Hallmark Fights 'AI Slop' with Design Gates and Thematic Libraries

Hassan El Mghari, the head of the developer experience team at Together AI, opened the session by highlighting the critical role of UI design in AI applications. He emphasized the objective of preventing AI agents from producing visually unappealing user interfaces. El Mghari mentioned his extensive experience, having built approximately 10 applications annually for the past five years, underscoring his practical insights into app development and design.

He expressed his enthusiasm to discuss strategies for improving the aesthetic quality of AI-generated interfaces. El Mghari's focus is on equipping developers with the knowledge and tools to create AI applications that are not only functional but also visually appealing, moving away from what he terms 'ugly UIs'.

Together AI operates as an AI-native cloud platform, offering a suite of services designed to support AI development and deployment. The platform enables users to run open-source AI models, providing a flexible environment for various applications. Additionally, Together AI facilitates fine-tuning models using proprietary data, allowing for highly customized AI solutions, and features a GPU cluster product specifically for high-performance hardware like H100s and B200s.

Hassan El Mghari showcased a portfolio of AI applications that prioritize simplicity and strong design, including a logo creator, 'Make Comics,' a subtitle generator, and an AI cloud agent. These applications are typically straightforward, one-page interfaces where El Mghari focuses on design and subtle animations to enhance user experience.

He attributes the success of these projects not to their inherent complexity, but rather to their emphasis on user experience (UX) and aesthetic design. This approach underscores his belief that impactful AI applications can be achieved through thoughtful design choices, even with minimal functional complexity.

El Mghari addressed the prevalent issue of 'AI slop,' where many AI applications share identical and uninspired design elements, making them instantly recognizable as AI-generated. He noted that within seconds, users can often discern if an interface was created by AI due to a set of common cliches. These include ubiquitous purple gradients, an overabundance of emojis, 'scroll to explore' prompts that users rarely request, all-caps pill-shaped buttons with wide letter spacing, and italicized headers.

This lack of distinctiveness stems from AI models frequently drawing from similar training data, leading to a homogenization of design. El Mghari asserted that investing even a small amount of extra effort, roughly 10% to 20% more, into refining the UI can provide a substantial competitive advantage.

He emphasized that good UI design is not merely an aesthetic choice but a crucial differentiator in the crowded AI application market. By avoiding these 'tells' and focusing on unique, thoughtful design, developers can make their AI products stand out and appear more polished and professionally crafted.

To counter the widespread issue of generic AI aesthetics, Hassan El Mghari introduced 'Hallmark,' a tool specifically designed to standardize and enhance design quality in AI outputs. Hallmark addresses the problem by codifying common 'AI slop patterns,' essentially creating design gates that instruct AI models to avoid these undesirable choices. This mechanism helps steer AI-generated interfaces away from predictable and often amateurish designs.

Crucially, Hallmark also provides AI models with an extensive library of diverse themes, offering rich inspiration beyond generic templates. This dual approach of blocking bad patterns and providing strong thematic guidance aims to produce more distinctive and aesthetically pleasing AI applications, fundamentally shifting AI outputs away from the homogeneity seen in much of today's AI-generated content.

Hassan El Mghari announced the successful launch of Hallmark, a design tool specifically engineered for AI agents, which has quickly gained significant traction. Since its release approximately six weeks ago, the platform has already garnered over 10,000 users. Hallmark's core mission is to eliminate 'AI slop' by implementing strict design gates and offering comprehensive thematic libraries to guide AI models toward better aesthetic choices.

This rapid adoption underscores the urgent market need for solutions that address the common problem of uninspired and generic AI-generated user interfaces. By providing structured design constraints and a wealth of creative options, Hallmark empowers AI agents to produce more sophisticated and visually appealing outputs, ultimately enhancing the overall user experience of AI applications.

Hallmark significantly enhances the quality of AI-generated landing pages by systematically replacing common 'AI tells'—those tell-tale signs of generic AI design—with more refined and aesthetically pleasing patterns. This improvement is achieved even with one-shot prompting, meaning a single, initial prompt can yield a complete website with a vastly improved baseline design. Hassan El Mghari highlighted that while one-shot prompts provide a strong starting point, much of the 'magic' in achieving truly high-quality results lies in iterative follow-ups.

By engaging in repeated refinement and adjustments, users can push AI outputs beyond initial generation to achieve a more polished and professional look. This iterative process allows for continuous improvement, moving away from typical AI-generated markers like excessive gradients and toward designs that are indistinguishable from human-crafted work, offering a significantly better aesthetic foundation than standard AI models.

Hassan El Mghari strongly advocates for the strategic use of smaller, faster open-source models during the iterative development cycles of AI applications. He believes that many developers undervalue the importance of these models, despite their significant benefits in terms of cost and speed. El Mghari suggests a workflow where the initial base of an application is built using larger, more comprehensive models like Claude, which are capable of handling complex foundational tasks.

Following this, subsequent iterative refinements and smaller-scale adjustments should be carried out using specialized open-source models. This approach strikes a crucial balance between leveraging the power of large models for initial generation and utilizing the efficiency of smaller models for rapid, cost-effective iteration, thereby optimizing both development time and expenditure.

El Mghari highlighted GLM 5.2 as a newly released open-source model that stands out for its exceptional design capabilities. Launched just two weeks prior to his talk, GLM 5.2 is presented as one of the first open-source models to genuinely excel in generating high-quality designs. According to El Mghari, GLM 5.2 is capable of producing results that are nearly indistinguishable from those generated by much larger and more expensive models, such as Opus 4.8.

This model offers significant advantages in both speed and cost-efficiency, being approximately five times cheaper to operate while delivering comparable performance. The emergence of GLM 5.2 signifies a crucial development in the AI design landscape, making high-quality design generation more accessible and affordable for developers and designers alike, potentially democratizing advanced design capabilities.

Hassan El Mghari emphasized that achieving design excellence, particularly for individuals who are not professional designers, is primarily a result of structured iteration rather than an innate talent. He argued that even basic AI-generated prototypes can be dramatically improved with just one or two focused follow-up prompts. This iterative process allows non-designers to refine outputs and bring them closer to a professional standard.

The ultimate goal, El Mghari clarified, is not to transform every developer into a professional designer. Instead, the focus is on ensuring that all applications, regardless of who designs them, possess a professional and cohesive feel. By consistently iterating and refining AI-generated designs, even those without formal design training can produce applications that are aesthetically pleasing and effective, bridging the gap between functional development and polished user experience.