When you build a web page or interface with AI, does it often look unmistakably "AI-made," ugly and generic? This problem has a name, "AI slop," meaning soulless, cookie-cutter output. As more people use AI to quickly build landing pages and product interfaces, this "AI look" is only getting more obvious. A GitHub open-source tool with over 64,000 stars, "Taste Skill," targets exactly this: it gives AI "taste" so the interfaces it builds have real layout, typography, motion and spacing, instead of just applying templates. This article unpacks in detail what it is, the concrete symptoms of AI slop, how it works, its style variants, who it's for, and what to watch out for.
What is Taste Skill?
Taste Skill is a set of "anti-slop" frontend Agent Skills for raising the quality of AI-built interfaces. It exists as a "SKILL.md," a portable instruction file the AI can load automatically and follow. It is not tied to any one framework; React, Vue and Svelte all work, because it targets "design intent" rather than a specific API syntax. The whole project is under the MIT open-source licence, free to use, with an active community and continuous updates; the main skill now ships a substantially rewritten second generation.
What are the concrete symptoms of AI slop?
"AI slop" is not just abstract criticism but a set of traceable flaws: layouts always centred and symmetrical with no visual hierarchy; spacing loose and uniform with no rhythm; type contrast too weak so headings and body blur together; safe but dull colour; missing motion or only basic hover. Together these make it instantly recognisable as "AI-made." The root cause is that the model tends to output the "statistically safest" choice, and safe usually means bland. Taste Skill uses rules to push the AI out of that safe zone.
How it works
At its core, Taste Skill is a research-informed set of "anti-repetition, anti-slop" design rules written into a SKILL.md. Once an AI agent loads it, it follows those rules while building interfaces: actively creating layout variation, strengthening type hierarchy, arranging whitespace and rhythm, and adding fitting motion. It even builds in "hard rules," such as banning certain instantly-tacky patterns, and requires inferring the brand's tone before designing. In effect, it turns a senior designer's judgment into instructions the AI can follow.
How to install and use it
Installation is simple: one npx skills add command installs it into mainstream coding agents like Claude Code, Cursor and Codex; you can also install just one specific skill. You can also copy the SKILL.md straight into your project, or paste it into a ChatGPT / Codex conversation. Once installed, the interfaces the AI builds have noticeably more taste. If you are improving an existing project, there is a dedicated "redesign" skill that audits your current UI first, then fixes layout, spacing, hierarchy and styling item by item, without starting over.
Three adjustable "dials"
The main skill offers three dials from 1 to 10 for precise style control. DESIGN_VARIANCE (layout experimentation): low is centred and clean; high is asymmetric and modern. MOTION_INTENSITY (animation depth): low is mostly hover; high is scroll-triggered and magnetic interactions. VISUAL_DENSITY (information per viewport): low is spacious whitespace; high is dense dashboards. Turn the three dials and the same skill moves from conservative enterprise style to bold avant-garde, fitting different brands and scenarios.
Multiple style variants
Beyond the default, it offers several specialised variants so you can pick by an already-decided direction: minimalist (the restrained editorial feel of Notion or Linear), high-end soft (soft contrast, generous whitespace, spring motion, an "expensive" feel), industrial brutalist (Swiss type, hard contrast, experimental layout), and a stricter variant tuned for GPT / Codex. There is also a "full output" skill that cures the old problem of the AI shipping half-finished work full of placeholder comments.
Not just code, it also outputs design images
Besides implementation skills that output code, Taste Skill has a set of "image-generation" skills: it can produce website reference comps, mobile interface flows, even brand kits (logo directions, palettes, type, identity applications). These skills output images only, not code. A common flow is to generate references and analyse them first, then hand them to a coding agent to implement, linking "design" and "development" into one pipeline. This "see it first, then build it" approach is especially useful for those who can't draw but want to lock a direction fast.
Who it's for
Anyone building websites, landing pages or product interfaces with AI can use it: founders building an MVP, marketing teams shipping campaign pages, developers who want speed without ugliness, freelancers wanting to deliver polished design with fewer hands. It is especially useful for non-designers, because it pre-encodes the judgment of "good design" into rules, so you can ship a decent interface even without design skill. For teams that already have designers, it works as a "first-draft accelerator," getting the AI to 70-80% before a human polishes it.
Limits and cautions
A few things to note: first, it raises "design taste," not product logic or information architecture, so the direction is still yours to set; second, the main skill is currently an experimental second generation and its behaviour may still shift, so if you depend on the exact behaviour of the old version you can pin the first generation; third, the output is a starting point, not a final draft, so important projects should still be human-reviewed and fine-tuned. Treat it as a "tasteful first-draft engine" rather than "one-click perfection" and it works best. Want to know how to pick the right variant and the actual rollout steps? Visit ai.ud.hk to explore UD's AI Staff solutions and see how to make your AI output more professional.
懂AI，更懂你｜UD相伴，AI不冷
