'The Missing Layer: Design Taste in AI Agents - Hassan El Mghari, Together AI'

AI app developers can significantly improve UI/UX quality by understanding and avoiding common AI-generated design patterns, using design frameworks like Hal...

By Sean Weldon

Design Taste in AI Agents: Systematic Approaches to Improving AI-Generated User Interfaces

Abstract

This research synthesis examines systematic approaches to improving user interface and user experience (UI/UX) quality in AI-generated applications. The analysis identifies recurring visual patterns termed "AI slop" that characterize low-quality AI-generated designs, which users can detect within 1-2 seconds of exposure. The paper presents the Hallmark framework, a structured methodology that codifies design anti-patterns into machine-readable constraints and provides curated reference libraries. Empirical findings demonstrate that strategic model selection, comprehensive reference provision, and iterative refinement processes substantially improve design outputs. Notably, the open-source model GLM 4 5.2 achieves near-parity with larger closed-source alternatives at five times lower cost. The research establishes that 10-20% additional effort focused on UI/UX provides significant competitive advantages, with design quality serving as a primary driver of user adoption in AI-generated applications.

1. Introduction

The democratization of application development through AI-assisted tools has fundamentally altered the accessibility of software creation. Natural language interfaces now enable developers to generate functional user interfaces through conversational prompts, substantially lowering technical barriers to entry. However, this accessibility has produced an observable phenomenon: a recognizable category of design output characterized by predictable visual patterns and suboptimal user experience decisions. These patterns, collectively termed "AI slop," represent systematic design choices that signal automated generation to end users within 1-2 seconds of interface exposure.

The implications of this phenomenon extend beyond superficial aesthetic considerations. Empirical evidence from production applications demonstrates that design and user experience constitute primary drivers of application adoption, with applications achieving millions of users succeeding primarily through intentional design focus rather than technical implementation alone. This observation creates a fundamental paradox in AI-assisted development: while generative tools lower barriers to functional implementation, they simultaneously risk commoditizing applications through homogeneous design patterns that undermine differentiation and user engagement.

This synthesis examines methodologies for transcending default AI-generated design quality through four primary mechanisms: systematic identification and avoidance of AI slop patterns, implementation of design frameworks that codify best practices, strategic model selection optimized for specific development phases, and iterative refinement processes incorporating high-quality design references. The analysis establishes that relatively modest additional effort - approximately 10-20% of development time focused on UI/UX refinement - provides substantial competitive advantages in an increasingly crowded landscape of AI-generated applications.

2. Background and Related Work

2.1 The AI Slop Phenomenon

AI slop refers to identifiable visual and interaction design patterns that characterize low-quality AI-generated interfaces. These patterns manifest across multiple design dimensions, including color schemes (disproportionate frequency of purple gradient backgrounds), typography (italicized headers and all-caps pill buttons with increased letter spacing), visual elements (gradient-based logos and excessive emoji usage), layout characteristics (systematic spacing and padding inconsistencies), and call-to-action patterns (formulaic phrases such as "scroll to explore").

Critically, users demonstrate rapid pattern recognition capabilities, identifying AI-generated websites within 1-2 seconds despite often lacking explicit vocabulary to articulate the specific design elements triggering recognition. This implicit detection capability suggests that AI slop patterns violate established design conventions or user expectations in systematic, machine-learnable ways that human perception processes efficiently detect.

2.2 Framework Approaches to Design Quality

The Hallmark framework represents a structured approach to addressing AI slop through two primary architectural components. First, it implements slop gates - explicit instructions that enumerate anti-patterns for language models to avoid during generation (e.g., "no purple gradients," "no italics in titles"). Second, it provides a curated theme library that serves as contextual reference material for models during generation processes. Launched approximately 1.5 months prior to the analysis period, the framework had attracted over 10,000 users and demonstrated capability to produce landing pages through single-shot prompts with minimal AI-generated visual tells. This framework serves as a reference implementation for developers lacking established personal design preference systems.

