# /taste, Yours to Fork

Run a binary visual review on any HTML file or URL. Ten pass/fail evaluations plus a 34-principle constitution sweep. Reports a scored table with specific violations and an optional auto-fix.

## Why fork it before you run it

A taste audit doesn't work out of the box. It works after you've fed it failure cases and corrected its read of each one. The binary evals (T1-T10) are universal enough to use as-is. The constitution principles are a different story, mine were calibrated against my own taste over 30+ wins and 40+ corrections. Yours should reflect what you find beautiful, not what I find beautiful.

The customize prompt walks you through 6 questions that re-aim the constitution at your aesthetic.

## What's in this folder

- `taste.md` - the skill itself, with `{{PLACEHOLDER}}` markers and the full 34-principle constitution
- `customize-prompt.md` - paste this into Claude Code along with `taste.md`. It asks 6 questions and writes a customized version to `~/.claude/skills/taste/SKILL.md`
- `tuning-history.md` - the iteration log. How my version of /taste got fine-tuned to where it is now. Useful as a model for what to capture as you iterate on yours.

## What was stripped

- References to specific shipped sites I use as exemplars
- The Pharadoxa palette (cyan / violet / amber / rose) is mentioned as an example only, not hardcoded
- My banned punctuation list (em dash, smart quotes, etc.) is a default but you can override
- File paths to my taste-library and constitution
- DESIGN.md integration with my project's `npx @google/design.md` lint command (kept as optional)

## What's still in

- The full T1-T10 binary eval set (these are universal taste checks: typography, contrast, density, restraint, motion)
- The 34-principle constitution, organized by domain (Visual Design, Code Architecture, UX Interaction, Content Quality)
- The Playwright screenshot capture logic at desktop (1440x900) and mobile (375x812)
- The auto-fix protocol that targets failing evals in priority order
- The scoring rubric (9-10 = ship, 7-8 = polish, 5-6 = enrich, < 5 = rework)

## How to install

1. Read `tuning-history.md` first. Five minutes. It shows you what kind of corrections actually moved the skill forward, useful before you customize.
2. Open `customize-prompt.md`. Copy it.
3. Open Claude Code. Paste the customize prompt. Then paste `taste.md` underneath. Send.
4. Answer the 6 questions in plain language.
5. Claude Code writes the result to `~/.claude/skills/taste/SKILL.md`.
6. Run `/taste path/to/file.html` or `/taste https://example.com` against any build to get a scored audit.

## When to use it

- Before you ship a build (catches AI-generated patterns reviewers don't always articulate)
- After a calibration session where you fixed corrections (validates the fixes landed)
- On a competitor's site (instant read on what they get right and where they're weak)
- As a self-audit on your own portfolio (uncomfortable, useful)

## A word on the constitution

The 34 principles in `taste.md` are mine. They're opinionated. Some examples:
- "Accent color is medicine, not seasoning. Below 15% surface area."
- "Custom easing curves, never CSS defaults. `ease`, `ease-in-out`, `linear` are tells."
- "Three similar lines beat a premature abstraction."
- "Vague headlines are a build failure."

You will disagree with some of these. That's expected. The customize prompt explicitly asks which principles you want to keep, edit, or strip. Bring your own taste. The skill works either way, the constitution is the layer that personalizes it.

The deeper play: every time you give Claude a correction during /pilot or /taste, that correction is a candidate principle. Capture it. Add it. Over time, the constitution becomes the residue of your judgment, written down so the next build inherits it for free.