The problem it solves is not a video problem

A film exists on paper. It needs money, or a cast attachment, or a network to say yes.

Everyone in the room can read. Nobody can see it. And a script does not survive being read by someone tired, on a plane, with four others.

Why this is the best fit in the whole course

Before generative tools, a sizzle was built from other people's footage. Scenes from films with a similar feel, cut together, with a temp score.

That has two permanent problems:

• It shows what your film is like, never what it is
• You can never use it publicly, because you do not own a frame of it

Generated footage fixes both. It is your world, your face, your light, and it is yours.

The three things it has to carry

Plot is the thing people put in and the thing that kills it. Ninety seconds cannot hold a story. It can absolutely hold a feeling.

The structure that works

Four movements, and it is closer to a music video than a film:

1. The world, established in three or four shots. No people yet, or one, small in frame.
2. The person, and the problem, without explaining either.
3. The turn. The thing that makes it a film rather than a mood. Faster, louder, or suddenly silent.
4. The image you want them to still have in their head afterwards. One shot. Hold it.

"From story/treatment.md, give me a sizzle structure in four movements. For each: what it has to make the viewer feel, and four to six shots that would do it. Do not tell the plot."

Voice over, and when to use none

If the world needs one fact to make sense, a single line of narration at the top buys it.

Two lines is a documentary. Four is a pitch deck being read aloud.

The title card is doing more work than you think

The last four seconds are the title, and often one line under it.

That line is the closest thing to a logline the room will accept, and it is what they repeat to someone else afterwards.

"Give me twelve one-line versions of the card under the title. No adjectives, no questions, under nine words each."

Then choose one yourself. Twelve options in seconds, one decision by you: the shape of every good use of this tool.

Nothing new to learn

Modules 2 through 8, on a shorter piece with a tighter constraint:

• Module 2 gives you the shot list, from the treatment instead of a script
• Module 5 matters most here: a drifting face reads as amateur instantly
• Module 6 carries the emotion, and the music is the biggest single lever
• Module 8 because it gets watched on a phone, at the wrong volume

A sizzle is the smallest complete version of everything you already built.

The constraint that makes it good

Twelve to eighteen shots. Not forty.

Fewer shots, held longer, cut on the music. A sizzle that breathes reads as confidence. One that cuts every twelve frames reads as hiding.

The one where the look file earns everything

A sizzle is judged almost entirely on whether it looks like one thing made by one person.

"Build look.md for this piece from these six reference frames, then assemble every shot prompt from it, using its exact words. Then list any shot whose prompt drifts from the look."

One look, twelve shots, no exceptions. This is the whole difference between something that gets forwarded and something that gets closed.

Deliver it for the room it will be watched in

Not the room you imagine. The real one:

• A phone, at arm's length, in daylight
• A laptop at forty percent volume
• Occasionally a real screen, which it must also survive

"Export the sizzle at my spec, plus a vertical version, plus a version with burned captions for the dialogue. QC all three and give me the loudness numbers."

Burned captions are not optional. It will be watched on mute more often than not.

Know what is in your film

You will be asked. By a financier, a network, a festival, or a platform.

"List everything in this piece that was generated: which model made each shot, which voice model, any reference image and where it came from. Write it to PROVENANCE.md with dates."

The exercise

Your own project, or one you wish existed. Two days, not two weeks.

1. *"From my treatment, a sizzle structure in four movements. What each must make them feel, and

four to six shots for each. No plot."*

1. Cut it to sixteen shots yourself. This is the hardest and most important step.
2. "Build look.md from these references, assemble every prompt from it, flag any drift."
3. Generate stills. Approve nothing you are not certain about.
4. Motion queue overnight, on the approved frames only.
5. "Music map: four cues, one per movement, with lengths."
6. Cut it to the music. Hold your shots longer than feels comfortable.
7. "Twelve versions of the line under the title, under nine words." Choose one.
8. Export, QC, burned captions, vertical. Write PROVENANCE.md.

What you should have now

• Ninety seconds that make a script feel like a film
• A look file and a bible that make the second piece for this project fast
• Every prompt and seed, so a note costs one row
• PROVENANCE.md, complete, before anyone asks for it
• Three exports that survive a phone, a laptop and a real screen

You can now make the thing that gets an unmade film made.

Verified against Claude Code 2.1.220 · 2026-07-30.