XNews

Anti-Slop Design Skills and Methodologies for AI-Generated UIs

Anti-Slop Design Skills and Methodologies for AI-Generated UIs

A former Figma engineer published seven methodologies to combat 'AI design slop' in user interfaces. The core framework is a three-step design cycle: list constraints, conceive solutions, and revise constraints holistically—warning against 'whack-a-mole' local patching that AI prompts exacerbate. Other principles include aggressively subtracting unnecessary elements AI tends to add, iterating in dedicated design tools (Figma, Cursor Design Mode) rather than production code to avoid 'prototype gravity,' using component libraries and a /showcase page for isolated UI building, deploying previews with real data for validation, collecting competitor screenshots as context for AI, and cultivating taste through team critique ('threshing' method). A separate collection of 13 'Taste Skills' was also highlighted, including a 1,200-line v2 skill that infers design briefs, configures three 'knobs' (variance, motion, density), enforces hard rules against AI clichés (banning em-dashes, default AI purple, specific serif fonts), and mandates a 60-item pre-release checklist.

Key insights

AI Accelerates Execution but Human Judgment Remains Central to Design Quality

The methodologies argue that while AI can speed up design execution, the critical judgment required for good design still resides with humans. The three-step cycle of listing constraints, conceiving solutions, and holistically revising constraints is a human-driven process that AI cannot replicate. The 'threshing' method, where a team collectively critiques a design until it is satisfactory, underscores that taste is built through human reflection and iterative feedback, not generated by a model. AI's tendency to add redundant elements means a human must constantly enforce subtraction to avoid cluttered, incoherent interfaces.

Dedicated Design Environments Prevent 'Prototype Gravity' and Codebase Contamination

Iterating on AI-generated designs directly within a production codebase is identified as a critical mistake. The concept of 'prototype gravity' describes how an initial AI-built implementation in code creates inertia, discouraging exploration of alternative designs and forcing new work to awkwardly graft onto existing code. The recommended practice is to use dedicated tools like Figma, Cursor Design Mode, or HTML prototypes, and to have the AI generate 3-4 variants for each concept. A /showcase page is also proposed as an isolated environment for building and debugging UI components before they are integrated into the main application.

Systematic Rules and Checklists Can Eliminate Recognizable AI Design Patterns

A detailed catalog of 'AI signatures' has been compiled to help engineers identify and eliminate machine-generated aesthetic flaws. These include a blanket ban on em-dashes as the 'number one tell,' prohibition of default AI purple (the 'LILA rule'), restrictions on specific overused serif fonts like Fraunces and Instrument Serif, and a rule limiting eyebrow labels to one per three sections. A 60-item mechanical pre-release checklist is enforced, with any single failure preventing the design from being considered complete. These rules are derived from real-world testing failures and are codified into the 'Taste Skill' for AI coding agents.

Component Reusability and Preview Deployments Are Non-Negotiable for Quality

The separation of view and logic through reusable component libraries is emphasized as essential to prevent visual fragmentation, such as buttons being repeatedly re-implemented with slight variations. Furthermore, evaluating a design with real data via preview deployments is considered mandatory. Even if an AI agent perfectly implements a specification, hands-on testing with actual data often reveals flaws, making refinement and rework inevitable. For large features, the practice of splitting frontend and backend into independent pull requests—with backend verified by tests and frontend by preview links—is recommended.

Source posts