How to make AI micro dramas that feel real

Every hit opens with tension already on screen in the first second. A few lines, a quiet middle for a song, a payoff you want to send to someone. Thirty seconds.

One document every prompt reads from, so characters and look never drift between parts.

Not "a woman". A 1994 Manhattan film still: year, city, lens, light, grain. Four options, pick the one that feels like a real person.

1994 Manhattan film still, a 27 year old woman, honey-brown hair, oversized cream cable-knit sweater, holding a chunky cordless phone, warm tungsten lamp light, cool blue dusk, 35mm film grain, soft halation, glossy 90s romance film, real photograph --ar 9:16

One photo is not enough for fifty shots. Front, back and close-up, same face, same clothes, made with Seedream 5 or Nano Banana 2.1. The sheet is the actor now.

Every place gets its own image with nobody in it. And the hero prop, the torn photo booth strip, gets a sheet like an actor, or it changes in every shot.

No lip sync. Give the character image to Seed Audio, describe who they are, and have them say a line that is not in the show. A clean studio clip is the voice identity.

The woman in the reference image speaks. She is Nora, 27, grew up in a small town in Ohio, two months in New York, on the phone with her sister, relaxed and a little amused. Clean studio recording. No telephone filter. No music.
She says: "No, the radiator still bangs all night..." (any line that is not in the show)

The sheets, the location, the prop and both voice clips go in. The golden line: @Audio1 is ONLY Nora's voice identity, never speak the words heard in it. Then the dialogue goes inside the shot, with accent and delivery. Acting and voice are born together.

6 connected cinematic shots, 15 seconds total, 9:16, Multishot ON
Total: 15s / 6 shots / 9:16

REFERENCES
@Image1 defines NORA's face, hair, body and wardrobe exactly. Do not use the image background.
@Image2 defines THEO's face, hair, body and wardrobe exactly. Do not use the image background.
@Image3 defines the hallway outside apartment 4B: layout, materials, light and grade.
@Image4 defines the inside of Nora's studio by the front door: layout, materials, light and grade.
@Image5 defines the torn black-and-white photo-booth strip: Nora holds the LEFT piece (four frames, torn top edge); Theo holds the RIGHT piece (one frame, torn bottom edge). Do not use the background.
@Audio1 is ONLY NORA's voice identity. @Audio2 is ONLY THEO's voice identity. Never speak the words heard in them.

[GLOBAL LOOK - LOCKED, APPLIES TO EVERY SHOT]
Manhattan, winter 1994, late evening. Glossy 90s film glamour: warm tungsten practicals inside, rich slightly desaturated midtones, 35mm film grain, soft halation around every lamp, round creamy bokeh, shallow depth of field. Vintage spherical cinema lenses, 40-50mm feel. Camera on a dolly or tripod: steady, slow push-ins only, no shake. Vertical 9:16 frame; faces fully in frame on every speaking or reacting shot.

Ten Minutes took seven versions. In the early takes she cries from the first frame, and it feels fake. Real people don't start crying. The fix: start calm, let it build, wet eyes only at the end.

Video is the expensive part. Every take starts as a 480p draft, then only the takes you keep get the full render and a Topaz 4K upscale. Then the edit: two parts, a song in the quiet middle, captions.