DEV Community

啦啦啦种太阳
啦啦啦种太阳

Posted on Originally published at skills.pub

ui-ux-pro-max Ships 2,600 Rows of Local CSV, Not a Design Model

The ui-ux-pro-max skill answers design questions by running a Python script against 14 CSV files stored inside the skill directory — no API, no model call, no network. The SKILL.md is 2,371 words and spends most of them on one thing: how to write a query for that script. On skills.pub it comes from a repository with 151,799 stars, and its own description enumerates the catalogue: 80 searchable styles, 194 product palettes, 74 font pairings, 130 UX guidelines, 105 icons, 17 GSAP presets, 25 chart types, 22 stacks.

How does the skill actually execute?

One shell command, repeated. The SKILL.md states the invocation and warns about the path:

python "${CLAUDE_PLUGIN_ROOT}/.claude/skills/ui-ux-pro-max/scripts/search.py" "<query>" --domain <domain>
Enter fullscreen mode Exit fullscreen mode

"Always invoke it by its full path — do not assume a particular working directory." The script takes a domain flag and a stack flag, and returns matching rows from the CSVs. Everything else in the skill is guidance on forming that query.

Flag Purpose Example
--domain Which catalogue to search style, color, product, typography, chart, ux, icons, gsap
--stack Restrict to a framework One of 22 stacks
--design-system Generate a full system for a new project Default on new work
--tokens Emit design tokens dtcg or css
--rtl Right-to-left layouts —
--variance, --motion, --density Tune the generated output, 1–10 —

What is in the catalogues?

Counted from the repository's catalog-summary.json and the CSV headers on 2026-11-20:

File Rows Columns
colors.csv 194 19
products.csv 194 10
ui-reasoning.csv 194 12
ux-guidelines.csv 130 10
icons.csv 105 11
styles.csv 89 30
typography.csv 74 11
react-performance.csv 44 11
landing.csv 34 10
app-interface.csv 32 11
charts.csv 25 15
motion.csv (GSAP) 17 12
google-fonts.csv 1,934 15
stacks/*.csv 1,278 total 13 each

The ux-guidelines.csv columns are the most useful to inspect, because they show what a "rule" looks like as data: No, Category, Issue, Platform, Description, Do, Don't, Code Example Good, Code Example Bad, Severity. Each row carries a good and a bad code example, which is what lets the agent quote something concrete instead of paraphrasing advice.

styles.csv has 89 rows but only 51 are active; 29 are marked supplemental and 9 deprecated. That distinction matters if you are comparing the skill's marketing copy against its behaviour — the "80 searchable styles" in the description is the rounded total, not the active count.

Where does the data come from, and is it fresh?

Local files, verified offline. The Python modules import only the standard library — csv, difflib, re, pathlib, math, json, os, tempfile — and contain no urlopen or requests. The one place a URL appears is fonts.googleapis.com preconnect links inside generated HTML previews.

Freshness is declared rather than enforced. data-provenance.json records verifiedAt: 2026-11-04 with a policy of manual-verifiedDays: 365 and needs-reviewDays: 90. The cited sources are W3C WCAG, OSMF, Mapbox, MapLibre and similar — citations for where the rules came from, not live fetches.

Question Answer
Does it need network access? No, not at runtime.
Does it call a model? No. The agent does the reasoning; the script does retrieval.
Can the data go stale? Yes. It is a snapshot with a declared review interval.
Can you edit it? Yes — the CSVs are ordinary files in the skill directory.

What does the SKILL.md forbid?

Four rules that read like they were written after real failures:

  • "Never assume a stack" — "a hardcoded default silently misroutes every recommendation."
  • "Do not persist unverified output."
  • "Never present a 0-result search as if it returned data." With a fallback: retry once with a narrower rewrite or an explicit domain.
  • "Never use --force without explicit user authorization."

The 0-result rule is the one worth copying into other retrieval-style skills. An empty search that gets summarised as if it were an answer is the most expensive failure mode in this category, because it looks like success.

How is it installed?

Two routes, and they differ in what you get:

Route Command Caveat
Claude Code marketplace /plugin marketplace add nextlevelbuilder/ui-ux-pro-max-skill then /plugin install ui-ux-pro-max@ui-ux-pro-max-skill Also registers the six sibling skills
CLI npm install -g ui-ux-pro-max-cli then uipro init --ai claude The README warns older uipro-cli releases are stale

The repository contains seven distinct skills — ui-ux-pro-max, design, brand, design-system, slides, ui-styling, banner-design — each duplicated into .claude/skills/ and cli/assets/skills/, except the main one, which has no CLI twin.

Is it better than just asking the model?

For consistency, yes, and for a specific reason: the same query returns the same row. Ask a model for "a palette for a fintech dashboard" twice and you get two plausible answers. Search a 194-row table and you get the row that was reviewed once.

The cost is coverage. Anything not in those CSVs does not exist as far as the skill is concerned, and the agent is told not to invent it. If your product type is not in products.csv, the skill's honest answer is a 0-result retry, not a design.

Three things to check before adopting it:

  1. Run the search script once by hand with a product type you know. Confirm you get a row, not an empty array.
  2. Ask for something absent from the catalogues and confirm the agent reports the miss instead of making something up.
  3. Check data-provenance.json's verifiedAt against today. If it is more than 90 days old, the catalogs are past their own review interval.

Catalogue sizes, column names, import lists and the SKILL.md directives were read from the repository on 2026-11-20: SKILL.md and catalog-summary.json.

Top comments (0)