Keeping the AI aware of its own earlier cuts as it works through the film

I kept one editable project document; the segments are views of it, not separate documents to merge later. Each local planning call receives its segment, ten seconds of context on either side, the film-wide outline and decisions, and what the pass has done so far.

Segments target four minutes, within two to five. They prefer a nearby sentence ending or scene change, fall back to a clock boundary and allow a short final segment, so the work stays bounded even when the footage does not divide neatly.

Separate documents would have needed merge logic and undo across segments; sending the whole transcript on every call kept the overload I was removing. The tradeoff is more planning calls and repeated shared context, in exchange for a smaller task per call that still serves one film.

I chose a server-controlled sequence: film-wide choices first, then local edits planned and applied in order, with the document re-read before each call for its current clip IDs and timing. Caption style, look, fades and music travel with each local prompt, and a film-wide operation proposed locally is removed and recorded.

Planning all segments at once from one copy of the timeline offered speed, but earlier edits could invalidate later plans. The model proposes edits while the server decides which step runs next, which keeps the workflow quotable, resumable and inspectable.

The cost is doing the steps one after another, which buys current references and a predictable order. Targeted segment tools let a small follow-up revision skip another full pass.