anti-slop-web-design

Anchor the builder to reference images and URLs you actually like, then generate 5 distinct aesthetic variants side by side instead of one-shotting. Narrow to a direction, create 3 layout variants, and iterate with a tweaks bar rather than requesting full re-renders.

Pass an aesthetic (design family), reference images or real site URLs, intent (audience and desired action), and guardrails (always-do and never-do rules). Keep it loose rather than writing a massive prescriptive design document.

The workflow references Impeccable for slop detection across 7 design axes, Taste Skill for layout and typography, Higgs Field MCP for hero image and video generation, and 21st.dev for component-level inspiration.

Hard bans include neon or glow effects, purple-to-blue gradient heroes, Inter as default font, shadow-lg everywhere, centered-everything layouts, emoji-as-icons, generic SaaS feature-tile grids, fake dashboard numbers, and stock-photo heroes.

Out-of-the-box generation regresses to the statistical mean of training data, so output stays average no matter how good the model gets. The fix is injecting your own taste through references and iterating across multiple variants.

Avoid the trap of stacking narrow, prescriptive skills, since each produces only one kind of output. Prefer flexible tools like Impeccable, Taste Skill, and Higgs Field whose results depend on your prompting rather than fixed templates.