3. Core Analysis

3.1 Model Selection and Task-Specific Optimization

The analysis reveals significant performance variations across language models when applied to design tasks, with implications for both output quality and computational efficiency. The open-source model GLM 4 5.2, released approximately two weeks prior to the analysis period, represents the first open-source model demonstrating genuine effectiveness at design tasks according to empirical evaluation.

Comparative assessment between GLM 4 5.2 and the closed-source baseline Opus 4.8 reveals near-indistinguishable output quality despite the former's significantly smaller parameter count. Critically, GLM 4 5.2 achieves this parity at five times lower computational cost and substantially faster inference speed. This performance profile inverts conventional assumptions about model scaling, suggesting that task-specific optimization and training procedures may matter more than raw parameter count for design-focused applications.

Furthermore, the analysis identifies task-specific model selection as a strategic consideration. Smaller, faster models prove preferable for iteration cycles despite being both cheaper and faster than larger alternatives. The closed-source model Cursor Composer 2.5, specifically post-trained for iteration tasks, demonstrates superior performance on refinement operations due to speed and responsiveness characteristics. This finding suggests an optimal workflow architecture: larger models for initial comprehensive generation, followed by smaller specialized models for rapid iterative refinement.

3.2 Reference-Based Generation and Inspiration Strategy

The provision of visual references and design examples to language models emerges as the single most impactful intervention for improving output quality. AI models provided with screenshots and reference materials produce substantially better outputs than those receiving text prompts alone, regardless of prompt sophistication or length.

The analysis recommends maintaining a curated inspiration vault - a personal collection of websites and applications exhibiting strong design characteristics for future reference provision. Optimal results emerge from explicit specification of design inspiration combinations, such as "mix of Duolingo and this app and this app," accompanied by multiple screenshots providing visual context. This approach enables models to interpolate between reference examples rather than defaulting to training distribution patterns that produce AI slop characteristics.

Empirical observation indicates that AI models "perform very very well when given really good inspiration as context," suggesting that model capabilities for design tasks may be more constrained by input quality than by architectural limitations. This finding has significant implications for workflow design: effort invested in curating high-quality reference libraries may yield higher returns than investments in prompt engineering or model selection.

3.3 Prompt Engineering and Task Decomposition

Systematic analysis of prompt characteristics reveals that longer, more specific prompts (2-3 paragraphs) consistently outperform shorter alternatives across design tasks. The analysis recommends voice note recording for 1-3 minutes as an efficient mechanism for generating comprehensive prompts that include context about user types, feature requirements, and design objectives.

For multi-feature applications, task decomposition emerges as a critical strategy. Breaking applications into separate prompts addressing 1-2 features per prompt produces superior results compared to monolithic prompts attempting to specify all features simultaneously. The recommended workflow architecture involves an initial comprehensive prompt with inspiration screenshots for core application generation, followed by sequential feature prompts queued for specific functionality additions.

This decomposition strategy addresses a fundamental limitation in language model generation: the difficulty of maintaining consistency across multiple constraints simultaneously. By reducing the constraint space per generation episode, sequential prompting enables models to allocate representational capacity more effectively to each feature subset.

3.4 Iterative Refinement Processes

The analysis establishes that AI-generated output should be conceptualized as base material requiring systematic refinement rather than final product. Single-shot generation produces functionally impressive but visually rough results; 1-2 follow-up prompts providing specific feedback dramatically improve logo quality, loading states, animations, and spacing characteristics.

Iteration cycles reveal systematic patterns - such as logos consistently requiring regeneration - that should be codified into preference files for future generation episodes. The analysis recommends maintaining design preferences in markdown files (e.g., agent.mmd) to build personal design systems that accumulate over time. This preference accumulation creates a feedback loop: observed deficiencies in generated outputs inform preference file updates, which constrain future generations to avoid previously identified issues.

The strategic implication is clear: keeping applications simple and iterating methodically produces superior results compared to attempting comprehensive one-shot generation. This workflow architecture aligns with human design processes, which similarly rely on iterative refinement rather than complete specification.

