Short Drama on an AI Pipeline: The Checklist That Kept Us Consistent

Short drama is cheap to shoot and expensive to keep consistent. An eighteen-scene series lives or dies on whether scene nine still looks like scene one. After running a few micro-series through AI video generation, this is the checklist that actually kept productions on track.

Lock the cast in pre-production

Not "generate a character once." Approve a reference frame per character and store it. Every scene starts from those frames, never from a fresh description. Changing a character's appearance between episodes is the most common reason a series looks amateur, and it is entirely avoidable.

Write scenes you can actually generate

Two people, one location, one emotional beat. Scenes with six characters in a crowded market are a different budget tier in every medium, and close to impossible in generated video. Writing toward the constraint is not a compromise: it is how the format has always worked on small budgets.

Generate in sets, review in sequence

Generate all shots of a scene in one sitting with the same reference frames, then watch them back to back before moving on. Problems invisible in a single clip are obvious in a sequence.

Cut as you go

Assemble a rough cut with placeholder audio as soon as two shots exist. Pacing issues show up immediately, and fixing one shot is cheap while regenerating a scene is not.

Track what you regenerate

Keep a log of which shots needed retries and why. After three episodes, the log tells you which scene types your pipeline handles badly, and that is what you write less of.

Keep the tool boring

The pipeline should be dull: same frames, same motion vocabulary, same export settings. Experimentation belongs in pre-production, not in episode five. A tool built for cast consistency rather than one-off clips makes this much easier — Pavo AI is designed around exactly that, with a free tier to test at https://pavoai.app/

Eighteen scenes is a lot. Eighteen scenes that look like one series is the goal.