A continuity ledger for AI short dramas: one character, three shots

A small story with an observable change

Google’s September 24 research post ↗ describes multi-agent orchestration and visual memory for coherent long-form video. This tutorial proposes a separate three-shot continuity exercise, without reproducing its system.

Our exercise is deliberately small: Lin makes tea, walks into the hallway, takes a sip, and returns the mug to the kitchen. Aim for about twenty seconds in total. There is one character, one important prop and two locations. Use tools you already have; this exercise does not require a particular model or a research implementation.

The objective is to preserve facts that should stay fixed while making one change understandable: a full mug becomes half full after the sip. Consistency does not mean freezing every detail. It means changes have a visible cause.

Make a continuity ledger before generating

Write a short reference sheet in your own words. Lin wears a plain blue apron and a yellow wristband on the left wrist. The mug is red, with a small white dot beside its handle. In the kitchen, the kettle stays on the right side of the counter and the window stays on the left.

Use your own drawings or reference images you are entitled to use. Treat them as working references rather than proof that a tool will reproduce every detail.

The mug moves from the kettle’s left to its right intentionally. Do not mark that change as an error. An unexplained green mug or a new window position would be a continuity problem.

Give each shot an entry state and an exit state

For each shot, describe only the action it must complete. Add the relevant fixed details and the required start and finish. For shot 2, a useful brief is:

“Lin enters the hallway wearing the established blue apron and yellow left wristband. The established red mug is full and held in the right hand. Lin takes one visible sip. End with the same mug half full, still in the right hand.”

This is a suggested writing pattern, not a universal model command. Adapt it to your tool’s supported references, clip length and controls. If those controls cannot express a detail reliably, simplify the scene instead of adding more instructions indefinitely.

Review the joins, then review the whole story

First compare the last frame of shot 1 with the first frame of shot 2, and repeat for shots 2 and 3. Check face, clothing, wristband, mug identity and fill level. Then watch the full sequence without pausing. A matching still frame can hide an action that never occurs.

Keep a short issue list: shot number, observed mismatch and the smallest correction to attempt. If the wristband switches arms, revise that shot’s reference or framing. If the sip is missing, correct the action. Revisit adjacent shots when the repair changes a shared state; otherwise avoid regenerating approved shots unnecessarily.

A practical twenty-minute exercise

Spend five minutes preparing the ledger and references, ten minutes generating or assembling a rough sequence, and five minutes reviewing it. Record each attempt and your actual time or cost.

Mark each required fact as “visible and correct,” “incorrect,” or “cannot judge.” A hidden wristband is not a verified match. If an important fact cannot be judged, choose a clearer frame or remove that requirement from the story before the next attempt.

Success is a sequence you can explain shot by shot, with every important change accounted for. This exercise has no claimed benchmark result or guaranteed output. Its value is a reusable record of what to preserve, what may change and what still needs correction.