic | 1-2 second cuts, whip pans, saturated accents, percussive sound design |
| Natural | Available-light look, soft shadows, minimal grade, ambient audio |
| Trustworthy | Steady framing, eye contact, clean typography, no exaggeration |
| Playful | Slight overcranking, bright primaries, snappy transitions, tactile props |
Keep this table in the project folder. When a reviewer says a cut "doesn't feel premium," you can point at four concrete attributes instead of arguing about vibes.
Storyboard and Shot-List Before You Generate Anything
Generative tools reward preparation and punish improvisation. A shot list created before generation reduces both timeline cost and re-render count.
Build the shot list as a spreadsheet with these columns: shot number, duration in frames, framing, subject, action, camera move, lighting, audio intent, and generation method. Fill every column. "Camera move: none" is a valid and useful entry.
The beat sheet comes first
Before shots, write beats. A 30-second spot might run: disruption, product reveal, proof, call to action. Each beat gets a time range. Then shots get assigned inside beats. This prevents the classic failure of a beautiful opening shot that consumes nine seconds of a fifteen-second film.
Animatic with stills, not clips
Assemble the shot list as a slideshow with rough timings and a scratch voiceover. Reviewing a twenty-second animatic costs minutes. Reviewing twenty seconds of generated footage costs hours and payout in rework. Fix pacing, ordering, and clarity in the animatic stage.
Mark the shots that must not be generated
Identify product close-ups, packaging text, faces of real employees, price figures, and legal text. Those shots belong to real photography, screen capture, or motion graphics. Generative video is strongest for environment, mood, transitions, texture, and abstract b-roll. Assigning it the wrong job is the single most common cause of wasted production time.
Prompting for Ads: Framing, Motion, and Product Fidelity
Prompts for advertising need more structure than prompts for experimentation. A reliable anatomy looks like this:
Subject → action → environment → camera and lens → lighting → grade → motion behavior → duration.
A worked example for a beverage brand's lifestyle insert:
A chilled glass bottle of sparkling citrus water on a weathered wooden table, condensation running down the glass, condensation droplets catching light, shot on a 50mm lens at close range, soft window light from the left, gentle handheld drift, cool neutral grade with warm highlights, shallow depth of field, background of blurred greenery, slow push in, four seconds.
Notice what the prompt avoids: brand names rendered as text, hand interaction, and the product label facing camera. Those are fidelity traps. Text on packaging and fine geometry are still weak points for generative video, and an advertising deliverable cannot ship with a melted logo.
Working with product plates
When the product must be accurate, generate the environment and composite the product. Practical sequence: generate a clean plate of the setting with deliberate space where the product will sit, shoot or render the product separately, then combine in the edit with matched perspective, shadow, and grain. This hybrid approach is faster than fighting for a perfect single generation and far more brand-safe.
Negative prompts that earn their keep
Useful exclusions for ad work include: extra fingers, distorted hands, unreadable text, warped packaging, duplicated limbs, flickering, jittery motion, watermark, oversaturated skin, fisheye distortion. Keep the list short and specific. A wall of negatives dilutes the prompt and often removes the very motion you wanted.
Iterate on one variable at a time
Change the camera move, keep everything else identical. Change the lighting, keep the camera. This sounds slow, but it is the fastest path to a repeatable house look, and it lets you document prompts that actually work for the campaign.
Match Each Shot to the Right Generation Approach
Not every shot deserves the same method. Choose deliberately.
| Shot need | Best approach | Why | Main risk |
|---|---|---|---|
| Establishing environment | Text to video | Fast, varied, easy to restyle | Generic look |
| Character performance | Image to video with a consistent reference | Maintains identity across shots | Face drift |
| Product detail | Real footage or render, then composite | Exact geometry and label | Perspective mismatch |
| Abstract transitions | Text to video, short duration | Texture and motion carry it | Overuse |
| Restyle of existing footage | Video to video | Keeps original motion and framing | Cost per second |
| Titles, offers, legal | Motion graphics in the editor | Perfect typography | None worth noting |
| Testimonial | Real capture | Authenticity cannot be faked reliably | Production time |
A practical ratio for a 30-second spot: roughly half real or rendered assets, a quarter generated atmosphere, and a quarter graphics and typography. Campaigns that push generation above seventy percent tend to show a recognizable synthetic sheen that audiences read as an ad they can skip.
Decision criteria in plain language
Ask three questions about each shot. Does it need exact product truth? Then it is not generated. Does it need a recognizable human? Then use a reference image or real capture. Does it exist only to create mood, motion, or transition? Then generation is the efficient choice.
Assembly: Editing, Sound, and Brand Consistency
Generated clips are raw material, not a film. Assembly is where the ad becomes an ad.
Edit for the hook
Modern feeds reward the first 1.5 seconds. Build the first cut so the promise or the most visually arresting frame lands before any logo. Many teams place the brand reveal at the end, which is correct for the end card, but the opening frame still has to earn attention on its own.
