TasteSkill v2: AI Agents That Self-Critique Web Design to Kill 'AI Slop'

TasteSkill, an open-source web design skill for AI coding assistants with nearly 100k GitHub stars, has released version 2. Author Leon Lin describes it as a complete rewrite that equips coding agents with three new capabilities: an art direction system, real image assets sourced from open archives, and the ability to inspect their own output in a real browser.

Six-Step Workflow

Read brief & list unique client assets – The agent ingests the project brief and extracts distinctive client materials.

Randomly select three research directions from 20 – The agent picks one direction to pursue.

Bind client assets and chosen direction into a core concept – A single sentence fuses the two into the guiding creative idea.

Fetch real images and videos from open archives – Assets are pulled from sources like Wikimedia Commons (public domain or CC-licensed) and automatically attributed on the page.

Browser-based self-check at three viewport widths – The agent renders the page in a real Chrome/Edge/Chromium instance and scans for "AI slop" (template-like, plastic-feel artifacts). Detection of slop marks the run as failed.

Self-critique loop – The agent reviews its own screenshots, fixes the three most severe issues, and repeats up to three iterations.

Each full build takes roughly 30–45 minutes and consumes hundreds of thousands of tokens. The system’s aesthetic baseline derives from 247 hand-picked reference sites, making visual quality a deterministic rule rather than luck.

Key Insight: The Mirror for Aesthetic Feedback

The author identifies step 5 as the breakthrough. The core problem with AI-generated web pages isn’t lack of coding ability but the absence of aesthetic feedback – models finish code without knowing whether it looks good, waiting for human criticism. TasteSkill v2 effectively gives the agent a mirror: it sees its own rendered output, detects slop, and iteratively repairs it.

Counter-Intuitive Design Choice: Real Photos Over AI-Generated Images

v2 deliberately avoids AI-generated imagery, instead searching open archives such as Wikimedia Commons for authentic photographs. The rationale is that synthetic images are themselves a primary source of "AI slop"; even a perfectly crafted layout is ruined by a single plastic-looking picture.

Installation & Requirements

Install the default skill via: npx skills add Leonxlnx/taste-skill Requirements: Node 22+ and Chrome/Edge/Chromium. No API key is needed. All image assets come from public domain or CC-licensed open archives with automatic attribution.

To install only the default design-taste-frontend skill:

npx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend"

Additional Skills & Desktop App

The repository also contains scenario-specific skills: redesign-existing-projects – overhaul existing projects minimalist-ui and industrial-brutalist-ui – distinct visual styles imagegen series – produce design mockups without writing code

A free, open-source desktop application called TasteCode packages the entire workflow; its download link is provided in the repository.

Repository:

https://github.com/leonxlnx/taste-skill