The AI Vertical Short Drama Pipeline: Why Order Matters More Than Prompts

Maosika Editorial ｜ Last updated 2026-07-25

Most AI short dramas fail not because the model is weak, but because the production order is wrong. Build story and continuity first, lock looks second, then split into scenes and shoot each as a controlled unit.

The core problem is not generation; it is sequence

Teams new to AI video usually start the same way: write a loose script, feed it straight into a video model, and hope the result looks like a show. What comes back is familiar: faces shift between shots, costumes change for no reason, props appear and disappear, episode 7 forgets a setup from episode 2, and the pacing dies in the opening seconds.

This is not a prompt-engineering problem. It is a production-management problem.

Traditional crews already know the answer: you do not roll camera before the brief is locked, the cast is set, the looks are approved, and the scene list exists. AI does not remove that discipline. It makes it more necessary, because every new generation is a fresh chance for the story to drift.

A reliable AI short drama pipeline is a controlled production sequence, not a single generate button. The order is the product.

The industrial order, from idea to deliverable

Below is the sequence that separates experimental clips from something a team can run at scale.

This is the same order a live-action crew uses, translated for AI-native production.

Stage 1: Intake — score the idea before you write it

Not every idea is ready to become a show. A useful intake stage asks the same questions a producer would ask in a development meeting:

• Who is the protagonist, and what changes about them across the series?
• What is the core conflict in one sentence?
• Where does the story start, escalate, and land?
• How long is each episode, and how many episodes total?
• Which platform is this for, and who is the target viewer?
• What is the hook rhythm: how often must a reversal, reveal, or payoff land?

A strong intake process does not just collect answers. It scores completeness and routes the project accordingly. If the brief is too thin, the team should not be allowed to jump into pages; it should be sent back to fill the missing dimensions. If the brief is already solid, the team can skip redundant questioning and move directly to a locked brief.

Think of this as the greenlight meeting. No live-action set starts shooting because someone had a cool logline in the elevator.

Stage 2: Lock the brief before any pages exist

A locked creative brief is the reference point every later stage obeys. It should contain at least:

• Logline
• Core conflict
• Story direction
• Ending direction
• Hook and payoff rhythm
• Platform and audience
• Episode-level outline
• Notes for the writer
• Protagonist arc

The protagonist arc matters especially in vertical short drama. A character who only reacts and never changes will stop carrying the series after a few episodes, even if the premise is strong.

Locking the brief does not mean creativity ends. It means later stages stop arguing about what the show is.

Stage 3: Build the story bible before scale

A story bible is a structured continuity record that travels with the whole series. It tracks:

• Character identities and stable traits
• Current state: injuries, hidden identities, changed relationships
• Relationship map
• Open and resolved plot threads
• Episode appearance plan
• Batch-level plot summaries
• Prop visual descriptions

This is the AI-era version of a writer's room continuity bible.

The reason this matters is simple: language models do not have reliable memory across hundreds of scenes. If you rely on the model to "remember" a scar, a secret, or a promise made in episode 3, you will eventually lose. A structured bible removes that risk. The writing process should consume only the slice it needs: current character states, unresolved threads, recent batch summary, and the current batch objective.

Continuity comes from assets and records, not from luck.

Stage 4: Write in batches, with beat sheets before pages

Vertical short drama has strict rhythm requirements that normal screenwriting habits often violate:

1. Golden 3 seconds: the first scene must open with conflict or suspense, not exposition.
2. Single-episode structure: opening hook, escalation, final-scene cliffhanger.
3. Payoff density: at least one small payoff per episode; a larger payoff every few episodes.
4. Dialogue discipline: short lines, no lecture-style speeches, minimal explanatory narration.

A strong AI script pipeline does not ask the model to "write an episode." It asks for the beat sheet first, including the cliffhanger, then asks for pages against that plan. After drafting, a rule-based QC pass should catch structural failures such as missing character lines, wrong scene counts, placeholder endings like "to be continued," or too little dialogue. If the draft fails, it should be rewritten with the error feedback before anyone sees it.

Writing should also happen in batches. After each batch, the bible is updated. Before the next batch begins, the next directional choices are confirmed: plot direction, focus characters, active threads, and end-of-batch hooks. Those confirmed choices become hard constraints.

This is how a 100-episode series stays coherent.

Stage 5: Lock looks before you shoot

Look development is not decoration. It is a consistency mechanism.

A production-ready style system should include:

• Character design rules: facial anchors, material, tone, view consistency
• Scene and environment rules
• Prop visual rules
• Video style tags shared across all downstream outputs

The important word is shared. If character art follows one style path and video generation follows another, you get the classic AI drama failure: anime-looking characters suddenly rendered as realistic people, or period costumes drifting into modern clothing.

