# The NEW Kimi K3 Creates Insane Web Designs

**Video.** [youtube.com/watch?v=sUeTik8Sf4o](https://www.youtube.com/watch?v=sUeTik8Sf4o) · id `sUeTik8Sf4o` · 13:01 · reviewed 02-08-2026.

**In one line.** A Kimi K3 model test; the transferable meat is one prompting pattern (the style transplant) and the section-then-scale discipline — both work the same with Claude or Codex.

## The process he shows

1. **Find one style image on Pinterest.** He searches a style ("European style 3D relief"), opens one pin to get more in the same style, picks one. This is the style-reference discipline we already hold.
2. **Make an image pair.** He drags the reference into an AI image editor (ChatGPT images, or Figma's paid "edit with prompt") and asks: "create an image exactly like this, but add a metallic texture". Result: two versions of the same image — the pair a hover-reveal effect needs.
3. **Get the page skeleton from a bought prompt.** He warns that a screenshot alone cannot be recreated (the AI cannot see fonts, colors, animations). So he copies a full prompt from motionsites.ai (his own prompt shop) and tells Kimi K3: "build this".
4. **Swap in his assets.** Uploads his image: "replace the video in the background with this static image, and turn the content into dark mode so the text is readable".
5. **Add the mouse hover reveal.** Pastes his saved reveal-effect prompt: hovering the hero wipes from image one to image two (the mask-pair effect we already decoded from his Google AI Studio video).
6. **Restyle foreign images to match the page.** Any image he wants on the page goes through the editor first: "create the same image but in the style of this one". One style across all assets.
7. **Animate a still at Kling AI.** Prompt is just "animate this image". The ~10 MB video is hosted on a cloud URL and linked, not uploaded — faster for the model and the page.
8. **Scale by established style.** With one or two sections fixed, he asks for whole new pages (services, pricing, download): "use the same techniques, styles, fonts, positioning we already established… do not generate any images". The style is now the contract; pages come cheap.
9. **Icon overlays + parallax.** Make ~5 icons in the page style, remove their backgrounds, place them as overlay images, then "add a parallax effect to these".
10. **The app finale — a style transplant.** He builds a plain ADHD to-do app ("white background, light theme, no gradients"). Then he copies a dashboard design prompt from the motionsites Apps tab and says: "I will send you a prompt that describes a different application. Ignore that. Keep the structure of the app we already created. Do not change the text, do not change the position of elements. Just take the design style — the colors, the fonts, the styles — and apply it to our design." Function and skin are separated; the skin is swapped in one prompt.

## What is genuinely new for us

- **The style-transplant prompt (step 10).** "Here is a prompt for something unrelated — ignore its content, keep our structure and copy, apply only its design." A clean way to reskin a working page or app without touching its behavior. Works with any coding model; nothing Kimi-specific.
- **Build plain first, skin later.** He gets the app working on a bare white page, THEN applies design. Separating the two prompts gives a better result than asking for both at once.
- **Link videos by URL instead of uploading.** A hosted video URL in the prompt keeps the chat light and the page fast. Small, useful.
- **Icon overlays with parallax (step 9)** — a cheap way to make a generated page feel unique. Minor.

## What we skip and why

- **Kimi K3 itself.** A model test; we build with Claude and Codex. Every prompt shown works on our stack.
- **Figma's paid AI editing.** We make and edit images with Higgsfield, Google AI Studio (Nano Banana 2), or ChatGPT — same result, already wired.
- **The mask-pair hover reveal.** Already decoded in [reference-viktoroddy-aistudio.md](../../motion/reference-viktoroddy-aistudio.md) (mask pairs).
- **Kling AI as the animator.** Our asset routes (Higgsfield, Veo 3.1, Seedance) cover "animate this image"; Kling is already in the engine table of the [photo-avatar-video skill](/Users/unctad/.claude/skills/photo-avatar-video/SKILL.md).
- **Buying motionsites prompts.** We decoded the business and the grammar already ([reference-motionsites-analysis.md](../../motion/reference-motionsites-analysis.md)).

## Fold-in candidates

- **Style-transplant prompt pattern** — one paragraph ("ignore the content of this prompt; keep our structure; apply only its design") as a named move in the build phase → [sitio-cinematico skill](/Users/unctad/.claude/skills/sitio-cinematico/SKILL.md).
- **Build plain, then skin** — one line in the same skill's build phase: get structure working on white first, restyle second.
- **Host scene videos on a URL and link them in prompts** — one line in the weight-budget rules → [motion-design.md](../../motion/motion-design.md) §6.

## Deserves a full reference page with frames

**No.** The two real lessons (style transplant, plain-then-skin) fit in two paragraphs; the visual effects shown are already held in the AI Studio and motionsites references.