Home/AI Design Analysis/Taste
T

Taste

Updated: Nov 17, 2026

Turn any website into a design brief your AI agent can build from β€” exact tokens plus the reasoning and trade-offs behind every choice, written as one file per tool.

Image not available

About Taste

Overview

Taste is a skill for Claude Code and Gemini CLI that turns any website URL into a complete design context for an AI agent. It reads a page through four separate agent passes and records more than tokens: it captures the decisions and trade-offs behind each design choice.

Running /taste on a URL writes two files per domain, a .md brief and a .json file. The output pairs a Design Map of measured tokens with a Taste DNA of four principles, each explaining why one choice was made and what it gave up. The skill lives in a public GitHub repository.

Key Benefits

  • Measures 20 design categories and cites every color, weight, spacing value, radius, and shadow with exact px, hex, or ratio.
  • Extracts 5–8 system-level rules, each carrying an Evidence line and a Design Goal that explains why the pattern exists.
  • Produces four Taste DNA principles with Trigger, Decision, Reason, Evidence, and Trade-off, and requires at least one Restraint principle.
  • Formats the result for each target tool, so Cursor, Windsurf, Copilot, Bolt, Antigravity, v0, Figma Make, and Lovable pick it up on their next run.
  • Runs a quality gate that greps for filler language and validates the JSON before writing final files.
  • Uses a single extra file per domain, with no pasting or manual handoff.

How It Works

You clone the GitHub repository into your Claude skills directory or your Gemini skills directory, add the Playwright MCP server, and restart the CLI. You then run /taste against any URL. Four agents take over from there: one extracts precise measurements, one detects systematic rules in those measurements, one infers the taste and its trade-offs, and one filters the analysis before a quality gate writes {domain}.md and {domain}.json.

Use Cases

  • Frontend engineers run /taste on a reference site to pull exact colors, type scale, spacing, and radius values into a build.
  • Design leads document why a competitor or inspiration site looks the way it does, capturing its rules as a reusable file.
  • AI agent developers drop the generated rules file into Cursor or Windsurf so the agent follows a target site's design language.
  • Prompt engineers load the output into v0 or Lovable to steer generated UI toward a specific design system.
  • Figma Make users take the generated CSS plus instructions to rebuild a visual style inside Figma.
  • Design researchers cite px, hex, and DOM evidence instead of describing a page by impression.

Why Choose This Product

Taste treats a token dump as a starting point rather than an answer. Every principle in the output ties back to evidence from the page and names the trade-off the design accepted, and the pipeline enforces at least one Restraint principle. The result arrives as one file per tool, which is what lets an agent act on it on the next run.

Taste Pros & Cons

Strengths
  • Measures 20 design categories with exact px, hex, and ratio values
  • Outputs both a .md brief and a .json file per domain
  • Writes pre-formatted rules files for Cursor, Windsurf, Copilot, and more
  • Pairs every principle with cited page evidence and a stated trade-off
  • Setup is a git clone plus one Playwright MCP configuration step
Limitations
  • Runs only in Claude Code and Gemini CLI
  • Playwright MCP setup downloads roughly 100MB of Chromium

Key Features

πŸ—ΊοΈ

Design Map Extraction

Records 20 measurement categories with exact px, hex, and ratio values for every color, weight, spacing value, radius, and shadow on a page.

🧬

Taste DNA

Delivers four design principles that each pair a Trigger, Decision, Reason, Evidence, and Trade-off, with at least one Restraint principle required.

πŸ€–

Four-Agent Pipeline

Runs extraction, pattern detection, taste inference, and a final filtering pass, with each agent reading the same page through a sharper lens.

πŸ“

System Rule Detection

Surfaces 5–8 system-level design rules, each with an Evidence line and a Design Goal explaining why the pattern exists.

πŸ“¦

Dual File Output

Writes both a {domain}.md brief and a {domain}.json file on every run.

πŸ”Œ

Tool-Ready Formats

Formats the output for Cursor rules, Windsurf rules, CLAUDE.md, Copilot instructions, Bolt prompts, GEMINI.md, v0 tokens, Figma Make CSS, and Lovable project knowledge.

βœ…

Quality Gate

Runs an anti-slop grep and validates the JSON before writing the final output files.

πŸ–₯️

CLI Skill Install

Installs with a git clone into the Claude Code or Gemini CLI skills directory and then runs as /taste on any URL.

Frequently asked questions about Taste

What does Taste output?

Every run writes two files: a {domain}.md brief and a {domain}.json file. Together they contain the full token set from the page plus 4 principles explaining why each design choice was made and what it gave up.

How is Taste different from a token extraction tool?

Extraction tools return a spec sheet β€” background #08090A, font Inter, radius 6px. Taste also records the decision (why this near-black instead of pure black) and the trade-off (what was chosen and what was deliberately left out), so an agent understands each choice instead of copying numbers.

How do I install Taste?

Clone the GitHub repository into your Claude skills directory or your Gemini skills directory, add the Playwright MCP server to your config, then restart the CLI. After that you run /taste followed by any URL.

Does Taste work with Gemini CLI?

Yes. The page documents a Gemini CLI path: clone into ~/.gemini/skills/taste, add Playwright to the mcpServers section of ~/.gemini/settings.json, restart, then run /taste on a URL.

Which tools can use the generated output?

The page lists Cursor, Windsurf, Claude Code, GitHub Copilot, Bolt, Antigravity, v0 by Vercel, Figma Make, and Lovable. Each receives the same taste in a format it already reads, so the agent picks it up on its next run without pasting.

What happens in each of the four pipeline steps?

Step 1 extracts precise, objective design measurements from a webpage. Step 2 detects systematic rules in those measurements. Step 3 infers taste and the deliberate trade-offs a designer made. Step 4 ruthlessly filters that analysis before output.