4. Technical Insights

4.1 Computational Efficiency and Cost Trade-offs

The performance characteristics of GLM 4 5.2 relative to Opus 4.8 challenge conventional scaling assumptions in language model deployment. Achieving near-parity output quality at five times lower cost and faster inference speed suggests that task-specific model selection can yield order-of-magnitude efficiency improvements. For production deployments involving thousands of generation requests, this cost differential translates to substantial operational savings without quality compromise.

4.2 Framework Architecture and Anti-Pattern Codification

The Hallmark framework's implementation of slop gates demonstrates a generalizable approach to constraining language model outputs: explicit enumeration of anti-patterns in machine-readable format. The framework codifies 20-30+ identifiable AI slop patterns, providing a reference implementation for developers seeking to build custom design systems. This architectural pattern - negative constraints combined with positive reference examples - appears more effective than positive-only specification approaches.

4.3 Sequential Generation Workflows

Multi-feature application generation benefits substantially from sequential prompting architecture (1-2 features per prompt) rather than monolithic specification. This finding suggests that language model working memory or attention mechanisms may struggle with highly multi-constrained generation tasks, and that workflow decomposition can circumvent these architectural limitations without model modification.

4.4 Implementation Considerations

The analysis identifies several practical implementation patterns. First, preference files should accumulate iteratively based on observed generation deficiencies rather than being specified comprehensively upfront. Second, reference provision should be visual (screenshots) rather than purely textual (descriptions), suggesting that vision-language model capabilities are critical for design tasks. Third, model selection should match task phase: comprehensive models for initial generation, specialized lightweight models for iteration.

5. Discussion

The findings presented in this analysis converge on a central insight: AI-generated design quality is constrained less by model capabilities than by workflow architecture and input quality. The dramatic improvement achieved through reference provision, iterative refinement, and strategic model selection suggests that current language models possess latent design capabilities that default workflows fail to elicit.

This observation has broader implications for AI-assisted creative work. The patterns identified here - task decomposition, reference-based generation, iterative refinement, and preference accumulation - likely generalize beyond UI/UX design to other creative domains. The fundamental challenge in AI-assisted creation may not be model capability but rather the development of workflows that effectively guide models away from training distribution defaults toward user-specific quality criteria.

Several areas warrant further investigation. First, the specific mechanisms by which visual references improve generation quality remain underspecified; understanding whether models perform interpolation, style transfer, or other operations would inform optimal reference selection strategies. Second, the scalability of preference file approaches requires examination; as preference files grow, do they eventually degrade generation quality through over-constraint? Third, the transferability of curated preferences across models and tasks merits systematic study.

The competitive implications of these findings are substantial. In an environment where functional AI-generated applications are increasingly commoditized, the 10-20% additional effort focused on design quality provides differentiation that users detect within seconds of exposure. This suggests an emerging competitive dynamic where workflow sophistication - the ability to effectively guide AI tools - becomes more valuable than access to the tools themselves.

6. Conclusion

This research synthesis establishes that systematic approaches to AI-assisted UI/UX design can substantially improve output quality relative to default generation workflows. The Hallmark framework demonstrates that codifying design anti-patterns into machine-readable constraints, combined with curated reference libraries, enables single-shot generation with minimal AI slop characteristics. Strategic model selection - specifically the deployment of GLM 4 5.2 for design tasks - achieves near-parity with larger models at five times lower cost, challenging conventional scaling assumptions.

The practical implications are clear: developers should maintain curated inspiration vaults, provide visual references with all generation requests, decompose complex applications into sequential prompts, and treat AI-generated output as base material requiring iterative refinement. Preference files should accumulate observed deficiencies over time, creating feedback loops that progressively improve generation quality.

The central contribution of this analysis is demonstrating that relatively modest additional effort - 10-20% of development time focused on UI/UX refinement - provides substantial competitive advantages in user adoption and engagement. As AI-assisted development tools proliferate, the ability to transcend default AI slop patterns through systematic workflow design emerges as a critical differentiator. Future work should examine the generalizability of these patterns to other creative domains and investigate the specific mechanisms by which visual references improve language model generation quality.

Sources

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub