Direct answer first: a broadcast-density AI episode - ours runs 58 cuts in 2 minutes 22 seconds, about 25 cuts per minute - is a logistics problem, not a generation problem. One person ships it by treating the episode as a manifest of cuts that flows through fixed stages, where every stage produces an approvable artifact before the next stage spends money. This is the anatomy of one real episode from our catalog, with the numbers from our production logs.
The manifest is the episode
The episode starts as a cutsheet: 58 rows, each with duration, camera note, on-screen characters, action, and dialogue. Nothing downstream is improvised - if a cut is not on the sheet, it does not exist. Cut density is a creative decision made once: hit vertical dramas run 25-30 cuts per minute, and writing the sheet at that density is what makes the result feel like television instead of a slideshow (why density matters).
Stage by stage, with real counts
- Casting sheets - multi-angle reference sets per character, generated once per series. Every keyframe generation attaches them; no character is ever described from scratch (the method).
- Keyframes - one approved still per cut before any video money is spent. Acting-heavy cuts took 2-4 takes per keeper in our logs. A rejected still costs cents; a rejected video costs dollars - approval order is the budget.
- Video passes - image-to-video per cut, batched overnight. Failures are normal: GPU memory contention alone cost us several takes per batch, recovered by rerunning the manifest's incomplete rows. The manifest, not a human memory, knows what is done.
- Continuity audit - automated checks catch what eyes miss at volume. A pixel-level scan across a 329-segment reshoot caught exactly 5 clips with ghost artifacts from reference frames - a 1.5% defect rate invisible in normal review.
- Dubbing and mix - locked voice-plus-seed per character, dialogue on cut boundaries, sidechain-ducked bed (the dubbing chain).
- Master and publish - upscale to 1080x1920, overlay subtitles (never burned), publish to the series page where early episodes are free and later ones unlock (where the money happens).
What actually consumes the person
Not prompting - deciding. Fifty-eight cuts times four or five decisions each (framing, take choice, line read, timing) is around 300 decisions per episode. The pipeline's job is to make each one small and final: approve this still, keep this take, done. The moment a decision can be reopened later, one person drowns. That constraint - decisions close and stay closed - is the honest answer to "how does one person do this", and it is the design center of the studio we built around this workflow.
We have shipped 30+ series this way as of August 2026, all public on our catalog.
What we could not verify
Our defect and retake rates are from our own logs on our own genre mix (thriller and romance vertical drama); we have no comparable logs from other studios to benchmark against. And "broadcast density" here means cut rhythm, not broadcast picture standards - our masters are 1080x1920 web deliverables, not TV-spec.