generating-storyboards

Files (2)

• SKILL.md
• eval_cases.json · evals/

SKILL.md (first lines)

# The storyboard

One sheet per clip, vertical panels in a row, each panel one action, left to right. Panels are numbered
within a sheet; sheets are numbered across the video.

**The sheet is always `16:9` and its panels are always vertical**, whatever shape the finished video
is. Neither follows the delivery ratio, and neither is the plan's to set.

The plan settles how many sheets there are, what happens in each clip and how long it runs, which
images come in as references, and the look. How a clip divides into panels is settled here.

Anything the plan specifies wins. The defaults below fill gaps only.

## Workflow

**One sheet at a time.** Run all six steps for the first sheet, then run all six again for the
second, and so on. Sheets cannot be built in parallel or in one call — each prompt needs the previous
sheet's finished image in hand.

### Step 1: Collect what's needed

- **An image of the person** — wherever a person appears in any panel.
- **Product image or images** — where the video has a product.
- **The previous sheet** — from the second sheet on.
- **What happens in this clip**, and how many seconds it runs.
- **How many sheets there are**, and which one this is.
- **The look**, where the plan names one.

What a supplied image is for comes from the brief, not from its contents — a person in a frame
doesn't make it the person reference, and one image can't serve as both the person and the product.

**Read every supplied image** with `image_analysis` before deciding anything: what the place looks
like, what they're wearing, and for a product, its colours, its shape, its finish, and how big it is.
Skip this where the image is already known.

Where any of this is missing, ask for it rather than assuming it.

### Step 2: Settle the sheet

- **References** — every one goes in twice: attached as `reference_images` on the generate call, and
  labelled Image 1, Image 2, Image 3 in the prompt text. **Same order both times** — the product
  image or images, the person's image, then the previous sheet once there is one. Say once what each
  is, then refer to it by label. A label with nothing attached points at nothing, and unlabelled
  attachments let the model pick its own subject out of the set.
- **The person** — pointed at by label, never described. State that wherever they appear it is the
  same person, with face, hair, build and skin tone identical throughout. Don't describe their age,
  ethnicity, build, hair, makeup or features — the reference supplies all of it, and words describing
  it fight the image. They need not be in every panel: a tight product panel may show only their
  hands, or nobody at all.
- **Setting** — where they are, when, and how it's lit. Unless the plan sends them elsewhere, carry
  the location, the hour and the direction of the light straight from the person's image.
- **Light** — from a named direction. The plan's look sets its quality — soft and neutral, dramatic
  and low-key, or clean studio light. No golden hour, sunset, or amber cast unless the plan asks for it.
- **Wardrobe** — the garments and their colours, named. From the person's image by default.
- **Look** — the look the plan names, carried into every sheet. Where the plan names none, a clean,
  naturally-lit photographic still.
- **Product name** — one short noun phrase, used word for word in every panel that mentions it.
- **Product size** — a measurement, so it can be held against the hand at its real size.

Skip the last two when there is no product.

From the second sheet on, write two lines: what carries over from the previous sheet, pointed at by
its label, and what the plan deliberately changes. Anything not named as changed matches it exactly.

### Step 3: Divide the clip into panels

The plan gives the clip's story and its seconds. How many panels carry it is worked out here.

**One panel is one primary action** — one continuous motion at one object. Take the clip's story and
split it wherever a second motion starts or a second object is taken up: *picks it up*, *tilts it to
the light*, *puts it on* are three panels, never one line.

**Each panel runs two seconds of the clip where possible, three at most**, so the seconds set how many there are — **up to five panels, never more**. Five vertical panels are the most one sheet holds before each shrinks and its detail — faces, hands, product labels — begins to distort. Where the clip's story has more than five primary actions, don't crowd them onto the sheet: keep the five that carry the clip — its key moments, always including the opening and the closing beat — and let the rest be carried as motion inside those cuts rather than getting their own panel.

### Step 4: Write the panels

**Settle the row before writing any panel.**

The panels read left to right as one continuous story, each picking up where the last left off. On the
second sheet and after, the first panel continues from the final panel of the sheet before it.

*Across the row, these vary:*

- A different action in every panel — a genuinely different thing being done, not the same pose
  wearing a new expression.
- The camera position changes between panels, and never repeats back to back.
- The panels are shot at a mix of camera distances rather than all close, all mid or all wide; a row
  that sits at a single distance throughout needs reworking.
- The expressions move through an arc rather than repeating.

*Across the row, these hold:*

- Setting and wardrobe hold unless a beat changes them, and the change lands between panels, never
  inside one.
- A panel shows a state, not the change into it — already open, already worn, already out of the box.
  The change itself falls on the cut.
- At most one state change lands in any single panel.
- State runs one way. Once something is open, worn or emptied it stays that way for the rest of the
  video — never re-closed, never taken off again, never put back where it came from.
- Once a part is taken off the product, it leaves the picture entirely — never set down nearby, on a
  surface, or in the other hand.
- Nothing appears without arriving. Anything in a panel was either in the panel before it or came in
  on the cut just before.
- What leaves is gone from every panel after it and is never described again, on this sheet or any
  later one — a cap taken off, a wrapper torn away, a box emptied.

**Then write each panel, with all six of these.**

**1. Camera position** — held at arm's length, locked off, over the shoulder, down their own eyeline,
or whatever else the look calls for.

**2. Framing** — name the distance outright: macro, tight close-up, medium close-up, medium,
waist-up, three-quarter, full-body wide. Where the plan names a drawn look, state the distance as
composition — how much of them is in the panel — and keep every other camera word out.

**3. Action** — one physical thing happening: a single continuous motion at a single object. Never
two hand actions at once. Where it handles a product, write the hand mechanics as a sequence — which
part is gripped, what moves, what that produces — never a verb that only names it.

From SupercmoHQ/superCMO-skills, Apache-2.0.

More from SupercmoHQ/superCMO-skills