The 12-Stage AI Vertical Short Drama Pipeline: From One-Line Idea to Deliverable Cut

Maosika Editorial ｜ Last updated 2026-07-25

Most AI short drama failures are not model failures — they are pipeline failures. The fix is not a better one-click button; it is a staged production line where each stage leaves an inspectable artifact before the next one starts.

Why most AI short drama runs fall apart

Teams usually blame the model when a run goes wrong: faces change, costumes drift, props disappear, episode 7 forgets episode 2, and the ending no longer matches the premise. But the deeper cause is usually a missing stage. Someone skipped the brief, skipped look development, skipped shot-level prompting, or asked one long generation to do the work of an entire crew.

A production-worthy AI vertical short drama workflow is not a single button. It is a sequence of handoffs, each with something you can point at, edit, approve, or roll back.

The 12-stage pipeline, in order

This is the order that matters. Moving stages around — for example, generating video before character looks are locked — is where most of the visible damage happens.

Stage 1 — Idea intake & scoring

The intake is not a formality. A concept that is too thin — no clear protagonist, no conflict, no hook rhythm, no platform — will get routed into deeper guidance instead of being allowed straight into writing. The point is to catch the vague pitch before it becomes 80 episodes of drift.

Stage 2 — Guided creative development

For concepts that are not yet production-ready, the system asks the missing questions: genre, lead, core contradiction, arc direction, episode count, episode length, tone, hook cadence, ending direction, and target platform. For vertical short dramas, episode length usually lands in the 1–2 minute range, sometimes extending to 3–4 minutes.

Stage 3 — Brief lock

The brief is the first hard gate. Until it is locked, writing does not start. It includes the logline, central conflict, story direction, ending direction, hook and payoff rhythm, target platform and audience, episode breakdown, and notes for the writer. The lead character field must include a character arc, not just a job title.

Stage 4 — Cast & visual confirmation

Before art or writing can safely proceed, the cast must be complete: names are valid, visual fields are filled, and leads match the brief. This is also where you catch problems like a story about three siblings where only two have been defined.

Stage 5 — Story archive build

The story archive is a structured continuity record covering identity, stable traits, current state (injuries, revealed identities, changed alliances), relationships, open and resolved plot threads, episode-by-episode appearance, batch summaries, and prop visual descriptions. It is the AI-era equivalent of a writer's room continuity bible. The principle is simple: do not rely on model memory; rely on a structured archive.

Stage 6 — Batch script writing

Scripts are written in batches of several episodes, not as one endless run. Each batch starts with intent lock: where this batch goes, who carries it, which threads move, and how the batch ends. Within each episode, the beat sheet comes before the dialogue. The final scene of each episode is planned as a cliffhanger before the body is written.

Vertical short drama writing has its own internal rules:

• The first scene must open with strong conflict or strong suspense — no slow exposition.
• Each episode follows: opening hook → escalation → end-of-episode cliffhanger.
• Every episode needs at least one small payoff; every few episodes need a larger one.
• Dialogue stays short, generally under 20 words per line; no lecture-like narration.

After writing, a rule-based checker catches broken titles, scene count mismatches, missing character lines, too little dialogue, and placeholder text like "to be continued." Failing drafts are sent back for rewrite with the error attached.

Stage 7 — Style selection

There are 17 built-in style manuals covering 2D, 3D, and realistic directions — urban realism, period realism, mature urban romance animation, 90s anime, Chinese ink style, xianxia, 3D donghua, stop-motion clay, cyber-Chinese, and others. The chosen style is shared across character art, scene art, prop art, and video prompts, so you do not end up with anime characters cut into realistic footage.

Stage 8 — Look development

Characters, scenes, and props are each drawn as reference assets before shooting. Scene art follows an empty-frame rule: no people in scene plates. Props are extracted episode by episode from the script using the original on-page names, then merged into a show-wide catalog. Human overrides are expected: you can swap a look, bind a nickname to a character, include or exclude a prop, or pick a historical versus modern version for time-crossing stories.

Stage 9 — Episode splitting into scene blocks

Each episode is cut into scene blocks targeting roughly 10 seconds each, with a soft text cap around 200 characters before further splitting. The result is not one big video per episode; it is a clip list — scene 1, scene 2, scene 3 — each with its own prompt, its own references, its own takes, and its own history.

Stage 10 — Shot-level prompt build

This is where many generic AI video tools stop being useful. A production prompt is not a paragraph of novel prose. It is built around eight components: precise subject, action detail, scene environment, lighting and tone, camera movement, visual style, image quality, and constraints.

A few prompt disciplines matter a lot:

• One camera move per shot — no stacking push, pull, pan, and tilt into one instruction.
• Shot numbers, not absolute timestamps.
• A mandatory constraint package for face stability, image quality, no watermark, and no duplicate-person artifacts in multi-character scenes.
• Dialogue in {}, sound effects in <>, score cues in ().
• Only the current scene's assets are visible to the prompt builder — no cross-scene leakage.
• If a character has a reference image, the prompt must not re-describe their clothes or appearance in text; text only describes action, expression, and injury state.

That last rule is one of the most important consistency controls in the whole pipeline. Most face-and-costume drift happens because the text prompt and the reference image are fighting each other.

Stage 11 — Multimodal shoot

Shooting is per scene block. You choose model tier, aspect ratio, resolution, and duration; 9:16 vertical is the default, not a post-hoc crop. Tasks run on an isolated queue separate from writing and art, with no parallel submissions for the same in-flight block, pre-deducted credits released on failure, and timeout recovery for stuck jobs. This is what makes it an operating production queue rather than a toy script.

Stage 12 — Review, reshoot, select

After render, you review takes, edit the prompt or swap references, and reshoot. The system does not auto-score finals or auto-pick the best take. Final quality judgment stays with the creator; what the system provides is a clean way to iterate: change the instruction, keep the references, shoot again, and compare takes.

The checkpoints that actually prevent disasters

If you only remember four gates from this pipeline, remember these:

1. Brief locked before writing. No approved brief, no script.
2. Archive updated after every batch. Continuity is an asset, not a memory.
3. Looks locked before shooting. No approved references means you are choosing lower quality on purpose.
4. Prompt reviewed before render. The model shoots what is on the page, not what you meant.

What this pipeline does not do

It is worth being explicit about the limits, because honest boundaries are what make a production tool trustworthy:

• There is no automatic final-score engine that picks the best take for you.
• Reference images are not a hard block — you can skip them and shoot from text only, but quality usually drops, which is why the professional path is to lock looks first.
• The ~10-second scene block is an engineering heuristic, not a frame-accurate timecode cut; long scenes may still need further trimming in edit.
• There is no current productized workflow for auto-extending video across scene boundaries; the unit of work is one scene block to one clip.
• Character consistency depends on the look-development chain and prompt discipline, not on a separate face-embedding verification step.

The positioning is straightforward: a scalable AI short drama operating system bends proven live-action production order into software — story and continuity first, then looks, then scenes, then shot instructions, then multiple takes — instead of promising zero-human, zero-error output.

About Maosika

Maosika (猫斯卡) is an AI production operating system for vertical short dramas. It is built around the staged pipeline described above: each stage leaves an inspectable artifact, humans make the approval and taste decisions at the gates, and the system is responsible for carrying those decisions consistently across episodes, scenes, and takes.

About Maosika — Maosika · Professional AI Video Production System. It connects briefing, scripting, look development, shot prompts and delivery into one reviewable pipeline. www.maosika.com