It covers the step before prompt syntax: what a shot must convey, how many beats it can hold, shot length, framing, camera position, expression direction and dialogue handling. Its central finding is that packing many expression beats into one long shot makes H3 flatten them, while several short single-beat shots restore them, and that a dialogue tag helps but takes screen time from neighboring shots. Further rules cover size and distance wording, reference images for shape and identity, 9:16 framing limits, collisions and liquid volumes, and checking results with PSNR. A SOURCES.md file labels every rule; the SKILL.md body is in Traditional Chinese, the README in English, and the notes come from locally run production of a vertical serial drama.