When people come across a short drama with millions of views, their first instinct is often “I'll make one with AI.” They open a video generator, type in a description, and get a few clips with mismatched art styles and characters whose faces change in every shot. The real barrier with AI short dramas isn't generation. It's the breakdown before you generate and the consistency after. Here is a workflow that produces something you can actually deliver.
Step 1: Break down the hit, don't copy it
Start by breaking the viral video down shot by shot: what hook keeps viewers in the first three seconds, how often a twist lands, how long each shot runs, how shot sizes change, and how the dialogue is paced. Record all of this in a spreadsheet. What you want to replicate is its structure and rhythm, not its visuals and lines. Copying those carries copyright risk, and it isn't fresh either.
Step 2: Rewrite the script on that structure
Keep the rhythm skeleton you extracted and swap in your own characters, settings and conflict. Have an LLM write several versions and pick the best. Revising the script ten times costs almost nothing; once you're generating footage, every revision costs real money.
Step 3: Turn the script into a storyboard and prompts
For every shot, spell out the shot size, character action, expression, setting, lighting and camera movement. Lock in your main characters' looks first and make fixed character reference images, then generate every shot from them. This is the key to fixing the “new face in every shot” problem. The more specific your prompts, the fewer times you'll need to regenerate later.
Step 4: Generate the visuals and limit retries
A common approach is to generate keyframe images first and, once you're happy with them, animate them with image-to-video, rather than rolling the dice over and over with text-to-video. This is where most of the cost goes: which model you use, how many generations per shot, and the resolution and duration all directly determine how much a video costs. It's a good idea to get the whole video working end to end with a cheaper model first, then redo the key shots with a better one.
Step 5: Voiceover, music and editing
Keep AI voices consistent: the same character should use the same voice throughout. When editing, cut to the rhythm sheet you made in Step 1, and be especially precise in the first three seconds and at the twists. Details like subtitles, sound effects and transitions decide whether viewers see it as a real piece of work or as AI-stitched filler.
Step 6: QA and delivery
Before delivering, run through a checklist: are characters consistent from start to finish, are there any shots with mangled fingers or garbled text, do lip movements match the lines, is the volume even, and is there any infringing material. If it's client work, also include the prompts, model settings and a cost breakdown, so the client can reuse them and you can quote the next job.
Viral Short Drama Remake lays out the whole process across 7 lessons of podcast audio plus illustrated notes (lessons are in Chinese), using four real remake case studies covering breakdown, rewriting, generation, voiceover and delivery, with the actual prompts, model settings and costs included. The first lesson is free.