Deep Research: ui-ux-pro-max-skill, the Project Ranked Ahead of taste-skill

Forum topic · 二一 · 2026-08-23

Summary

A detailed technical review of nextlevelbuilder/ui-ux-pro-max-skill, the most-starred (120K) anti-slop frontend skill repository, which outranks the more-discussed taste-skill (79K). The author audits the entire repository tree and verifies every headline number: a Python BM25-plus-regex retrieval engine, 37 CSV datasets, 54 TTF fonts, 11 test files, and 22 technology stacks spanning web, desktop (JavaFX, Avalonia, WPF, UWP), and native mobile (SwiftUI, Jetpack Compose). Unlike taste-skill's 87KB always-in-context prompt, ui-ux-pro-max keeps its SKILL.md to only 16KB and loads knowledge on demand. Data freshness is enforced via tests that fail CI when stack versions or links go stale. The article also analyzes its freemium structure (brand/slides/banner directories are thin Premium teasers), compares it feature-by-feature with taste-skill, and concludes the two are complementary: one is a design-intelligence engine, the other an aesthetic-discipline rulebook. It ends with per-scenario recommendations and needed patches for Chinese typography.

Key points

Ranking correction: The top repo in the anti-slop frontend space is not taste-skill (79K stars) but nextlevelbuilder/ui-ux-pro-max-skill (120K stars, created 2025-11-30) — a mistake widely repeated in Chinese-language coverage.
Different species: taste-skill is an 87KB aesthetic-discipline prompt fully resident in context. ui-ux-pro-max is a running product: a Python retrieval engine, 37 CSV datasets, 54 real font binaries (5.2MB), 11 test files, and an npm CLI installer — 662 files, 20.5MB across 7 skill directories (ui-styling 5.6MB, ui-ux-pro-max 3.4MB flagship, design, design-system, brand, slides, banner-design).
Mechanism: core.py performs hybrid BM25 + regex retrieval over CSVs, returning deterministic ranked rows with explicit fallback on zero hits — no model judgment in the retrieval step. Data splits into 35 horizontal CSVs (styles, colors, typography, ux-guidelines, charts, motion…) and 22 stack-specific CSVs (react, nextjs, svelte, swiftui, flutter, jetpack-compose, javafx, wpf, etc.).
Verified claims: All ten advertised numbers check out exactly row-by-row — 79 searchable styles (50 active), 192 product palettes/types/reasoning profiles, 74 font pairings, 119 UX guidelines, 105 icons, 17 GSAP presets, 25 chart types, 22 stacks. (Contrast: taste-skill's five academic citations trace to mismatched or unverifiable sources.)
Freshness tests: test_web_stack_freshness.py hardcodes a current-version table (react 19.2.x, nextjs 16.2, vue 3.5.x, svelte 5, astro 7.1.6, angular 22.x, tailwind 4.3, shadcn cli 4, nuxtjs 4.5, nuxt-ui 4.10) and validate_data.py's STACK_OFFICIAL_HOSTS requires official-domain links — knowledge rot becomes a failing CI check.
Context strategy — the core insight: flagship SKILL.md is only 15,969 bytes vs taste-skill's 87,253 bytes (~21.8K tokens) fully resident. Rules live in quick-reference.md (24.5KB) and pro-rules.md (10.9KB), loaded on demand. taste-skill's own issue #67 proposes exactly this progressive-disclosure refactor (~12KB core) — still open with zero comments.
Three knobs, actually wired: --variance, --motion, --density (1–10) feed the Python search — high variance retrieves Brutalism/Bento Grids, low retrieves Minimalism — reproducible, unlike taste-skill's advisory-only knobs.
Design system persistence: --design-system aggregates product type, style, palette, typography, motion; --persist writes a MASTER.md plus per-page overrides for reuse.
Business structure: Basic (this repo) is a deliberate free tier feeding a Premium product (uupm.cc pricing, PayPal). The brand (87KB), slides (19KB), and banner-design (13KB) directories are thin shells with no CSVs — matching Premium's selling points — while core UI/UX gets 9MB of data. Openly documented, but reviewers should know where the paywall sits.
What it does NOT do: no anti-slop negative-fingerprint bans (no "anti-slop" in README, no mention of taste-skill, em-dashes freely used). It answers "what style/palette/font should I use," not "why does my page look AI-made."

Comparison with taste-skill

| Dimension | ui-ux-pro-max (120K★) | taste-skill (79K★) | |---|---|---| | Form | Python engine + 37 CSVs + fonts + CLI, 20.5MB | 13 markdown files, 87KB flagship | | Context cost | 16KB SKILL.md, on-demand retrieval | 87KB / ~21.8K tokens always resident | | Determinism | High — ranked rows, fallback markers | None — same prompt varies per run | | Claim accuracy | 10/10 verified | Citations partially fabricated | | Data freshness | CI-enforced version tables | None | | Tests | 11 files incl. data contracts | None | | Design system | Generates + persists MASTER.md | Single-page only | | Frameworks | 22 stacks incl. desktop/native mobile | Hardbound to React/Next (issue #86) | | Anti-homogenization | None | Core capability, bans down to hex/characters | | Business model | Freemium funnel | Fully free, sponsor wall | | Dependencies | Python, 5.2MB fonts | Zero, plain text |

Recommendations by scenario

Single-page marketing/portfolio: taste-skill — the negative bans directly target "looks AI-made," zero dependencies.
Multi-page products, dashboards, admin panels: ui-ux-pro-max — taste-skill's §13 explicitly excludes dashboards/data tables; here the density knob and app-interface/charts datasets fit.
Cross-platform: only ui-ux-pro-max covers desktop and native mobile.
Project-wide design system: ui-ux-pro-max's --design-system --persist is unique.
CI quality gate: neither works. ui-ux-pro-max's determinism is at retrieval, not output validation. You'd need your own regex linter in PR checks — taste-skill's §14 62-item preflight list is effectively a ready-made linter spec.
Best combo: distill taste-skill's §9 negative list into ~30 lines of project conventions; delegate retrieval to ui-ux-pro-max. Running both heavy skills in context is too expensive.

Chinese-locale patches (needed for both)

Font stacks: neither covers Chinese. Suggest Source Han Serif SC / Songti SC for display headlines (rare in training data, hence effective de-AI-flavor), PingFang SC for body.
Typography: avoid leading/trailing punctuation at line breaks, ~0.15em spacing between CJK and Latin, text-spacing-trim.
Metrics: raise body line-height to 1.8+, reduce display sizes ~10% vs Latin settings.
Translate the bans: full-width quotes, no English ellipsis, no tilde as range marker — the em-dash ban matters little in Chinese.

Conclusion

ui-ux-pro-max is the most engineering-mature project in this space: a working retrieval engine, fully verified numbers, freshness enforced as failing tests, 16KB context footprint, 22 stacks, persistent design systems. But it does not cure homogenization — it lacks negative-fingerprint bans and doesn't suppress the model's default aesthetic. One-line verdict: ui-ux-pro-max is a design-intelligence engine; taste-skill is an aesthetic-discipline codebook. One tells you what to use; the other tells you what not to use. Their gaps are complementary, and neither replaces the other. The real vacancy is a third thing: a deterministic validator that can exit 1 in CI — across 200K combined stars, no project here can block a single commit.

--- *References: github.com/nextlevelbuilder/ui-ux-pro-max-skill ｜ uupm.cc ｜ comparison repo github.com/Leonxlnx/taste-skill ｜ data snapshot 2026-08-24*