You supply the source text and the runtime of one episode; if the runtime is missing, the skill stops and asks instead of guessing. It builds a fact ledger of the original, budgets seconds per story beat, turns narration and inner thoughts into visible action or sound, and splits the result into shots with size, angle, movement, performance, dialogue and AI-generation cautions. Shot durations must add up exactly to the target runtime, output defaults to 9:16 vertical, and multi-episode requests get a continuity ledger across episodes. The repo ships reference docs and sample before-and-after comparisons, plus a script that only checks the skill package's own structure.