How AI Vertical Short Dramas Actually Get Made: The 12-Stage Production Pipeline From Idea to Deliverable

Maosika Editorial ｜ Last updated 2026-07-25

Most AI short drama failures are not model failures—they are pipeline failures. A usable vertical drama is built stage by stage, with locked handoffs between story, continuity, look, scene blocks, prompts, shooting, and review.

If you ask for "a short drama" in one prompt, you usually get a smooth-looking first clip and a collapsing second episode. Faces change, costumes drift, props disappear, hooks soften, and episode 5 forgets what episode 1 promised.

That is not because the model is bad. It is because the job was treated as a single generation instead of a production process.

A professional AI vertical short drama pipeline is closer to a real crew than to a magic button. Each stage produces something inspectable. Each stage can be revised before the next one spends time and money. The human keeps taste, story judgment, and final approval; the system handles repetition, continuity, and the boring but deadly consistency work.

The 12 stages, in order

This is the production order that separates a deliverable vertical drama from a demo reel.

This order is not arbitrary. It mirrors the way live-action crews avoid expensive mistakes: lock the brief, cast the roles, build the bible, write in batches, approve the look, block the scenes, then shoot.

Stage 1–3: From idea to locked brief

A vertical short drama idea usually starts as a sentence: "A disgraced heiress returns as a mysterious investor." That is a logline seed, not a production brief.

The first job is to turn that seed into something a writer can execute episode after episode.

What a locked brief should contain

• One-sentence logline
• Core conflict
• Main character arc
• Story direction and ending direction
• Hook and payoff rhythm
• Platform and target audience
• Episode breakdown plan
• Notes for the writer about tone, taboos, and must-hit beats

For vertical dramas, the brief also needs format constraints: episodes are usually 1–2 minutes, with some productions extending to 3–4 minutes. The first scene must hook immediately. Background explanation cannot sit in front of the conflict.

Creative intake scoring is a useful production habit. If an idea is too thin, it should go through deeper guidance. If it is already complete, it can move straight to brief lock. The point is not to score creativity; it is to prevent under-defined stories from entering writing.

Stage 4–5: Characters and the story archive

A short drama can run for dozens or hundreds of episodes. The model will not remember everything. Neither will a tired writer working at batch speed.

A story archive is a structured continuity record for the whole show. It tracks character identities, stable traits, current state, relationships, open and resolved plot threads, episode appearances, batch summaries, and prop descriptions.

This is the key difference between "chatting with a model" and "running a show." When a new batch is written, the system should feed only the relevant archive slice: current character state, unresolved threads, recent batch summary, and the new batch objective. That reduces both forgetting and irrelevant noise.

Continuity does not come from asking the model to "remember." It comes from storing decisions in a structured archive and reading from it before every batch.

Stage 6: Batch script writing for vertical rhythm

Writing one giant script file is a bad idea for vertical drama. It is better to write in batches, with each batch planned, written, checked, and archived before the next batch begins.

Vertical script rules that should be enforced

1. Golden 3 seconds: the first scene must open with conflict, shock, mystery, or a strong power shift. No slow setup.
2. Single-episode structure: opening hook, escalation, then a cliffhanger in the final scene.
3. Payoff density: at least one small payoff per episode, such as a reversal, identity hint, evidence reveal, or face-slap beat; larger payoffs every few episodes.
4. Short-line dialogue: lines should be tight and speakable, usually under 20 characters in Chinese-style pacing logic and adapted naturally in English. No lecture-like exposition.

Before writing the actual dialogue, each episode should have a beat sheet and a planned cliffhanger. Then the script can be written against that plan. After writing, a rule-based check should catch obvious failures: missing character lines, wrong scene counts, placeholder text like "to be continued," too little dialogue, or format errors.

A failed script should be sent back with specific feedback, not handed to the user as a draft that needs forensic cleanup.

Stage 7–8: Style and look development

One of the most common AI drama failures is style mismatch: a character is generated in anime style, but the video render looks realistic; a scene is painted as cyberpunk, but the props look like modern office stock.

The solution is to choose one style path and let characters, scenes, props, and video prompts all follow it.

A production-ready style system should include:

• Character appearance guidance: face anchors, material,气质-like presence, view consistency
• Scene and prop guidance
• Video style tags that carry the chosen look into generation

Reference art should be produced in stages:

1. Text is polished into image-generation prompts using the style manual and hard constraints such as gender.
2. Images are generated, versioned, and saved.
3. Only finished, readable assets are used for video generation.

Scene references have a simple but important rule: scene art should not contain people. Empty scenes are scene references. If a person is in the frame, the model may treat that person as part of the scene cast.

Props should be extracted from scripts using the script's own names, then cataloged across episodes. If episode 8 calls it "the old locket," the prop system should not rename it into something generic.

