Overview
Design-taste is a research layer that grounds UI decisions in real products before an agent writes code. It uses Kage MCP (free, keyless at https://kage.design/mcp) to find screenshots and stored layout prompts for comparable apps, extracts concrete decisions (layout, typography, spacing, colour), then synthesizes a brief for the build. Research is mandatory — an agent that skips it defaults to the same generic Tailwind layout every time.
Refero (paid MCP, MIT skill) teaches the same workflow with 150k+ app screens and flows; Kage covers the same shape of research with 600+ web products and no auth. Use whichever endpoint you can reach.
When to Use
- User says "design X like Y" — a named reference product
- User says "in the spirit of" or "similar to" without naming a product
- User asks "what should this look like" or "how should this feel"
- New surface, no existing design language in the repo
- Competitor teardown: "what do Linear, Notion, Arc do here?"
- Before delegating a build to
hallmark, impeccable, or anti-slop-frontend
When NOT to Use
- DESIGN.md already exists and tokens are locked — go straight to
impeccable document + build
- User has a Figma frame or a screenshot of the target — use
image-to-code directly
- Task is pure component work in an existing system — use
shadcn-ui or 21st-dev
- User wants accessibility fixes only — use
accessibility-audit
- User wants a rendered audit of existing code — use
ui-critique
Workflow
1. Frame the brief (no tools yet)
Write three lines before calling anything:
- What — the surface (landing, pricing, onboarding, dashboard)
- Who — the audience and the job they hire the product for
- Constraint — stack, brand colour, platform, deadline
Ambiguous brief? Ask one question. Do not guess product positioning.
2. Query Kage MCP
Endpoint: https://kage.design/mcp (Streamable HTTP, no auth). Add to .cursor/mcp.json or claude mcp add --transport http kage https://kage.design/mcp.
Tools:
design_brief — closest matches to a text brief; returns stored prompts + screenshots
search_designs — search the 600+ library by product/style/industry
get_design — one product's full prompt, taxonomy, extracted components with prompts, and the screenshot
get_component — one extracted component (hero, pricing, footer) with its prompt and crop
list_taxonomy — vocabulary of page types, styles, industries, stacks, colours
{
"mcpServers": {
"kage": { "url": "https://kage.design/mcp" }
}
}
Example call: "Design a landing page like Linear for a dental clinic" → design_brief returns Linear's stored prompt (layout, type, spacing, colour only — no brand or copy) plus 2–3 comparable screenshots.
3. Synthesize, do not average
Pick one primary reference. Borrow at most one narrow detail (a hero rhythm, a nav treatment) from a secondary. Do not blend three products into a beige average.
For each borrowed decision, write one line:
Layout: hero left-aligned + dense feature bento ← from Linear
Type: Public Sans, 72/16/12 scale ← from Stripe
Colour: #0A0A0A surface, single accent ← from Vercel
Motion: 200ms ease-out, no parallax ← justified restraint
If no MCP is available (Kage down, restricted environment), fall back to the methodology alone: name 2–3 comparable products from memory, extract layout/type/colour from what you know of them, and state clearly in the brief that no live research was run.
4. Lock the direction, then build
Save the synthesis as a design-brief.md beside the surface (or in PRODUCT.md per impeccable). Then delegate the actual build:
- Full page →
hallmark (Floor enforcement) or impeccable (24 commands)
- Craft constraints only →
anti-slop-frontend
- Component assembly →
21st-dev + shadcn-ui
Verification of the built page is owned by the downstream skill, not this one.
Anti-Rationalization Table
| Rationalization |
Reality |
| "I already know what Linear looks like" |
Memory of a product is a vibe, not a spec. Query MCP; you'll find spacing and type values you'd have gotten wrong. |
| "MCP is down, I'll skip research" |
Do the synthesis from memory and say so in the brief. Silent skip = default Tailwind layout. |
| "Three references make it richer" |
Three references average into generic. Pick one primary, borrow narrowly from others. |
| "The screenshots are the design" |
Prompts describe layout/type/colour, not brand or copy. Building a clone is a ToS and craft violation. |
| "Research is overhead for a small component" |
For a single button, fine. For any surface with a hero, pricing, or onboarding flow — research first. |
Verification
References
1---2name: design-taste3description: Use when building a UI and need to ground taste in real products — query Kage MCP for screenshots and layout prompts of comparable apps, then synthesize a design brief before writing code. Use when "design a site like X", "UI in the spirit of Y", or "what should this look like" — a taste-first research layer for agents, not a template gallery.4domain: content5author: oyi776license: MIT7subdomain: design8category: content9tags:10 - design11 - ui12 - taste13 - inspiration14 - research15 - mcp16 - screenshots17 - prompts18version: 1.0.019triggers:20 - "design a site like"21 - "ui in the spirit of"22 - "design taste"23 - "ui inspiration"24 - "like linear for"25 - "what should this look like"26 - "find design references"27 - "competitor ui"28 - "design brief from references"29argument-hint: "[product_or_style] for [audience]"30metadata:31 source: https://skillmd.com/skills/oyi77/design-taste32---3334## Overview3536Design-taste is a research layer that grounds UI decisions in real products before an agent