A practical workflow for producing a cinematic AI product review with consistent B-roll, shot plans, prompt craft, sound design, and edit pacing.
A product review video is decided by the shots between the talking. The close-up of a knob turning with a faint click, the slow drift across brushed aluminium, the hand lifting the device off a workbench and turning it toward the light — that is the B-roll that makes a viewer believe the verdict. Generative video can now produce those shots deliberately rather than by accident, but only if you run the job like a small film production instead of a slot machine.
This guide walks through the whole workflow: scripting the review spine, designing a consistent product look, prompting B-roll that actually cuts together, building presenter segments, and finishing with sound and pacing. It is written for creators who want a cinematic product reel — the kind that reads as deliberate craft rather than a demo of a model's quirks.
Why the B-roll carries the review
Trust in a review comes from evidence. A viewer wants to see the hinge travel, the port layout, the texture of the grip, the way the lid closes. A talking head asserts; an insert proves. Cinematic B-roll also does a second, quieter job: it covers the cuts where your argument changes direction, so the edit feels continuous even when the reasoning jumps.
Generative video is unusually strong at exactly this class of shot. Short, object-focused, controlled lighting, no dialogue, no second character — a four-second dolly across a product is a far easier generation problem than a twenty-second scene with two people talking. That mismatch between what models do well and what reviews need is the entire opportunity. You are not trying to generate a film. You are generating twenty small, precise fragments that a human editor assembles into a film.
Build the review spine before you prompt anything
Most weak AI product reels fail at the script stage, not the generation stage. If you cannot say what each shot is proving, no prompt will save it.
Write five to seven beats
A review is an argument with a shape. Typical beats: the claim (what this product is and who it is for), the context (what it replaces), the first proof moment (the feature that actually matters), the friction (what annoys you), the comparison, and the verdict. Write one sentence per beat. Everything else — every generated clip — attaches to one of those sentences.
Turn the beats into a three-column shot plan
This is the single most useful document in the project. Column one is the beat. Column two is the proof shot: what the camera must physically show for that beat to land. Column three is the generated B-roll that fills the gaps around it.
| Beat | Proof shot | Generated B-roll |
|---|---|---|
| Build quality claim | Macro of the seam between panels | Slow push-in on textured metal edge, shallow depth of field |
| Battery complaint | Hand swapping the pack | Top-down insert of the pack seating with a click |
| Verdict | Product on a desk, wide | Slow orbit at eye level, warm window light behind |
The proof shots you can film for real should be filmed for real. Save generation for the shots that are expensive, impossible, or simply not worth a lighting setup — which, for most creators, is most of them.
Designing a consistent product look
The hardest problem in AI product video is not realism. It is consistency. A device that changes finish between shots destroys the illusion faster than any artifact.
Lock a hero frame first
Generate a still image of the product before you generate a single second of video. Treat that still as your reference plate: it defines silhouette, colour, logo placement, material sheen, and lens character. Feed it into every video prompt as an image reference, and describe it in words too — "matte graphite body, single silver ring around the lens, no visible branding on the front face." Words plus reference beat words alone.
If you are building that plate from scratch, start with the AI image generator and iterate on the still until it matches the product in your hand. Ten minutes on the still saves an hour of regenerating video.
Speak in lighting language, not adjectives
Models respond to cinematography vocabulary far better than to mood words. "Beautiful lighting" produces mush. "Large soft source at 45 degrees camera left, subtle rim from behind right, dark grey seamless background, 3:1 ratio" produces a repeatable look you can carry across twenty clips.
Keep a short style block that you paste into every prompt in the project:
50mm macro, shallow depth of field, soft key camera left, cool rim light, matte background, subtle film grain, no text overlays.
That block is your visual contract. Change one word and you have a different film.
Prompting B-roll that actually cuts together
Individual clips can look gorgeous and still be unusable, because they do not cut. Editing rhythm is designed at the prompt stage.
Give every clip a motion budget
One motion idea per clip. Pick one: a slow push-in, a lateral dolly, a rack focus, a static frame with the product rotating, a hand entering frame. When you stack an orbit plus a zoom plus a product rotation, the model has to invent geometry, and geometry is where artifacts live.
Plan pairs, not singles
Editors think in pairs. Generate clips that can be cut against each other: a wide and its matching insert, a push-in and its reverse pull-out, a left-to-right dolly and a right-to-left answer for the return. If you generate twelve unrelated clips, you will spend the edit hunting for anything that matches.
Keep clips short
Three to six seconds is the sweet spot. Long generations drift, soften, and invent details. Short clips are also cheaper to discard — and you will discard a lot. A three-minute review typically needs eighteen to twenty-five usable B-roll moments, which means generating somewhere between sixty and a hundred attempts.
The five-part prompt template
A reliable product prompt answers five questions in order. Once the pattern is memorised, you can write a new one in under a minute.
- Subject and state — what it is and how it sits in the world: "a compact mirrorless camera body resting on a concrete workbench, lens attached."
- Surface and material truth — "matte black magnesium alloy, fine grain texture, faint fingerprints near the grip."
