Taste-Skill

Taste-skill is the most-cited anti-slop frontend skill in the ecosystem. It pairs an opinionated design vocabulary with a hard set of "no" patterns — the failure modes that make AI-generated sites look generic. Two variants are in active use:

• taste-skill-v1 — the original, MIT-licensed, opinionated on hierarchy/type/space
• gpt-tasteskill — a GPT-tuned variant; verify license before embedding the full text on a public page

The SKILL.md files for both variants are publicly available in their respective GitHub repos. Discover the canonical publisher via skills.sh before installing.

When to use

• You're building a landing page or portfolio with an agent and want a guardrail against the generic look
• You're running a design-system bake-off and want one arm to have a known anti-slop baseline
• You maintain a design vocabulary and want to encode it as a skill rather than a doc

When not to use

• The job is a dashboard or data table — taste-skill is for marketing/landing/portfolio, not information-dense UI (the standard caveat in the source)
• You want the agent to be unconstrained — taste-skill is opinionated by design

How it works

The SKILL.md defines a small set of design vocabulary tokens (type scale, spacing rhythm, color usage rules) and a longer list of "do not" patterns — generic gradients, stock-photo hero treatments, AI-looking illustration defaults. The agent loads it on landing-page work and applies both sets as constraints. The result is a site that looks designed, not assembled.

Install

npx skills add <taste-skill-repo>

Full SKILL.md

The full SKILL.md files live in the canonical publisher's GitHub repos. Discover the publisher via skills.sh and read from the verified source:

• taste-skill-v1/SKILL.md (MIT) — the original
• gpt-tasteskill/SKILL.md — verify license before redistributing

The embedded copy is intentionally not pasted into this page at launch: the second variant has not had its license independently verified for public embed, and the first is best read in context of the upstream repo.

Verify it yourself

Quick static check — taste-skill is text-heavy and (intentionally) opinionated. Static red-flag scan still applies: confirm no curl|sh, no base64 -d, no process.env. The opinionated content is not the same as injected content; the scan distinguishes them.

Two-model review — paste the SKILL.md into two different models with the Vetting Guide prompt. The interesting case for taste-skill is whether the two models agree on the "no" patterns — if one flags a pattern as an injection class, dig in.

Review transparency

Related

• Skills Collection Overview
• Vetting Guide
• Building Antigravity + OpenCode Skills — first-person build log