Kill the 'AI Look': taste-skill's Aesthetic Guardrails for Coding Agents

It's Not a Component Library, It's a Prohibition Engine

taste-skill does not touch code, modify models, or bind to any component library. It is a collection of rules written into SKILL.md that change the agent's aesthetic behavior — zero intrusion, pluggable, framework-agnostic. Its core mechanisms are threefold:

Anti-slop prohibitions, not suggestions. The Anti-Default Discipline section explicitly bans LLM default tropes: AI purple gradients, centered hero on dark grid, three equal-width feature cards, pervasive glassmorphism, infinite micro-animations, Inter + slate-900. These are hard constraints that trigger failure, not gentle advice.

Three knobs parameterize aesthetics. Three 1–10 dials control output: DESIGN_VARIANCE (layout experimentation, 1 = absolute symmetry, 10 = artistic chaos), MOTION_INTENSITY (motion depth, 1 = static, 10 = cinematic physics), VISUAL_DENSITY (information density, 1 = gallery whitespace, 10 = cockpit data). A single skill spans minimal editorial to brutalist without swapping rule sets. The variable names are document-wide anchors; rules reference them repeatedly and forbid the agent from inventing aliases.

14-section decision chain, not scattered prompts. From §0 brief inference (audience, mood, existing brand assets) through §2 design-system mapping, §8 dark-mode protocol, §9 master prohibitions, §11 legacy project audit, to §14 hard pre-flight checks. Each section is context-triggered, forcing the agent to read requirements before acting instead of immediately applying a template.

These examples show what the prohibitions remove: AI's default symmetry, whitespace, and template feel vanish, replaced by artifacts unique to each client.

The Most Valuable Section: Honest Rules

§2's design-system mapping embodies rare restraint. Many prompts tell the model to create a Material style or add glassmorphism; the model then hand-rolls CSS to fake it. taste-skill reverses this:

If the brief reads Microsoft enterprise, install official Fluent UI; if Google flavor, use @material/web ; if Atlassian platform, use @atlaskit ; if UK public service, use govuk-frontend .

The rule explicitly forbids hand-writing CSS that official packages already provide, and forbids importing a system's tokens only to override 90% of them. Even more telling, it admits limitations: Apple's Liquid Glass has no official liquid-glass.css on the web; anyone mimicking it must label the result as an approximation. This honesty — refusing to bluff — builds trust in a field where every tool claims omnipotence.

A 30-Minute Build Flow

The default v2 skill operates on the author's research across 247 curated sites , executing a six-step loop:

Read the brief, list what only this client possesses.

Script draws 3 directions from 20 research angles; agent picks one.

Bind the client's own assets to that direction in a single sentence.

Assemble the page with real images and assets from open libraries (must cite sources, not AI-generated).

Inspect the page in a real browser at three viewport widths; detect slop and fail immediately .

Review own screenshots, identify the three worst issues, iterate up to three rounds.

Steps 5 and 6 are the most valuable: the agent becomes its own QA engineer. Generation is not the endpoint; the agent screenshots, tests in-browser, then critiques its own output — closer to a real designer's revision loop than any generation-quality prompt.

Installation and Configuration

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

Default skill requires Node 22+ and local Chrome/Edge/Chromium; no API keys needed (network only for reference images). Individual skills can be installed via --skill "design-taste-frontend". The repo contains 14 skills: default taste-skill (v2), redesign for visual audits of existing projects, image-to-code for reference-image-first workflow, gpt-taste emphasizing layout variance for GPT/Codex, style variants soft / minimalist / brutalist, and three pure image-generation skills imagegen-frontend-web / mobile / brandkit. A desktop app TasteCode (tastecode.dev) bundles the design agent. Claude Code users add via /plugin marketplace add Leonxlnx/taste-skill; Codex users via codex plugin add taste-skill@taste-skill.

Who Benefits

Developers using Cursor, Claude Code, Codex, or v0 to generate landing pages, product UIs, or portfolios but frustrated by generic AI aesthetics.

Content teams wanting to codify aesthetic standards into version-controllable, diffable, reviewable rule sets.

Practitioners needing a reference-image-first workflow; image-to-code fills that gap.

Conversely, the default skill explicitly targets landing pages, portfolios, and marketing pages — not data-heavy back-office systems, multi-step product flows, or interaction-dense dashboards.

Known Limitations

Cannot rescue weak models. taste-skill is a rule set; its ceiling is the underlying model's compliance ability. Weak models ignore rules; it raises the floor for capable models, not a silver bullet.

Single-build cost is high. Official docs state ~30–45 minutes and hundreds of thousands of tokens per build. It saves rework costs, not token spend.

Version instability. v2 marked experimental; wording may change before stabilization. Official advice: lock v1 for precise-behavior dependencies. Issues channel closed (0 open issues); feedback routed to Discord — no public bug tracking.

Commercial sponsorship visible. README sponsors include Sidrune AI, Kimi, Vercel. Author declares no tokens or crypto, but readers must distinguish rules from sponsor placements.

Author's Take

taste-skill succeeds because it targets the most painful, least quantifiable point in vibe coding: AI flavor . Everyone complains about lookalike AI interfaces; few translate that into an executable, reusable, pluggable constraint layer. Lin chose the right level — not a new component library, not a new model, just an aesthetic behavior constraint layer. 8 months and 95.7k stars validate that judgment.

Expectations should be tempered: it's a tug-of-war between rules and model execution; winning depends on the model underneath. What makes it worth long-term attention isn't the three knobs or 14 sections, but §2's honesty — use official packages when they exist, label approximations when they don't. In a domain where every tool claims unlimited capability, a tool that writes its boundaries into the rules is more trustworthy.

For newcomers: run it on a real landing page once. Watch the browser self-check and self-critique steps. See how much rework they eliminate. The bill will tell you if it pays off.

GitHub repository: https://opengraph.githubassets.com/1/Leonxlnx/taste-skill