Stage 9: Cut episodes into scene blocks

A 60-second episode should not be treated as one video prompt. It should be cut into scene blocks, usually around 10 seconds each.

A scene block is the smallest controllable production unit. It has its own script section, references, prompt, shot plan, output clip, and retake history.

Why this matters:

• If one moment fails, you reshoot that moment, not the whole episode.
• Each block can have the right references without cross-scene contamination.
• Producers can review a clip list like an editing timeline.
• Costs are easier to control because failures are isolated.

Longer text sections can be split by action beats, punctuation, and paragraph breaks. The target is not broadcast-accurate timecode; it is a shootable unit.

Stage 10: Build shot prompts like a shot list

AI video models do better when prompts sound like a shot list, not a novel.

A strong shot prompt usually includes eight elements:

A few prompt disciplines matter a lot:

• One shot, one camera move. Do not stack push-in + pan + shake in one clip.
• Use shot numbers, not absolute timestamps like "0–3s."
• For complex scenes, use a three-part structure: overall setup, shot-by-shot beats, then constraints.
• Keep motion continuous and moderate; extreme fast action is more likely to break.
• Mark dialogue, sound effects, and music with clear symbols so they are not confused with visual instructions.
• Feed only this scene's references. Do not let another scene's props or characters leak in.

There is also one rule that prevents a huge amount of character drift:

If a character has a reference image, do not describe that character's clothing or appearance again in words. The reference image is the source of truth. Text should only describe action, expression, and injury or state changes.

When text fights the reference image, the model improvises. That is how faces and costumes change between shots.

Stage 11: Shoot scene by scene

The actual production step should feel familiar to anyone who has worked on set.

For each scene block:

1. Confirm the reference set is complete, or knowingly proceed with text only.
2. Generate the shot prompt.
3. Read and edit the prompt if needed.
4. Choose model tier, aspect ratio, resolution, and duration.
5. Submit the shot.
6. Wait for render.
7. Review the take.

The default deliverable shape should be 9:16 vertical, not a horizontal video cropped after the fact. Vertical framing changes composition, headroom, gesture, and subtitle space. It should be built in from the start.

A serious production queue also needs operational discipline:

• No duplicate submissions for the same scene while a task is running.
• Failed tasks should be visible and retryable.
• Credits or compute should be reserved on submission and released on failure.
• Stuck tasks should time out and recover.
• Video tasks should be separated from script and art tasks so one workload does not block everything.

This is what turns a script into production infrastructure rather than a toy.

Stage 12: Review, reshoot, and select

AI does not replace the director's eye.

After a scene is rendered, the creator should be able to:

• Compare multiple takes of the same scene
• Edit the prompt and reshoot
• Swap a character, scene, or prop reference
• Manually bind a script name to the correct character asset
• Include or exclude props to control visual focus
• Switch style if the whole look needs to change

A good system does not silently decide which take is "best." It logs takes and lets the creator choose. There is no reliable universal automatic scorer for drama quality. The final judgment belongs to the person who understands the audience, the hook, and the story beat.

Where teams get the order wrong

The fastest way to waste budget is to shoot too early.

Common wrong orders:

• Writing a full script before locking the brief
• Generating video before character designs are approved
• Using one long episode prompt instead of scene blocks
• Describing costumes in prompts after reference art exists
• Letting episode 12 read raw chat history instead of a structured archive
• Treating every bad clip as a model problem instead of a prompt, reference, or blocking problem

High-quality short drama is not one long generation cut into pieces. It is many controllable units stacked into a coherent show.

Honest limits

It is important to say what this kind of pipeline does not do.

1. There is no automatic "best clip" judge. Final quality review remains human.
2. References are not a hard gate. You can shoot without them, but quality usually drops; professional process should approve look dev first.
3. Prompt reference markers still need human review. Good structure reduces mistakes, but creators should still read the prompt.
4. Scene blocks are heuristic, not frame-accurate edits. Some blocks may run long; final trimming and assembly still belong in editing.
5. There is no cross-scene automatic video extension workflow. The production unit is one scene block to one clip.
6. Character consistency depends on the asset chain, not face embedding magic. Era-specific costume routing helps, but final look still depends on reference quality and prompt discipline.

These are not flaws to hide. They are the boundaries that make the tool usable for professionals.

About Maosika

Maosika (猫斯卡) is an AI production operating system for vertical short dramas. Instead of offering a single text-to-video button, it structures the workflow from idea to deliverable through locked brief, story archive, look development, scene blocking, engineered shot prompts, multimodal shooting, and take-by-take review. Its position is clear: repeatable, drift-prone, failure-prone steps become a constrained pipeline; aesthetic judgment stays with the creator.

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