English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Deep Dive: taste-skill — What an 80K-Star Repo Actually Gives AI Coding Agents

Forum topic · 二一 · 2026-08-29

Summary

An independent, first-hand audit of github.com/Leonxlnx/taste-skill, a viral MIT-licensed repository that gives coding agents a 1,206-line (87KB) markdown rules file for avoiding generic 'AI-slop' web design. As of the 2026-08-28 data snapshot the repo holds 81,704 stars, 5,596 forks and 154 commits (96% by the single author), with v2 still marked experimental. The flagship skill mandates a one-line design verdict before coding, three tuning knobs (variance/motion/density) that hard-gate rules, a named blacklist down to hex codes and fonts (Fraunces, #f5f1ea), a 62-item pre-delivery checklist, and a declared scope of six categories it refuses to handle. The author's empirical analysis finds 69.4% of the checklist items are mechanically verifiable, and demonstrates a ~150-line Python linter (24 rules) catching all 23 seeded violations with zero false positives, arguing for an upstream-generation-plus-downstream-CI architecture. The report also flags caveats: decorative academic citations, ~17K tokens of resident prompt cost, an unreleased v2, sponsorship by Kimi/Vercel, and unresolved Chinese-typography gaps. Original post in Chinese on zhichai.net.

Key points

This is a structured digest of a long-form Chinese investigation of github.com/Leonxlnx/taste-skill (data snapshot 2026-08-28; all figures from the repo source and GitHub API). Full repo: github.com/Leonxlnx/taste-skill · site: tasteskill.dev · MIT license.

What taste-skill is

  • A pure-markdown rules file (13 SKILL.md files; flagship 87,253 bytes / 1,206 lines) that tells coding agents what *not* to build: no generic AI-looking pages.
  • Repo health: 81,704 stars / 5,596 forks (6.8% fork ratio), 58 open issues, 7 contributors (author has 148 of 154 commits), created 2026-02-19, last push 2026-08-24. v2 is still experimental; the entire changelog sits under [Unreleased].
  • Anatomy of the v2 flagship

  • §0 Brief Inference: before writing code, the agent must declare in one sentence how it reads the brief (page type, audience, language, design system). One clarifying question max.
  • §1 Three knobs: DESIGN_VARIANCE / MOTION_INTENSITY / VISUAL_DENSITY on 1–10 scales, with real gating — e.g. variance > 4 bans centered hero sections; density > 7 bans generic card containers; motion > 3 makes prefers-reduced-motion mandatory. Claimed motion but static output is judged broken.
  • §9 Negative list: binary, named bans — zero em-dashes (U+2014), banned serif fonts (Fraunces, Instrument Serif), 16 specific beige-family hex codes (#f5f1ea, #b08947, …), fake names (John Doe, Acme, Nexus), banned copy tropes, and hard bans on window.addEventListener('scroll') and div-stacked fake screenshots. The spec itself notes graded wording ('use sparingly') gets ignored by models; only countable, binary criteria survive.
  • §14 Pre-flight checklist: 62 items, any failure means output is incomplete.
  • §13 Explicit scope surrender: dashboards, data tables, multi-step forms, code editors, native mobile, realtime collab — use Fluent / Carbon / Polaris / TanStack Table instead.
  • Versioning care: v1 kept installable, install names separated from folder names.
  • Core empirical finding: most of the checklist is CI-enforceable

    Decomposition of all 62 checklist items by mechanical verifiability:

    | Tier | Items | Share | Examples | |---|---|---|---| | A · regex-able | 29 | 46.8% | dash ban, hex/font blacklists, scroll listener, h-screen, eyebrow count | | B · semi-checkable | 14 | 22.6% | theme lock, contrast, CTA intent dedup, bento counts | | C · model/render needed | 15 | 24.2% | copy self-review, motion motivation, Core Web Vitals | | D · procedural | 4 | 6.5% | verdict declared, knobs justified |

    Conclusion: ~69.4% of checks can be pushed down into deterministic CI validation. The criticism that taste-skill 'has no enforcement' is wrong in degree, not direction.

    PoC linter

    The author ported 24 high-frequency rules to a ~150-line Python linter and ran it against a landing page seeded with AI-design tells. All 23 violations were caught, zero false positives, exit 1 — including banned dashes, Fraunces, #f5f1ea, scroll listeners, h-screen, duplicate CTA intents, and eyebrow-count over the published formula (count ≤ ceil(sectionCount/3)).

    Recommended architecture: taste-skill upstream to prune the statistical mode out of the candidate pool; a distilled linter downstream in CI to hard-fail. Upstream alone leaks; downstream alone means endless rework.

    Ecosystem and author (2026-08-28)

  • Competitor standings: nextlevelbuilder/ui-ux-pro-max-skill (122,278★), Leonxlnx/taste-skill (81,704★), Nutlope/hallmark (27,374★), alchaincyf/huashu-design (23,640★), hardikpandya/stop-slop (16,493★). Notably, the only regex-linter competitor (Laith0003/ux-skill) has just 65★ — the market buys convenience, not reliability.
  • Author Leon Lin (Munich) also ships unlazy (2,676★) and agentic-ai-prompt-research (2,520★). Sponsors include Kimi (Moonshot AI) and the Vercel OSS Program; no paywall, and the README explicitly disclaims any official token.
  • Chinese community activity: issue #87 (full bilingual translation of 15 skills), #69 (Chinese aesthetic preset 'cijian'), #33 (Chinese AI-writing bad-taste patterns need their own list).
  • Caveats

    1. The research directory is decoration: references.md cites five works with no links; several check out as misattributed or unverifiable (Compounding Error Avoidance has no trace; Winter Break Hypothesis is folklore). Ironically, the citations are written with em-dashes — the very character the skill bans. Verify any cited claims against original literature. 2. Token cost: the flagship is ~17K tokens, the full family ~50K. Issue #67 proposes progressive disclosure (12KB core + on-demand loading) but has sat unanswered for two months. Long instruction blocks are the first to be silently truncated. 3. v2 instability: everything is [Unreleased]; issue #86 notes §3's framework bindings (React/Next/Motion/GSAP) contradict the README's framework-agnostic claim. Pin v1 or wait for v2.0.0 stable.

    Adoption advice

  • Landing pages / portfolios: distill §0 + §1 + §9 + §14 (~12KB) for ~90% of the effect at ~30% of the cost.
  • Backends / forms / data tables: skip it (§13 surrenders these); use Fluent / Carbon / Polaris.
  • High-consistency needs: distill upstream + custom linter downstream.
  • Chinese-language output needs custom patches: Chinese-appropriate dash/quotation bans, line-break rules, a Chinese serif stack (Songti SC / Source Han Serif SC) and body line-height ≥ 1.8.

Bottom line

taste-skill's real value is a transferable engineering grammar: decompose undecidable 'taste' into decidable mechanical assertions, use binary bans to close the model's interpretive wiggle room, gate rules with knobs, and externalize quality as a checkable list. The hype has foam, the citations are garnish, but ~70% of the discipline demonstrably compiles into deterministic checks. Use the distilled version upstream, run a linter downstream, and don't worship or debunk it.

Tags

#taste-skill#ai-coding-agents#prompt-engineering#design-linting#open-source#ci-enforcement#web-design#github

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178634199