Week 3 Set: The shot list exercise

Third set — translating a one-paragraph scene into a structured prompt sequence, then checking what the model actually did with your blocking

8 March 2027

Before the session#

Bring a one-paragraph scene description — two people, one location, one small action (a handoff, an argument, a decision). Write it like a scene, not a prompt: what happens, in what order, who moves where.

In the session#

Round one — break it into coverage (25 min). A director doesn’t hand a crew one paragraph and hope; they break it into shots and specify how each one is framed. Do the same to your scene as a numbered prompt list: a wide establishing shot, at least one closer coverage shot on each character, and one shot that carries the scene’s turn (the handoff, the reaction). For each prompt, name the shot type and camera move explicitly — “static wide,” “slow dolly in,” “handheld tracking behind” — the same vocabulary a shooting script would use, because that’s the vocabulary these models were trained against captioned footage that used it too.

Round two — generate and compare (25 min). Run your prompts through whatever generation tool the set has access to this week. For each output, check it against your blocking: did the character end up on the side of frame you specified? Did the second shot’s lighting and wardrobe match the first one’s, or did the model quietly restart the scene from a slightly different prior? Log every divergence with a timestamp, the same way week 2’s tell-logs worked — “shot 3, 0:04, character’s jacket changes colour from shot 1” is an entry; “it doesn’t really match” is not.

Round three — diagnose the gap (10 min). For your worst divergence, decide whether it’s a prompting failure (you didn’t specify something the model needed) or a continuity failure the medium can’t currently fix (no memory between separate generations of what a specific character or object looked like a shot ago). Both are useful findings; only one of them a better prompt can solve.

Afterwards#

Keep the shot list and the divergence log — week 7’s pipeline-mapping exercise asks you to reverse-engineer someone else’s generation process, and today is the version of that exercise where you already know the answer because you wrote the prompts yourself.