How Vertical AI Short Dramas Actually Get Made: A 12-Stage Production Pipeline

Maosika Editorial ｜ Last updated 2026-07-25

Most AI short drama failures are not model failures—they are pipeline failures. The reliable path is to lock the brief, build a continuity bible, approve looks, block scenes, then shoot and retake field by field instead of generating one long video.

The pipeline matters more than the prompt

A vertical short drama is not a single text-to-video job. It is a sequence of controlled handoffs: idea, brief, characters, continuity record, script batches, look, assets, scene blocks, shot instructions, rendered takes, review, and revision. When teams skip stages, the result is the familiar list of problems: faces change between shots, costumes reset, props disappear, episode 8 forgets episode 2, and the hook lands too late.

An AI short drama production pipeline is an ordered workflow where each stage produces an inspectable deliverable before the next stage starts. That definition matters because AI models do not maintain long-term consistency on their own. The pipeline does.

This guide walks through 12 stages using the working language of a real crew: development, writers' room, art department, camera, and post. The goal is not to remove human judgment. The goal is to move repetition, drift, and avoidable errors into a repeatable system while keeping taste decisions with the creator.

Stage 1: Score the idea before development

The first checkpoint is not writing. It is intake completeness. A raw idea is scored on the dimensions that actually affect vertical drama writing: genre, protagonist, core conflict, story direction, episode count, episode length, tone, hook rhythm, ending direction, platform, and target audience.

A strong intake can move straight to a locked brief. A partial idea needs guided development. A vague idea needs a full development pass. The point is simple: do not let a half-formed logline become 80 episodes of script.

Stage 2: Lock the creative brief

The brief is the first hard gate. It should contain the logline, central conflict, story direction, ending direction, beat and hook rhythm, platform and audience, episode-by-episode outline, protagonist arc, and notes for the writer.

Think of this as the development meeting result. Until the brief is locked, no script should be written. In vertical drama, the brief also enforces format constraints: episodes are usually 1–2 minutes, with some projects extending to 3–4 minutes, and the default deliverable is 9:16 vertical, not a landscape video cropped after the fact.

Stage 3: Approve the cast and visual identity

Before writing scales, define who is in the show. Characters need names, roles, visual fields, relationships, and a clear place in the brief. The protagonist must have an arc, not just a costume and a job title.

This stage prevents a common production leak: the writer invents new characters episode by episode, the art department cannot keep up, and the final video introduces random lookalikes. Cast approval should be treated like a real pre-production gate: lineup complete, names valid, visual fields complete, leads aligned with the brief.

Stage 4: Build the story bible

A story bible, or continuity bible, is a structured record of character identities, stable traits, current states, relationships, open and resolved plot threads, episode appearance tables, batch summaries, and prop visual descriptions.

This is one of the most important assets in AI drama production. The model should not be expected to "remember" the show across dozens of episodes. The bible is the memory:

• Who knows what
• Who has been injured, exposed, promoted, or disguised
• Which threads are still open
• Which props must appear later
• What happened in the previous batch

The rule is: do not rely on model memory; rely on a structured archive.

Stage 5: Write scripts in batches

Do not write an entire season as one long generation. Write in batches, with each batch covering several episodes. Before each new batch, lock four things:

1. The dramatic direction of the batch
2. The key characters in focus
3. The conflicts and threads to advance
4. The episode-end cliffhangers

These become hard constraints. If a new writing decision conflicts with the archive, the confirmed creative intent wins. This is how a show stays coherent while still allowing plot turns.

Stage 6: Plan beats before dialogue

For each episode, produce a beat sheet first, then write the pages. Vertical drama has strict rhythm rules that should be enforced before the script gets pretty:

• Golden 3 seconds: the first scene must open with conflict or suspense, not exposition
• Episode structure: opening hook, escalating conflict, final-scene cliffhanger
• Payoff density: at least one small payoff per episode; a larger payoff every few episodes
• Dialogue: short lines, generally under 20 Chinese-equivalent words in the source format; no lecture-style narration

The beat sheet is not bureaucracy. It is the cheapest place to fix a boring episode.

Stage 7: Run script quality control

After writing, scripts should pass a rule-based check before they move toward production. Typical failures include wrong episode titles, mismatched scene counts, missing character lines, too little dialogue, and placeholder text such as "to be continued." A failed script should be sent back with specific feedback and rewritten until it passes, within a limited number of retries.

This matters because bad script structure cannot be saved by pretty visuals. If the cliffhanger is missing or the scene count is wrong, the video stage will only make the problem more expensive.

Stage 8: Choose the visual style and do look development

Select one visual style for the whole show and use it across character art, scene art, props, and video prompts. A style manual should cover face anchors, materials, mood, view consistency, scene treatment, prop treatment, and video style tags.

