Script to Short Drama Pipeline

Script to Short Drama Pipeline

What it does

Organizes story/novel/script input into a repeatable short-drama production line: story beats, character and scene setup, storyboard prep, and video generation handoff.

When to use

• User has a script, novel excerpt, or story brief and wants a short drama
• Need a structured production manifest before generation
• Building Film School lessons on script-to-video workflows

Inputs

• Source text (script, novel excerpt, or story brief)
• Target duration, aspect ratio, language, audience
• Optional style references or brand constraints

Outputs

• Production manifest (beats, characters, environments, props)
• Storyboard prep per beat
• Generation handoff checklist

Steps

1. Intake — Parse source text; confirm duration, ratio, language, audience.
2. Beat breakdown — Split into 3–8 beats based on target runtime (~5–15s per beat).
3. Asset registry — List recurring characters, environments, props with stable IDs.
4. Storyboard prep — For each beat: action, dialogue, camera note, required assets.
5. Generation handoff — Ordered beat list, asset references, quality checklist.

Quality checklist

• Every beat has clear action and visual anchor
  
• Recurring characters/environments have stable IDs
  
• Beat count matches target duration
  
• Dialogue fits spoken pacing for short video
  
• No orphan props or unnamed characters in later beats
  

What we found running it

Four tasks, each run with the skill installed and again with skills disabled, identical task text. Both arms produced 22 files in total, so volume tells you nothing here. What the two arms actually built is different in kind.

The skill arm assumes the thing will be generated; the control arm assumes it will be shot. Every skill run produced an asset registry with stable IDs — 03_asset_registry.md (1,631 B) for the café scene, 03_人物与场景资产清单.md (2,323 B) for the woodworker, 03_场景道具清单.md (3,186 B) registering five ENV and eight PROP identifiers with a cross-shot consistency table. That registry is the mechanism the skill uses to keep a character's face and a prop's shape stable from beat to beat, and it is followed in every run by a generation handoff (05_generation_handoff.md, 03_generation_handoff_checklist.md, 05_视频生成交接清单.md). The control arm never produced a registry in any of the four. It produced a casting brief, a shooting schedule, an equipment-and-budget list, and a risk-and-contingency sheet instead — a call sheet for a crew.

Neither framing is wrong; they answer different questions from the same brief. But if you are feeding this into a video model, the control arm's output leaves the consistency problem entirely unaddressed.

Two smaller things the skill arm did that the control arm did not: it diagnosed the source material in task 1, noting the excerpt does not carry enough dialogue to fill 60 seconds before proposing two invented beats to cover the gap; and it landed task 3's beat table on exactly 100 seconds and added 05_采访提纲.md (1,821 B), an interview outline — correct for a documentary-style piece where the subject has to actually say the lines.

Cost was close and not consistently in one direction: $0.39 / $0.29 / $0.51 / $0.44 with the skill against $0.49 / $0.23 / $0.50 / $0.30 without.

What we did not test: nothing was shot and nothing was generated. Every artifact in both arms is a document. The asset-ID scheme is the skill's central claim and we have no evidence it actually holds a character consistent, because no frames were ever produced from these manifests. All four tasks are Chinese-language vertical short form between 8 and 100 seconds.

Reference repos

• Toonflow, ArcReel, AIComicBuilder, LocalMiniDrama