writes code. It uses Kage MCP (free, keyless at `https://kage.design/mcp`) to find screenshots and stored layout prompts for comparable apps, extracts concrete decisions (layout, typography, spacing, colour), then synthesizes a brief for the build. Research is mandatory — an agent that skips it defaults to the same generic Tailwind layout every time.3738Refero (paid MCP, MIT skill) teaches the same workflow with 150k+ app screens and flows; Kage covers the same shape of research with 600+ web products and no auth. Use whichever endpoint you can reach.3940## When to Use4142- User says "design X like Y" — a named reference product43- User says "in the spirit of" or "similar to" without naming a product44- User asks "what should this look like" or "how should this feel"45- New surface, no existing design language in the repo46- Competitor teardown: "what do Linear, Notion, Arc do here?"47- Before delegating a build to `hallmark`, `impeccable`, or `anti-slop-frontend`4849## When NOT to Use5051- DESIGN.md already exists and tokens are locked — go straight to `impeccable document` + build52- User has a Figma frame or a screenshot of the target — use `image-to-code` directly53- Task is pure component work in an existing system — use `shadcn-ui` or `21st-dev`54- User wants accessibility fixes only — use `accessibility-audit`55- User wants a rendered audit of existing code — use `ui-critique`5657## Workflow5859### 1. Frame the brief (no tools yet)6061Write three lines before calling anything:62- **What** — the surface (landing, pricing, onboarding, dashboard)63- **Who** — the audience and the job they hire the product for64- **Constraint** — stack, brand colour, platform, deadline6566Ambiguous brief? Ask one question. Do not guess product positioning.6768### 2. Query Kage MCP6970Endpoint: `https://kage.design/mcp` (Streamable HTTP, no auth). Add to `.cursor/mcp.json` or `claude mcp add --transport http kage https://kage.design/mcp`.7172Tools:73- `design_brief` — closest matches to a text brief; returns stored prompts + screenshots74- `search_designs` — search the 600+ library by product/style/industry75- `get_design` — one product's full prompt, taxonomy, extracted components with prompts, and the screenshot76- `get_component` — one extracted component (hero, pricing, footer) with its prompt and crop77- `list_taxonomy` — vocabulary of page types, styles, industries, stacks, colours7879```json80{81 "mcpServers": {82 "kage": { "url": "https://kage.design/mcp" }83 }84}85```8687Example call: `"Design a landing page like Linear for a dental clinic"` → `design_brief` returns Linear's stored prompt (layout, type, spacing, colour only — no brand or copy) plus 2–3 comparable screenshots.8889### 3. Synthesize, do not average9091Pick **one** primary reference. Borrow at most one narrow detail (a hero rhythm, a nav treatment) from a secondary. Do not blend three products into a beige average.9293For each borrowed decision, write one line:94```95Layout: hero left-aligned + dense feature bento ← from Linear96Type: Public Sans, 72/16/12 scale ← from Stripe97Colour: #0A0A0A surface, single accent ← from Vercel98Motion: 200ms ease-out, no parallax ← justified restraint99```100101If no MCP is available (Kage down, restricted environment), fall back to the methodology alone: name 2–3 comparable products from memory, extract layout/type/colour from what you know of them, and state clearly in the brief that no live research was run.102103### 4. Lock the direction, then build104105Save the synthesis as a `design-brief.md` beside the surface (or in `PRODUCT.md` per `impeccable`). Then delegate the actual build:106- Full page → `hallmark` (Floor enforcement) or `impeccable` (24 commands)107- Craft constraints only → `anti-slop-frontend`108- Component assembly → `21st-dev` + `shadcn-ui`109110Verification of the built page is owned by the downstream skill, not this one.111112## Anti-Rationalization Table113114| Rationalization | Reality |115|---|---|116| "I already know what Linear looks like" | Memory of a product is a vibe, not a spec. Query MCP; you'll find spacing and type values you'd have gotten wrong. |117| "MCP is down, I'll skip research" | Do the synthesis from memory and say so in the brief. Silent skip = default Tailwind layout. |118| "Three references make it richer" | Three references average into generic. Pick one primary, borrow narrowly from others. |119| "The screenshots are the design" | Prompts describe layout/type/colour, not brand or copy. Building a clone is a ToS and craft violation. |120| "Research is overhead for a small component" | For a single button, fine. For any surface with a hero, pricing, or onboarding flow — research first. |121122## Verification123124- [ ] Brief names one primary reference and at most one secondary detail-borrow125- [ ] Layout, typography, spacing, and colour each traced to a source (live MCP or named-from-memory)126- [ ] No product copy, logo, or brand asset reproduced127- [ ] Build delegated to the correct downstream skill (`hallmark` / `impeccable` / `anti-slop-frontend` / `21st-dev`)128- [ ] If research was skipped or MCP failed, the brief says so explicitly129130## References131132- Kage library and taxonomy: <https://kage.design>133- Kage MCP setup: <https://kage.design/mcp-setup>134- Kage terms (prompts free for commercial use, no attribution; no bulk scraping): <https://kage.design/terms>135- Refero skill (MIT, same methodology, paid MCP): <https://github.com/referodesign/refero_skill>136- Google DESIGN.md spec (format the brief can target): <https://github.com/google-labs-code/design.md>