owntent
owntentowntentCreator-owned AI studio
BlogDocsPricingLog in

25–30 cuts per minute: the short-drama editing grammar we measured, then automated

Published 2026-08-04

When we started producing vertical short drama with AI, the first versions felt wrong in a way that was hard to name. The images were fine. The voices were fine. The episodes still felt slow.

So instead of guessing, we did the boring thing: we sat down with hit vertical dramas — the kind topping the short-drama apps — and counted cuts with a stopwatch.

What the measurements said

Three findings repeated across everything we measured:

  • 25–30 cuts per minute. That's an average shot length near two seconds. For comparison, modern feature films average around 4–6 seconds per shot. Vertical drama runs at roughly double that pace.
  • Hard cuts, essentially 100%. No dissolves, no wipes, no fancy transitions. The grammar is: cut, cut, cut. Any transition longer than a frame reads as hesitation on a phone screen.
  • Dialogue carries everything. Cuts land on lines and reactions — speak, cut to reaction, cut back. Scenery breaths that a TV drama would allow simply don't exist. If nobody is talking or reacting, the shot is usually gone.

Four consecutive shots from our series, seconds apart — hard cuts only

The counterintuitive part: a beautiful four-second shot performs worse than two ordinary two-second shots. Pacing beats polish. Viewers on a phone decide in seconds whether something is dragging, and "dragging" starts around the third second of any single image.

Turning numbers into pipeline defaults

A measurement is only useful if the production system enforces it. We baked the grammar into the pipeline as defaults, not guidelines:

  • Shot breakdown targets ~2-second beats. When a script scene gets decomposed into shots, the planner splits action and dialogue into beats sized for the measured rhythm — a 60-second scene should come out as roughly 25–30 shots, not 12.
  • Hard cuts are the only default. The edit assembler simply concatenates. Transitions exist as an explicit, rarely-used override rather than a temptation.
  • Every shot must justify itself with a line or a reaction. Shots that carry neither get flagged during breakdown, before anything is generated — the cheapest possible place to kill a slow moment.

We ran a full five-episode romance series through these defaults (an enemies-to-lovers contract marriage story — yes, we know, the genre exists because it works). The difference against our early slow drafts was immediate and embarrassing: same story, same art quality, completely different watchability.

Why this matters for AI production specifically

Hand editors internalize pacing over years. An AI pipeline has no instincts — it has defaults. That's usually framed as a weakness, but it inverts nicely: once the right grammar is measured, the pipeline applies it with perfect consistency across every episode, every series, at batch scale. Episode 47 keeps exactly the discipline of episode 1.

The craft moves up a level: instead of cutting each scene by hand, you decide the grammar once and then direct — reviewing output, rejecting takes, adjusting the spec when a genre demands it.

You can judge the results yourself — the series produced under this grammar are free to start at owntent.com/watch. And if you want to direct your own with these defaults already in place, start at owntent.com.

← All posts
owntentowntentCreator-owned AI studio

Make it with AI. Own your fans. A creator-owned short-drama & webtoon studio.

Use cases

  • How it works
  • Script to film
  • Webtoon to film
  • Compare services

Product

  • Pricing
  • Account

Resources

  • Privacy
  • Terms
  • Acceptable use
  • Copyright
© 2026 owntent, All Rights ReservedSupport: support@owntent.com

owntent's AI features are powered by third-party AI models (such as Seedance, Kling, and GPT-Image). owntent is an independent product, not affiliated with or endorsed by these model providers; model names are trademarks of their respective owners. Generation results vary with model capability.