Teach AI what good design looks like to you.

Taste started as a joke: an app where you could lick your screen to save a design you loved. Then people actually used it, and it turned into a product used by thousands of designers and builders.

The first job was practical. AI coding tools default to generic UI, so Taste turns the screenshots you admire into a structured profile that Codex, Claude Code and Cursor design from, and your agents ship something closer to what you would have.

Then I realized what all that activity adds up to. Every upload, poll, rating, judgment and annotation is a human expressing taste in a structured way. That makes Taste a source of high-quality design preference data, the kind frontier labs need to teach models what good design looks like.

Design decisions

01
Save from wherever you find design
Taste lives on five surfaces: the web, a Mac app with a global hotkey, an iOS app on the share sheet, a Chrome extension and a Figma plugin. Saving is deliberately boring. No tagging, no folders, no curation. Everything lands in one library, and every save becomes a labeled example.
Teach your AI agents how you think about design
Five surfaces, one library.
02
Taste as structured data
A GPT-4o vision pipeline extracts more than 40 signals from every capture: colors with semantic roles, the type scale, spacing rhythm, shape language, density, WCAG contrast on real text pairs, and bounding-boxed UI elements. Across a library, those collapse into a versioned Taste Profile with per-category scopes.
Taste for iOS: capture, rate, and browse community designs.
03
Capturing judgment, not just likes
A save tells you what someone likes. Judgment tells you why. I've built several ways to capture it, each trading a little effort for richer signal: short interviews about a design, ratings, Tinder-style swipes, and now polls. A poll shows you a design and asks what you'd change. You pick between two options, or choose a third to make your own change. Together they turn taste into structured preference data instead of a pile of bookmarks.
04
What you reject matters as much as what you love
Every profile has an explicit anti-preferences block. Rate any AI-generated screen nailed it, close, or missed, and that feedback feeds the next synthesis pass. It's prompt-level RLHF for your taste, with no fine-tuning.
~/.claude/skills/taste/SKILL.md
```
## Anti-Preferences
- Dislikes overly muted or low-contrast designs
- Dislikes busy layouts with too many buttons
- Dislikes sharp corners
```
An excerpt of my own exported profile. This site was designed with an agent reading it.
05
Live where the AI reads it
Taste should never be the thing you stare at. The profile exports as a Codex or Claude skill and runs as an MCP server with OAuth 2.0, rotating tokens and a full audit log, so it becomes live context on every prompt. Calibration, not prompting.

Highlights

Shipped across macOS, iOS, web, a Figma plugin, a Chrome extension and an MCP server
Vision pipeline extracts 40+ structured design signals per screenshot
Preference data as 6-dimensional taste vectors and Elo-scored pairs
Supabase backend: Postgres with RLS, pgvector, twelve edge functions

SuperlocalA map that reveals itself as you explore

Save from wherever you find design

Taste as structured data

Capturing judgment, not just likes

What you reject matters as much as what you love

Live where the AI reads it