design-taste-frontend (Leonxlnx/taste-skill)

评测

在 Codex CLI（SOTA 档）上跑 3 个任务、对照、一句话提示、装 skill 三组，共 18 次运行，由规则脚本打分（没有 LLM 评委），断网沙箱、每次运行独立临时目录。具体做法见下方英文评测原文。

三组对比

Codex CLI 实测，18 次运行，规则脚本打分，没有 LLM 评委。总分 16.2 / 30，结论 视情况而定。

下面是评测原文，英文。

Verdict: Situational (16.2 / 30). Source: Leonxlnx/taste-skill, skills/taste-skill, pinned at commit 12af471. Install: npx skills@latest add Leonxlnx/taste-skill --skill taste-skill. Licence: MIT (repo LICENSE).

What it is

The design-taste-frontend skill (Leonxlnx/taste-skill): an art-direction workflow for web pages that runs a local script to pick a design 'direction' from the brief (for example 'the page is the object'), writes a direction file and report into a .taste-skill/ folder, then builds the page against it, with rules against generic AI-looking layouts, a fixed set of sections and visual checks.

How we tested

Codex CLI 0.166 (gpt-6-astra, low reasoning) on a dedicated Windows test PC, a fresh temporary project and temporary HOME/TMPDIR/caches per run, skill installed (treatment) or absent (control) or absent with the skill's core requirement squeezed into one sentence appended to the prompt (one-liner group), 2 runs per group per task, network off, /tmp writes blocked. The prompts never mention the skill (except the wording noted below).

Same three tasks for all groups as for ui-ux-pro-max and frontend-design: a landing page for the Tidepool seaweed-snack brand, a municipal water-quality dashboard, and a redesign of the Granite Peak climbing-gym homepage keeping its content. The one-liner asks for one specific concept and palette, a distinctive type pairing and asymmetric layout, no purple gradients/default Inter/three-card grid/numbered markers/uppercase eyebrow labels, imagery drawn in CSS/SVG, and to check desktop and mobile widths. Pages were screenshot offline (desktop and mobile) and scored blind from shuffled contact sheets.

1. New landing page from a brief — Create a single-page landing site for Tidepool, a small-batch seaweed snack brand that sells by subscription (three flavours, $14/box, ships monthly). Write one self-contained index.html with inline CSS and minimal JS only if needed. No frameworks, no build st…
2. Data-dense dashboard UI — Build a dashboard page for a municipal water-quality monitoring team: current readings for 6 stations (pH, turbidity, chlorine, temperature), a 7-day trend for each, and an alerts panel. Single self-contained dashboard.html with inline CSS/JS and hard-coded re…
3. Reshape an existing bland page — old.html is the current homepage of Granite Peak Climbing Gym (Boulder, CO). Redesign it into new.html: keep all of its real content (pricing, classes, hours, address) but make the page look like it belongs to this specific gym. Self-contained HTML/CSS, no fra…

Scores come from a blind read by Grok Bot (shuffled ids, no condition shown) against a fixed rubric, 0–5 per run; an AI judge, so treat small gaps as noise.

Results

The one-liner given to the third group: "Art-direct against AI slop: choose one specific concept and palette for this subject, avoid purple gradients, default Inter, the generic three-card feature grid, numbered 01/02 markers and uppercase eyebrow labels, use a distinctive type pairing and an asymmetric composition, draw imagery in CSS/SVG, and check the page at desktop and mobile widths and fix what looks off."

Mean score: 3.96 without vs 3.58 with the skill (delta -0.38). Tokens per run: 115k vs 431k (3.74x); time 175s vs 198s. One-liner group: 4.29 (the skill's gain over the one-liner: -0.71; one-liner's own gain over no skill: +0.33); one-liner tokens 298k per run, so the skill costs 1.44x the one-liner.

What happened

• Blind mean: one-liner 4.29, control 3.96, skill 3.58 (n=6 each). Per task the skill gave 4.5 / 4.5 on the redesign (control 4.0, one-liner 4.1), 3.5 on the dashboard (control 3.75, one-liner 4.4) and 2.75 on the landing page (control 4.1, one-liner 4.4).
• The low landing-page mean comes from one broken page: a treatment run's page shows garbled text (the flavour selector writes markup as text, so '</span>' appears inside a button), a collapsed navigation and scattered shapes; the other treatment landing page was fine (4.0). The skill's own workflow reported 'browser verification was blocked', so it shipped without a visual check. Without that run the skill would average 4.0, level with control.
• The skill does change what pages look like: the slop tells it names are gone. Uppercase eyebrow labels per page: control 0-5, skill 0-1; numbered 01/02 markers: control 3-6 in 3 of 6 pages, skill 0 in all 6. The one-liner removed the same tells (0-1 labels, 0 markers) so the sentence achieves that part alone.
• Cost is very high: 431k tokens per run for the skill vs 298k one-liner vs 115k control (3.7x and 1.4x), with 25.7 shell commands per run (control 7.7). Each run also drops about 40 skill and direction files into the project (the skill's direction.json and report.md are written under .taste-skill/).
• Treatment dashboards and landing pages with .taste-skill/direction.json were distinctive in look (editorial paper-and-ink layouts) but not consistently better than the plain dark-green, serif-italic style that control and one-liner converge to.

Dimension scores

Limits of this test

• Blind visual judging by the reviewer from full-page contact sheets (desktop 600 px, mobile 300 px per page); a 0.25 difference is within judge noise. One broken page moves the skill's mean by about 0.3.
• n=2 per cell, three tasks; the one broken page could be bad luck rather than a systematic defect.
• Fonts are not loaded offline, so type pairing choices are not visible.
• The skill's own browser-screenshot step could not run in the sandbox, which probably hurts it more than the other groups.
• n = 2 per group per task: differences under about 1 point are noise.
• Blind AI judging: single judge, no inter-rater check; the original control/treatment runs and the later one-liner runs were judged in separate sessions (one-liner on 2026-11-02), so cross-session drift is possible.

Security

See security.md. Scan: 10 CRITICAL / 44 HIGH / 9 MEDIUM. Verdict: safe to run (no-network sandbox). A clean scan does not prove a skill safe.

Bottom line

It removes the generic-AI look and produced the best redesign here, but cost 3.7x tokens and in 1 of 6 runs shipped a visibly broken page without noticing; a one-sentence art-direction prompt scored higher at about a third of the tokens. Try it for art-directed marketing pages where a human checks the result, not as an unattended default.