How an AI Agent Made a Viral Micro Drama: 500K+ Views on One Episode

Case Study

Wrong Number is a six-part micro drama. We posted it as short reels on Instagram. One episode, The Coat, passed 500,000 views. The episode before it, made with the same characters and the same tools in the same week, stopped at 18.7K.

This post covers how the series was made, and why one reel won and the other lost.

What a micro drama is (and what it is not)

A micro drama here is not a series of chapters. It is one scene, 30 to 45 seconds, two people, one turn.

• Standalone beats serial. People share a scene. They do not share "part 2".
• End on a payoff, not a cliffhanger. The Coat ends on both faces, close and warm. Viewers rewatch a payoff.
• Leave a silent middle. The music goes in during the edit, so the scene needs room for it.

The workflow, step by step

1. The story bible

Before a single image there is one document: who Nora and Theo are, what they want, where they live, how 1994 looks (warm interiors, grain, snow, wet streets), and every location and prop. Every prompt after this pulls from the bible. That is why the characters stay the same people across reels.

2. Casting with Visual DNA

Faces were cast in Midjourney until they felt like real people, not models. Each face became a character sheet, saved as a Visual DNA in Kolbo.AI, so the same face comes back in every shot. A character changes clothes only when the story says so.

3. Locations and props

Every location gets one graded frame in the 1994 look before any video: the stoop, the apartment, the payphone. Props get the same treatment, because props carry the plot. In The Coat, the coat is the story.

4. Voices: the step most people skip

This is the difference between "AI video" and a scene you believe.

• Each character's voice is designed once with Seed Audio, then kept as a short, neutral sample.
• The dialogue is not generated separately and lip synced on top. It is written inside the Seedance 2.5 prompt, with each voice sample passed as a reference. Seedance 2.5 generates the picture and the soundtrack together, so mouths, breath and room tone match.

5. The Seedance 2.5 prompt

Each reel is two Seedance 2.5 parts of up to 15 seconds each. Each prompt locks who is in frame (the Visual DNAs) and where they stand, the camera and blocking beat by beat, the lines in order, and an audio block with dialogue and room tone only, no music.

6. Drafts, takes and 4K

Work in Draft at 480p. Most shots need 1 to 3 takes, so you only pay for HD once a take works. Keep the take, upscale it to 1080p, then to 4K with Topaz. Cut on shot changes, never in the middle of an action.

Why one reel hit 500K and the other stalled at 18.7K

The Door opened on a chain on a door: slow, unclear, nobody to care about yet. The Coat opened on two faces already in tension. The first second decides the reel. Everything after it only matters if people stay.

Other lessons from the production:

• One scene took 7 versions. Emotion has to build. Wet eyes in the first second read as fake.
• One character was cut entirely. No prompt fixes a character who feels wrong.
• Judge the look from the actual frames, never from the bible text.

The agent's part

Zohar directed. An AI agent ran the pipeline through the Kolbo.AI MCP, from the bible to the 4K files: casting, Visual DNA, locations, voices, prompts, drafts and the edit. The human job was the notes. Every rule the notes produced went back into the agent's skills, so the next series starts with them already in place.

Bible, cast, Visual DNA, locations and props, voices, a Seedance prompt with the dialogue inside, Draft takes, 4K, then the edit. That is the whole system.

Go make something worth watching. - Zohar Founder, Kolbo.AI