Style Contract

Style Contract - Visual Taste Guardrail

This skill creates a written contract for how a storyboard or launch-video system should look and feel.

It is used before generation or regeneration so the agent does not chase one frame at a time and lose cohesion.

Activation

Use this skill for requests like:

• "get the vibe right"
• "this looks amateurish"
• "take inspiration from these references"
• "do not copy this verbatim"
• "make it more product launch SaaS motion graphic"
• "the UI is nice but the generated text kills it"
• "background should be cleaner / more gradient / less graphic"

Do not use this skill for file batching, video prompting, or frame QA unless paired with those skills.

Inputs To Study

Inspect only the relevant materials:

• product screenshots
• accepted storyboard frames
• rejected storyboard frames
• brand kit and logo assets
• reference images or videos
• written taste notes from the user
• target video model limitations

Summarize what the product itself already does well before adding external style.

Contract Sections

Write the contract in these sections.

Product Visual Truth

Capture what must stay faithful:

• actual UI components
• real logo and mark
• button shapes
• border radius
• color palette
• gradients and glows
• type weight and scale
• density and spacing
• any distinctive product modules

Borrow From References

Name transferable traits only:

• pacing
• spaciousness
• gradient handling
• glass depth
• type hierarchy
• negative space
• camera restraint
• clean reveal behavior
• module choreography

Do Not Borrow

Name what must not be copied:

• reference brand names
• reference product concepts
• unrelated UI controls
• fake icons
• text strings
• layouts that do not fit the user's product
• 3D/realistic footage if the desired style is product motion

Composition Rules

Define recurring structure:

• full-bleed stage or clean gradient background
• one primary UI element per frame when possible
• copy can sit separate from UI
• no nested cards unless product UI actually has them
• use crop, scale, and edge placement for motion implication
• preserve breathing room

Text Policy

Set a strict text mode:

• no generated text
• minimal exact text
• real UI text only
• copy added later in editor

For AI image generation, prefer fewer words. For AI video, treat important text as risky unless it is already baked into a stable frame.

Motion Feel

Describe the implied motion:

• smooth ease-in/out
• controlled scale
• opacity reveals
• soft glow pulses
• subtle parallax
• card/module slide
• chat bubble stagger
• final hold

Avoid vague terms like "dynamic" unless paired with exact motion mechanics.

Avoid List

Build a project-specific avoid list from user feedback:

• amateur collage
• random background graphics
• fake dashboard labels
• placeholder bars
• mutated logos
• gibberish text
• unrealistic stadium/crowd footage when product UI should lead
• over-rotated UI if the video model cannot resolve it
• generic centered SaaS cards

Output Format

Default:

Style Contract: <project / sequence name>

Product Visual Truth:
- ...

Borrow:
- ...

Do Not Borrow:
- ...

Composition:
- ...

Text:
- ...

Motion:
- ...

Avoid:
- ...

If the user is about to generate frames, end with a compact generation directive that can be reused across prompts.

Regeneration Use

When a frame fails, do not rewrite the whole style. Identify which contract rule it violated:

• too graphic
• too literal UI screenshot
• not enough product truth
• copied reference
• fake/mutated logo
• text pollution
• bad density
• video-model unstable angle

Then produce a narrower regeneration prompt.

What Not To Do

• Do not generate images.
• Do not manually edit images.
• Do not turn every reference into mandatory rules.
• Do not copy reference compositions verbatim.
• Do not let a single nice frame overwrite the contract for the whole sequence.