Defeating AI Slop: A Workflow for Tasteful Web Design
Chase AIgo watch the original →
the gist
Avoid generic AI-generated web design by curating a personal inspiration library, utilizing specialized design MCPs, and adopting an iterative 'wide-net' build process instead of one-shot prompting.
Cultivating Design Taste
AI models default to 'slop'—generic, overused aesthetics like purple gradients and standard sans-serif fonts—because they regress to the mean of their training data. To break this, you must move from passive prompting to active curation. Build a personal 'Inspiration Library' using screenshots from platforms like Dribbble, Pinterest, and Twitter. Organize these by design type (e.g., dither mono, vast quiet, print tech) and use Claude Code to build a local app that catalogs these with associated keywords and design briefs. This library serves as your foundational reference, ensuring your output is grounded in specific, intentional aesthetics rather than generic defaults.
Integrating Specialized Design Tools
Standard LLM outputs lack the nuance of professional front-end polish. Enhance Claude Code by integrating specialized MCPs and CLI tools that handle the heavy lifting of design refinement. The 'Impeccable' CLI is recommended for its 23+ commands that address typography, spacing, and responsiveness, effectively identifying and stripping away 'slop' patterns. Similarly, the 'Taste' skill provides a higher-level framework for layout and motion. For visual assets, the 'Higgsfield' MCP is essential, as it bridges the gap between code and generative AI, allowing you to pull high-quality imagery and video directly into your development workflow without leaving the terminal.
The Iterative Build Sequence
Abandon the 'one-shot' prompting mindset. Instead, use a wide-net approach: generate five distinct variations of your site in different aesthetic families simultaneously. Once a direction is chosen, generate three sub-variations to refine the layout. This comparative process allows you to see what works in practice rather than guessing in a vacuum. When prompting, provide four key inputs: the aesthetic family, a reference image from your library, the project intent, and specific negative constraints (e.g., 'no purple gradients').
Refinement and Live Tweaking
Once the structure is set, focus on the hero section as the anchor for the site's mood. Iterate on imagery and color palettes by requesting variations from your chosen aesthetic. To finalize the polish, implement a 'tweaks bar' on your dev server—a UI overlay that allows you to adjust fonts, accent colors, and spacing in real-time. This mimics professional design tools and allows you to make granular aesthetic decisions without needing to re-prompt the model for every minor change.