"AI Video Prompts for B-Roll: How I Get Clean Coverage Models Won't Dream Up"

A client launch video, two weeks ago: the A-roll interview was fine, the music was fine, and the one shot I actually cared about — a taxi idling through a wet Brooklyn side street — came back with the car sliding across the asphalt like it was on black ice. I watched it four times before I believed it. That one physics break was enough to make a viewer doubt every other frame in the film. That's the thing about b-roll: it's where AI video still stumbles, and it's where most of your edit actually lives.

Why most generated b-roll looks "AI"

It's not the quality. It's the consistency. Run the same scene prompt three times and you get three different cars, three different skies, three different years. Nobody complains about that in an A-roll — a talking head has a face, so the model anchors. B-roll has no anchor, so the model improvises every frame.

The second tell is energy. Real b-roll is about coverage density, not spectacle. I've shot coffee bars in East London where nothing "happens" — and the footage is still usable, because there's a grinder, a steam wand, a timer, a hand. Generated b-roll skips the boring middle: every clip is either a dramatic push-in or a static postcard. That gap between "postcard" and "coverage" is exactly what you have to write into the prompt.

The five shot types I actually use

Everything else is decoration. These five cover a product film, a location feature, or a travel piece:

1. Wide establishing — the whole scene, nothing moving except ambient stuff

2. Medium action — one person doing one thing, frame from the waist up

3. Close texture — a surface, a material, a hand, macro-ish

4. Detail macro — the tiny moment that doesn't make sense on its own

5. Transition insert — light, shadow, motion blur, something the cut can land on

Each one gets its own prompt. Never "cinematic shots of a city street." That sentence is what produced the sliding taxi.

Prompt patterns per shot

Wide establishing. What I write: "Wide static shot from a fixed tripod position, slow morning light moving across a quiet side street in Brooklyn, brick facades, parked vehicles, a bus passing in the far background, gentle haze, 35mm look, natural color, no people in the foreground, camera remains locked off for the entire clip." The load-bearing words are "locked off" and "for the entire clip." Models love a slow drift that was never asked for.

Medium action. "Medium shot, barista pulling an espresso shot in a small London coffee bar, worn brass machine, steam rising, hands visible, camera at chest height, shallow depth of field, morning window light, realistic handheld micro-movement." One subject, one verb, one location detail. Two subjects in the frame and one of them will morph.

Close texture. "Extreme close-up of freshly poured coffee in a ceramic cup, steam lifting, a single fingerprint on the rim, macro lens, f/1.4, warm window light from the left, shallow focus, no text, no logos." The "no text, no logos" line is not optional — leave it out and you will get an invented café name spelled wrong, and that is the fastest way to break a viewer's trust.

Detail macro. "Close insert, steam leaving the top of a paper cup in slow motion, grey concrete wall out of focus behind, early morning light, 90mm macro, gentle drifting steam." These are the shots that make a cut feel professional. They carry no story on their own, which is exactly why the model can't get them wrong — there's nothing to get wrong.

Transition insert. "Fast motion blur of light streaks through a train window at dusk, warm orange and cold blue tones, one second of movement, usable as a cut transition, no readable text." I generate these in batches of six and keep the one where the blur shape matches the color of the next scene.

Keeping the subject and place stable across shots

The model does not do cross-prompt memory. What it does do is copy-paste memory — within a single prompt, a repeated phrase is treated as a strong constraint. So I write one block, about 30 words, describing the subject or the place, and I paste that exact block into every shot's prompt. Same jacket, same wall color, same sky. Change one adjective between shots and you've built a jump cut your audience can feel but not name.

Same move for location. Pick two or three landmarks and repeat them: "wrought-iron fence, red brick, a single streetlamp." Lighting gets its own line — "overcast afternoon, soft diffused light, no harsh shadows" — because color temperature is the cheapest way to make five different generations look like one afternoon.

Timing, length, and what I still shoot by hand

Generate long. A 10-second generation gives me maybe 3 seconds I'll actually use; a 4-second one gives me 1.5. B-roll needs breathing room, and the breathing room is almost always in the seconds you'd cut for a "safety" frame. So: request 8–10 seconds, cut to 2–3.

And here's the honest part, because I've lost a deadline to pretending this wasn't true: hands. Anything involving two hands coordinating — pouring, tying, operating a machine — still breaks about half the time. Text still renders as gibberish. Physical contact between two objects (a cup setting down on a saucer) is a coin flip. For a real shoot I now generate maybe 70% of the b-roll and still hand-hold a phone for a day. That's the split I'd bet on today, and I'd be surprised if it holds in six months.

Negative space matters too. "No text, no logos, no watermark, no people in the foreground" costs four short phrases and saves an entire re-render when the model decides the scene needs a billboard.

How I batch it

For a shoot I list the coverage before I generate: three wides, four mediums, two textures, two details, three transitions — eighteen takes. I generate all of them in one sitting, then review on a laptop at full volume with the audio off, which is weirdly the best way to spot the stutters. My keep rate hovers around 40%. The rest go back into the pile as "reference" prompts, which is how I keep improving them — every rejection tells me which phrase to tighten.

One more thing I do before any client project: a five-generation test of the widest establishing shot. If the sky and the street disagree twice out of five, the location block isn't specific enough and I rewrite it before I burn the batch. That test costs fifteen minutes and has saved me at least two full days.

The sliding taxi taught me the whole lesson in one frame: b-roll isn't one prompt, it's a shooting plan you happen to type.