This avoids the classic AI drama fracture: characters look like 2D anime, but the video output suddenly turns photorealistic. The style path should be shared, not re-decided scene by scene.

A practical range of style directions includes urban realism, period realism, mature urban romance animation, 1990s Japanese anime, Chinese ink style, xianxia period fantasy, 3D Chinese animation, clay stop-motion, and cyber-Chinese fantasy, among others.

Stage 9: Generate approved character, scene, and prop art

Create assets before shooting. Character art should be produced in two steps: first refine the text description into an image prompt using the chosen style and hard gender rules, then generate the image and store it as a versioned asset. Scene art should follow the same system, with one strict rule: empty scene plates must not contain people.

Props should be extracted from the script using the script's original naming, then merged into a show-wide catalog. The prop list should follow the script, not a generic art department guess.

For time-jump or flashback stories, route character looks by scene context: modern scenes use modern styling, period scenes use period styling. This is a rule-based asset decision, not something left to the model's imagination.

Stage 10: Cut episodes into scene blocks

Do not treat an episode as one render target. Cut it into scene blocks, roughly 10 seconds each, with a soft text-length limit and further splitting when action beats, paragraph breaks, or sentence endings require it.

Each block becomes a production unit:

• Own prompt
• Own reference images
• Own rendered takes
• Own revision history

This is closer to a clip list on an editing table than a single "make episode" button. It is also the reason revisions become affordable: if one shot fails, you reshoot that block, not the whole episode.

Stage 11: Build shot prompts with reference mapping

For each scene block, assemble references in a fixed priority: scene image → prop images → character art. Only fill slots when an asset exists. If something is missing, mark it as text-only instead of inventing a binding.

The most important consistency rule is simple: if a character has reference art, do not describe that character's clothing or appearance again in words. The reference image is authoritative. Text should only describe action, expression, and injury.

This single rule prevents a huge share of AI video face and costume changes. When the prompt re-describes a character who already has a reference, the model receives conflicting instructions and starts redrawing the person.

Shot prompts should use cinematic structure, not prose. A strong prompt includes eight elements:

Use shot numbers, not absolute timestamps. Keep actions continuous and controlled. Mark dialogue, sound effects, and music with clear symbols. The system should be teaching the model to read a shot list, not a novel.

Before rendering, check for missing references. Missing assets should trigger a reminder. Creators can still proceed with text-only prompts, but that is a conscious trade-off, usually with weaker consistency.

Stage 12: Render, review, and retake

The shoot stage should behave like a production queue, not a toy button:

• Confirm or edit the prompt
• Select model tier, aspect ratio, resolution, and duration
• Submit the scene block
• Render in a dedicated queue
• Store the result as a take
• Compare historical takes for the same block
• Change the prompt or references and reshoot when needed

The default output is vertical 9:16. Duration can be intelligent or fixed within a short shot range. Audio can be generated, and watermarks should be off by default.

A mature workflow does not try to auto-pick the "best" take. Final quality judgment stays with the creator. The system's job is to make retakes cheap, traceable, and isolated.

What this pipeline does not do

Honest boundaries make the process more trustworthy, not less.

1. There is no automatic quality scoring or auto-retake engine. Final review stays with the creator.
2. Reference images are not a hard gate. You can skip them, but text-only shots are usually less consistent.
3. Reference markers depend on prompt discipline. Creators should still check that references are present before rendering.
4. Shot grammar follows the platform's camera rules. The full art manual controls asset creation; video uses style tags and cinematic prompts.
5. The ~10-second block is an engineering heuristic, not a timecode-precise cut. Some blocks may run longer; final assembly still needs editing.
6. There is no productionized cross-shot automatic extension flow. The unit of work is one scene block to one finished clip.
7. Character consistency depends on the asset chain, not a face-embedding lock. Era routing helps, but final quality still depends on good art and disciplined prompts.

The production principle

High-quality vertical AI drama is not one long generation cut into pieces. It is a stack of controllable units. The pipeline enforces the order that professional crews already know: story and continuity first, then look development, then scene breakdown, then shot instructions, then multiple takes.

Maosika (猫斯卡) is built around this staged approach: an AI production operating system for vertical short dramas, designed to turn repeatable, drift-prone steps into a constrained pipeline while keeping审美 judgment in human hands.

If you want to go deeper into the writing side, read the [vertical short drama script process breakdown](/blog/vertical-short-drama-script-structure/). For the visual consistency problem specifically, see [why AI short drama characters change faces and how to fix it](/blog/ai-short-drama-character-consistency/). For prompt construction, use the [AI short drama shot prompt checklist](/blog/ai-short-drama-shot-prompt-checklist/).

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