Before shooting starts, confirm:

• The full cast is present
• Names are valid and consistent
• Visual fields are complete
• Lead character designs match the brief

Then build scene art and prop art with discipline. Scene art should not contain people, because it is meant to be an environment reference, not a cast photo. Props should be extracted from the script using the script's own names, then merged into a show-level catalog so nothing disappears between episodes.

Stage 6: Split episodes into scene blocks

Do not treat one episode as one giant generation job.

A better unit is the scene block: a short segment, roughly 10 seconds, with its own script slice, its own reference set, its own prompt, and its own render history. If a block is too long, split it further by action beats, paragraph breaks, or sentence boundaries.

This gives you three advantages:

• You can fix a bad moment without re-rendering an entire episode.
• Each prompt stays focused on one action.
• Reference images map cleanly to the scene that actually needs them.

The result should look like a clip list on an editing timeline, not a single opaque video file labeled "Episode 3."

Stage 7: Build shot prompts like a shot list, not a novel

A production-ready shot prompt is not a paragraph of poetic description. It is a structured instruction set. At minimum it should cover:

A few rules matter a lot:

• One shot, one camera move. Do not stack push-in, pan, and tilt into the same shot.
• Use shot numbers, not absolute timestamps.
• For complex scenes, use a three-part structure: overall setup, shot-by-shot instructions, constraint pack.
• Favor slow, continuous motion over high-action chaos.
• Mark dialogue, sound effects, and music with clear symbols.
• Feed the prompt builder only this scene's assets. Do not let props or characters from other scenes leak in.

The goal is to make the model read like a camera department, not a novelist.

Stage 8: Render, review, and re-render selectively

A mature production workflow treats each scene block as a take.

When you render, the system should use the prompt you approved and the references you locked for that scene. If you want a different result, you change the prompt or swap a reference, then render a new take. That is the same logic as a film set: you do not ask the crew to magically improve without changing the instruction.

Professional controls at this stage include:

• Editing the shot prompt like a director revising shot notes
• Swapping character, scene, or prop references
• Manually binding script names to correct cast entries
• Including or excluding props to control visual focus
• Switching style paths when needed
• Choosing model tiers for quality, cost, and speed tradeoffs
• Reviewing historical takes and selecting the best one

There is no need to re-invent the whole scene because one facial expression failed. Re-render the unit that failed.

Stage 9: Assemble for the platform, not just the model

Vertical short drama is a platform-native format. The default output should be built for 9:16 vertical, not cropped later from a horizontal master. Episode length usually sits in the 1–2 minute range, with some productions extending to 3–4 minutes depending on platform and audience.

After scene renders are selected, you still need assembly:

• Ordering clips for rhythm
• Smoothing transitions between blocks
• Sound, music, and effects alignment
• Final pacing pass
• Export to platform specification

AI generates units. Human editorial judgment still turns those units into a watchable show.

The reference image rule that prevents most face drift

One rule deserves special attention because it solves a large share of AI drama consistency problems:

When a character has an approved reference image, the prompt should not re-describe that character's clothing or appearance in text. The reference image is the source of truth. Text should describe only action, expression, and temporary state such as injury.

This sounds simple, but it attacks the root cause of face and costume changes. If the prompt says "woman in red dress" while the reference shows a blue coat, the model receives conflicting instructions. Remove the conflict by letting the image own appearance.

Reference mapping should follow a fixed priority: scene first, then props, then characters. If an asset does not exist, mark it as text-only rather than inventing a false binding. For time-jump or costume-change stories, route the correct character look to the correct scene based on setting.

What this pipeline does not do

Honest boundaries matter. A real production system should not pretend to solve everything.

• It does not automatically score final creative quality or auto-choose the best take. Human review still decides what is good enough.
• Reference images are not a hard gate. You can skip them and render from text, but quality usually drops; professional teams lock looks first.
• Scene blocks are an engineering heuristic around 10 seconds, not frame-accurate timecode. Final editorial control still matters.
• The unit of production is the scene, not an automatically extended multi-scene video stream.
• Character consistency depends on the full look-development chain, not on a magic face-lock claim.

That is not weakness. That is what makes the system usable for real production.

Why order beats prompt tricks

Most teams waste time looking for the perfect prompt that will make AI behave like a full film crew. That prompt does not exist.

What exists is process:

• Approve the brief before pages.
• Build the bible before scale.
• Lock looks before shooting.
• Split into scenes before rendering.
• Give each scene only its own assets and instructions.
• Keep multiple takes and re-render selectively.

High-quality short drama is not one long, impressive generation cut into pieces. It is many controlled, reviewable units assembled into a show.

AI does not replace the crew. It makes the crew's existing discipline executable at speed.

---

*Maosika (猫斯卡) is an AI production operating system for vertical short dramas. It encodes the sequence above into a structured workflow: intake scoring, brief lock, story bible, batch scriptwriting with rule-based QC, look development across 17 style manuals, scene blocking, shot-prompt generation, multi-take rendering, and selective re-renders. It is designed for teams that want a repeatable production pipeline, not a one-click miracle.*

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