Match shots to each other
Synthetic shots rarely share a look. Normalize them: unify color temperature, lift or crush contrast toward a single reference frame, add the same subtle grain, and check motion blur direction. A shared look-up table and a single grain pass do more for cohesion than any prompt trick.
Sound design does most of the persuasion
Promo videos live or die on audio. Layer three elements: a bed (music or ambience), spot effects (whooshes, impacts, cloth, liquid), and voice. Keep the voiceover at a consistent loudness, cut music on the call to action, and design for sound-off viewing with burned captions that sit inside safe zones.
Brand system checks
Verify logo clear space, minimum logo size, approved color values, and typography in the end card. Confirm the legal line is legible at the smallest delivered resolution. These details break more campaigns than any visual artifact in a generated shot.
Quality Control: The Checklist That Prevents Rework
Review every clip at full resolution and on a phone screen before it enters the timeline. Phone review catches legibility problems that a large monitor hides.
- Anatomy: hands, fingers, limbs, teeth, eyes.
- Text: packaging, signage, screens, captions, legal lines.
- Geometry: product shape, bottle necks, labels, edges, wheels, architecture lines.
- Continuity: wardrobe, props, light direction, weather, time of day.
- Temporal artifacts: flicker, morphing, popping, sudden resolution shifts.
- Motion: speed ramps, unnatural acceleration, floating objects.
- Audio: sync drift, clipped dialogue, hum, inconsistent ambience.
- Compliance: claims, disclaimers, regional restrictions, talent consent.
Score every item pass or fail. A clip with two fails goes back for regeneration rather than into an edit you will later rebuild. Keep a short artifact log per clip so a reviewer's note maps to a specific shot and timestamp.
Versioning and Deliverables Across Channels
One master timeline, many outputs. Build the longest, widest version first — typically 16:9 at the maximum duration — then derive shorter and vertical cuts from it.
A clean versioning scheme
Name files by campaign, asset, ratio, duration, and language: spring-citrus_master_16x9_30s_en. Never rely on "final_v3_reallyfinal." Versions multiply quickly once you add 9:16, 1:1, 6-second bumpers, and localized captions.
Hook variants, not full remakes
Produce three opening variants against one shared body. Test different first frames, different first lines of voiceover, and different pacing. This gives meaningful creative learning for a fraction of the cost of three separate productions.
A two-week rhythm that works
Days one and two: brief, beats, shot list, animatic. Days three to five: generate and review in batches of four to six clips. Days six and eight: assembly, sound, and grade. Days nine and ten: quality control, versions, captions, and delivery. Keep a buffer day; regeneration requests always appear late.
Ten Mistakes That Sink AI Promo Videos
- Generating the product instead of compositing it. Labels and fine geometry warp; brand teams will reject it.
- Writing the shot list after the footage exists. Pacing problems become unfixable.
- Skipping the animatic. The cheapest stage is the one teams skip.
- Chasing a perfect single generation. Composite two good pieces instead.
- Using one prompt style across all tools. Each model rewards different phrasing.
- Ignoring sound until the end. Half of perceived quality is audio.
- Leaving captions outside safe zones. Vertical platforms crop aggressively.
- Overloading the first second with the logo. Attention is spent before the hook lands.
- Skipping color unification. Mixed grades read as amateur assembly.
- No artifact log. The same defects get re-approved in the next review round.
Each mistake has a cheap preventive action. Print the list, tape it to the production checklist, and cross items off during pre-production rather than during the final review.
Frequently Asked Questions
How many generated shots should a promo video contain?
For a 15-second vertical, three to five. For a 30-second spot, five to eight, mixed with real or rendered product assets. Above that ratio, cohesion and product accuracy both deteriorate.
Can generative video replace product photography?
Not for hero product shots. Use it for environment, mood, and transitions, and reserve real capture or 3D rendering for anything where the label, geometry, or material finish must be exact.
What should be decided before the first prompt is written?
Format, duration, aspect ratio, the single promise, mandatory elements, and which shots are forbidden from generation. Teams that answer these six questions first report far fewer regeneration cycles.
How do you keep a consistent look across many clips?
Lock a reference frame, then match color temperature, contrast, grain, and lens character to it in the edit. Keep a written record of the prompt structure that produced the reference so later shots can be recreated.
Is it worth generating multiple hook variants?
Yes. Opening variants cost little because the body is shared, and the first two seconds drive most of the performance difference in a feed.
What is the biggest time saver in an AI ad workflow?
The shot list plus animatic. Reviewing pacing as a slideshow with scratch audio catches structural problems before any generation time is spent.
How do you handle localization?
Design the master with text-free plates wherever possible. Add typography and captions in the editor per language so a single visual master serves every market without regeneration.
What should be checked on a phone before delivery?
Caption legibility, logo size, legal line readability, safe-zone cropping, and whether the hook lands within the first two seconds.
