writing-video-prompts

Files (8)

• SKILL.md
• eval_cases.json · evals/
• prompt-gemini-omni.md · references/
• prompt-grok.md · references/
• prompt-kling.md · references/
• prompt-seedance.md · references/
• prompt-veo.md · references/
• prompt-wan.md · references/

SKILL.md (first lines)

# Clip prompts

Turn a clip into a prompt its video model will honour. **Text only — nothing is generated here.**

**A video may be one clip or several.** Write a prompt for every clip in it, then hand them all back
together. Clips in one video share a model, an aspect ratio and a look.

Write only what the clip actually has. No person on camera means no performance direction, no product
means no product rules, no speech means no dialogue.

## Workflow

### Step 1: Read what the clip has

Ask only for what would change the prompt and can't be defaulted.

- **The model and the duration.** Where no model was named, use `seedance-2.0-fast` and say so. The
  aspect ratio is a call parameter rather than prompt text — carry it if it was given, don't ask.
- **What happens in the clip** — as beats, or as a single action.
- **The media, and what each one is for** — a storyboard sheet, a start frame, an end frame,
  reference images, a source video, an audio track. **Any of them may be absent**, and they differ from
  clip to clip.
- **The dialog segment**, where anyone speaks, and whether they are on camera or heard over the
  picture.

### Step 2: Pick the clip's form

**With a storyboard sheet, one cut per panel** — panel 1 is the first cut, panel 2 the second, hard
cut between them. 

Without a sheet, cut only where the brief asks for one.

**Where a beat's opening frame is fixed** — by a panel or by a start frame — the hands, the grip, the
product, the props and the positions are already there when it opens. Write what happens next, not
how it starts.
Write "a hand reaches in" or "she brings it into frame" only when the frame genuinely has no such
element. Introduce something that is already there and the model renders a second one — an extra
hand, a duplicate arm, a doubled person.

### Step 3: Read the chosen model's guide

| Model | Guide |
| --- | --- |
| `seedance-2.5` | `references/prompt-seedance.md` |
| `seedance-2.0` | `references/prompt-seedance.md` |
| `veo-3.1` | `references/prompt-veo.md` |
| `gemini-omni` | `references/prompt-gemini-omni.md` |
| `gemini-omni-1.1` | `references/prompt-gemini-omni.md` |
| `kling-3.0-pro` | `references/prompt-kling.md` |
| `grok-imagine-video` | `references/prompt-grok.md` |
| `wan-2.7` | `references/prompt-wan.md` |
| `wan-3.0` | `references/prompt-wan.md` |

The `-fast` / `-lite` / `-turbo` / `-standard` / `-mini` variants share their family's guide.

**The guide owns how the prompt is laid out** — the order the elements go in, how references are
labelled, how speech and sound are marked, and how to keep something out of frame. Follow it on all
of those, over anything the steps below imply about layout.

It does not own which sections the prompt is built from or the order they come in. Step 6 fixes those, and they are the same whatever the model.

**It does not decide whether the clip cuts.** Step 2 already did, from the media. A guide showing how
to write a sequence of shots is giving you the layout to use *if* the clip has one — not a reason to
add cuts to a clip that doesn't.

### Step 4: Label the media

Where more than one piece of media goes on the call, name each one in the prompt, in the order it
will be passed, and say once what each is. Unlabelled, the model picks its own subject out of the
set. **The labelling convention comes from the model's guide** — a numbered slot, a tag, or plain
words. A clip carrying a single start frame and nothing else has nothing to label — just say what it
opens on.

- **A storyboard sheet** — labelled once at the top as the storyboard for this clip. Don't contradict
  it and don't caption it: a sentence that would work as a caption for the panel is a description of
  a still, so write that panel's beat in motion instead.
- **A start or end frame** — named as the frame the clip opens or closes on.
- **The person** — labelled on first mention in a beat, then the plain noun. Repeat the label later
  in the same beat and the render snaps back to the reference, breaking continuity.
- **The product** — pointed at by its label and its name, never described; words about its colour, finish, shape or markings compete with the image and the sheet, both already on the call. A fuller description from the plan is for your reference, never for the prompt.
  Label it where it is shown intact: first introduction, a
  hero close-up, a label point-out. Drop the label during any verb that changes it — opening, tearing, pouring, scooping,
  applying, drinking, mixing, putting on — because carrying it through forces the product back to its
  reference state mid-beat, so an opened jar snaps shut. Pick the label up again once it is shown
  intact.
- **Several product images** — point at whichever shows the angle that beat needs.

**Never put a label inside quoted speech.** Quoted text is read aloud as written, so the label gets
spoken.

### Step 5: Read the storyboard

Skip this where the clip has no sheet.

Run `image_analysis` on the sheet before writing anything, and take these from each panel as it
actually rendered:

- the camera position — held at arm's length, locked off, over the shoulder
- the framing distance
- the action
- which hand is doing what
- where the product is — held, set down, worn, hidden, absent
- the expression

**Where the sheet and the plan disagree, the sheet is what the model will see, so the sheet wins.**
Write the cut to the panel, and say which panel disagreed and how.

### Step 6: Write the clip

**Every prompt is built from the same sections, in this order:**

1. **The look** — 6a below.
2. **What the clip is** — 6b below.
3. **The beats** — 6c to 6e below.
4. **The audio** — Step 7.
5. **The closing block** — Step 8.

The guide names each section and marks what goes in it. Leave out a section the clip has no use for;
never reorder them, and never fold one into another.

From SupercmoHQ/superCMO-skills, Apache-2.0.

More from SupercmoHQ/superCMO-skills