- Action — the smallest possible movement: "a hand enters from the right and rotates the body ten degrees."
- Camera — format, lens, and movement: "macro, 50mm equivalent, slow push-in from the front left, shallow depth of field."
- Light and grade — "soft key from camera left, cool rim, neutral grade, mild contrast, subtle grain."
Read the finished prompt back and ask one question: could a camera operator shoot this on a table in one take? If not, simplify until they could.
Negative constraints that fix real problems
Negatives are not decoration; they are corrections for failures you have already seen. The recurring ones in product work: drifting or duplicated logos, text that turns into gibberish, extra fingers, warped straight edges, physics-defying reflections, and the sudden zoom that models like to add unasked. Add those as explicit exclusions, and re-add them in every batch — a clean prompt library like the prompt library keeps the good strings from being lost between projects.
Building the A-roll around the B-roll
Presenter segments are what turn a mood board into a review. Keep them boring on purpose: static camera, consistent framing, consistent wardrobe, same background across sessions. The interesting part of the frame should be the product, not the setup.
Record your voice separately from a clean microphone and cut picture to it. This gives you the freedom to rewrite a sentence in post without reshooting, and it keeps audio consistent when your A-roll is captured across multiple days. For demo beats where you would normally reach for a camera, generate the cutaway instead and keep your hands out of the shot.
Once your first sequences exist, the video templates can carry the open, chapter cards, and end plate so you are not rebuilding graphics for every upload.
Sound and the pacing pass
AI B-roll almost always arrives silent, and silence is what makes generated footage feel synthetic. Three layers fix it cheaply:
- Foley — a click for the dial, a soft thud for the device landing, fabric for movement. Even subtle samples anchor a shot in physical space.
- Room tone — a continuous low bed under the whole edit so cuts do not create dead air.
- Music — one track, restrained, ducked under narration.
Then do a pacing pass with the sound off. Watch the cut muted and count how long each shot stays on screen. Anything under about a second does not register as information; anything over six seconds in a review starts to sag unless there is movement. Trim until the rhythm feels like a conversation.
A realistic three-day production schedule
Day one — script and plates
Write the beats, build the shot plan, generate the hero still, and lock the style block. Do not generate video yet. Ending day one with a document and one good image is a success.
Day two — generation sprints
Work in batches of eight to ten clips against a single beat, review them at half speed, and keep only the ones that match the style block. Name every file with its beat and prompt so the edit does not become archaeology. Expect a hit rate around one in three; plan for it and the failures stop feeling like setbacks.
Day three — edit, sound, finish
Assemble the spine, drop in B-roll, add foley, colour-match the clips so the blacks and whites agree, and export. If a shot is 80 percent right, fix it with a crop or a speed ramp rather than regenerating.
Mistakes that flatten AI product reels
- Generating before planning. The most common and most expensive error.
- Changing the product description mid-project. Consistency lives in repeated language.
- Over-long generations. They drift. Cut them short.
- Too many camera moves per clip. One idea, one movement.
- Mixing photoreal and stylised clips. Pick a register and hold it.
- Ignoring the reveal. Viewers need a beat to register what they are looking at before the cut.
- No texture. Perfect surfaces read as renders; a little dust, a fingerprint, or a scuff sells the object.
Pre-publish checklist
- Every beat has a proof shot.
- The product's colour and logo are identical across all clips.
- No clip exceeds six seconds without a reason.
- Foley is present on every object interaction.
- Blacks and whites match between generated and filmed footage.
- No unintended on-screen text survived generation.
- The first eight seconds state the claim.
- The verdict is spoken, not implied.
FAQ
Do I need the physical product to make this work?
For an honest review, yes — you need to have used it. The generated B-roll supports your first-hand observations; it does not replace them. If you are producing a preview or a concept video rather than a review, say so on screen.
How many clips does a three-minute review need?
Roughly eighteen to twenty-five usable B-roll moments, which usually means generating sixty to a hundred attempts. Presenter segments and screen recordings cover the rest.
Why does my product change shape between shots?
Almost always because the description changed. Freeze a single subject paragraph and a single style block, reuse them verbatim, and use the hero still as an image reference in every prompt.
Can I use AI B-roll in a sponsored video?
Yes, but disclose it and keep the claims truthful. Generated shots should illustrate real characteristics of the product, never invent features or imply performance you have not tested.
What resolution and aspect ratio should I generate at?
Match your delivery platform: vertical for short-form, wide for long-form. Generating once at the correct aspect ratio beats cropping later, because cropping throws away the framing you composed.
How do I stop the model from adding camera moves I did not ask for?
State the camera behaviour explicitly ("static locked-off frame, no zoom, no orbit") and add the unwanted movement to your negative list. Repeat both in every prompt of the batch.
Take the idea into motion
A cinematic product review is not a model showcase. It is a shot plan, a style contract, and a patient edit — with generation doing the heavy lifting on the twenty small moments that would otherwise take a week of lighting setups. Start with the hero still, write the style block once, and generate in short, single-idea clips. When you are ready to put the workflow into practice, generate your first clips in Orelon and build the reel beat by beat.

