How We Make an AI Short Drama: Billionaire’s Secret Bride

Making an AI vertical short drama is mostly a writing and review job; the generation step is the smallest part. We know this because we are doing it with Billionaire’s Secret Bride, a 60-episode contract-marriage series with 40 episodes produced so far. It is not out yet. It is coming soon to our drama catalog.

Here is the workflow in order. Where something is still a plan, we say so.

Before any image exists, the show has a bible. Ava Chen is 29, proud and strategically calm; she takes the marriage deal for her mother’s treatment, but her goal is autonomy, not rescue. Ethan Blackwood is 34, a feared heir who confuses control with safety. Selena Vale, 32, is his former partner and a media tactician, and the bible says in so many words that she is never to be reduced to jealousy.

The season is split into three acts. Episodes 1–15 are the bargain, in which Ava wins contractual equality. Episodes 16–42 turn the couple into reluctant investigators of a crime that connects their families. Episodes 43–60 expose the cover-up and end with a marriage chosen on equal terms.

The bible also fixes the hard limits. Every character is a 21+ adult. No nudity, no explicit sex, no coercive romance, no resemblance to real actors or celebrities, no logos or luxury brands. Ava keeps her agency in every episode.

Each episode runs about a minute and has a fixed shape: a hook in the first three seconds, one turn in the middle, a cliffhanger at the end. A summary has to fit one sentence containing an event, a conflict and a hook. If it does not, the episode is not ready.

Episode 1, The Secret Wedding, opens on a hospital bill stamped past due. Ava signs a marriage contract in a candlelit registry, hears footsteps, and turns to find the groom is Ethan Blackwood. The doors lock. Episode 2 puts her at a family dinner where she is humiliated, then plays a recording of the family’s own threat. By Episode 10 Ethan tears up a contract, and it is the wrong one, because Ava kept the binding copy.

Scripts are written shot by shot with timecodes. Episode 1 has five beats across 30 seconds, each with a picture note and a short spoken line. Episode 1’s music note: low piano, rain, one bass hit.

The biggest risk in AI video is that a character’s face drifts between shots. So every series carries a fixed text description for each lead, reused word for word in every image and video request, along with a palette and a world lock that names the place, era and signature spaces.

For this show the palette is rain, black cars, glass towers, ivory fabric, champagne gold and red evidence lights. Wardrobe moves on a planned path: Ava goes from soft ivory to structured ivory-and-black tailoring, and Ethan’s rigid black slowly picks up warmer details. Three props recur all season: the marriage contract, the missing page from Ava’s accident file, and Ethan’s evidence wall. There is also a rule that characters from our other series are never swapped in for Ava or Ethan.

Our image bible requires approved reference images of each lead, alone and together, before any episode media is made. A reference image does not count as a finished episode.

The poster is 2:3, with a dark lower third so the title and badges stay readable. Each episode gets a thumbnail and an illustration, vertical-first, showing one clear conflict and one prop: the contract, a ring, a file, a camera, an elevator. For a thumbnail we pick a single emotional moment from the episode rather than squeezing every beat into one frame. The first ten episodes of Billionaire’s Secret Bride already have their episode art.

Video is assembled from short generated clips that follow the script’s shot order and keep its last beat. The rule we set for ourselves is blunt: generate the same character three separate times from the same reference, show the clips to three people on the team without telling them which is which, and at least two must agree it is the same person. Otherwise it fails.

Our reject list for images and frames is specific: malformed hands, duplicate faces, stray text or watermarks, logos, a borrowed skyline, a generic interchangeable CEO office. On video, the review sheet covers 9:16 framing, sound, subtitle sync, and continuity of characters, costumes and sets from shot to shot. Anything that fails goes back for rework.

Subtitles are burned into the picture in English, centered in a safe zone, white with gold accents for the speaker. The series title, logline and synopsis are localized into Traditional Chinese, Japanese, Korean and Brazilian Portuguese, along with previews of the first ten episodes. They are rewritten per market, not translated line by line; the English script stays the source of truth. The plan is that no language version goes out until someone has reviewed it.

Every display image, episode thumbnails included, gets a translucent ShortFling corner mark before it is published. Images from the NSFW zone always get one, and since August 2026 ordinary display images get one too. The clean original never goes public. We do this because thumbnails get copied and reposted, and a mark is the cheapest way to show where they came from.

Quite a lot. People write and approve the bible and accept or reject every reference image. A person reviews each generated file before it is attached to an episode. An automatic validator checks the structure, such as continuous episode numbering, required fields and repeated character descriptions, but it does not judge content. The two hard lines, adults only and no real people, are checked by people.

The same reference-first approach is the basis for AI Star Mode, where you will be able to cast your own face as a lead. It is not live yet; there is a waitlist.