Loading summary...
7 min read (79% time saved)
This video compares the frontend design capabilities of various AI models, revealing that while models like Gemini can produce good initial designs, Opus 4.5, when combined with a specific 'frontend design skill' (a markdown file providing design guidelines), significantly outperforms others in generating high-quality, unique, and iteratively refined designs, effectively avoiding generic AI aesthetics.
Summarize another video
Press ⌘K to quickly paste a new URL
The presenter introduces the current state of frontier AI models like Gemini 3 Pro, Opus 4.5, and GPT 5.2, acknowledging their general capabilities but questioning their proficiency in frontend design. The core problem identified is the prevalence of 'AI slop'—generic designs with purple gradients and repetitive fonts. The presenter teases a surprising solution that elevates Opus 4.5 from the bottom of the design ranking to the top, showcasing examples from personal projects.
The video features a sponsored segment by Railway, a deployment platform. The presenter shares his personal connection to Railway, having almost joined the company. Key benefits highlighted include:
To compare design sensibilities, the presenter sets up six agents: two with Gemini, two with Claude Code (Opus), and two with GPT 5.2 (Codeex). All models receive a prompt to create five unique marketing homepage designs for an 'Imagen Studio' app called T4 Canvas. A crucial 'hack tip' is introduced: instructing the model to create five different designs, ensuring variety and revealing the underlying design templates the models were trained on. GPT5 models are estimated to have ~10 templates, Opus ~20, and Gemini ~1-2.
The main topic of the video is introduced: the 'frontend design skill.' This skill is a markdown file from the Claude Code GitHub, acting as a reusable context or description of desired behavior for the AI model. It guides the model to create 'distinct production grade front-end interfaces that avoid generic AI slop aesthetics,' emphasizing intentionality, creative choices, and avoiding clichés like purple gradients and common fonts. The skill explicitly instructs the model to make unexpected choices and vary themes.
The presenter first showcases the default designs from Opus 4.5, which he ranks as the worst. These designs are criticized for being generic, featuring disliked purple gradients, poor text visibility, and overall uninspired layouts. Examples include designs with awkward box placements, repetitive structures, and 'awful noise backgrounds,' confirming the initial low ranking of Opus without the skill.
Next, GPT 5.2's designs are reviewed. The presenter notes that GPT 5.2's system seemed to override his instruction to not use the frontend design skill, likely applying it anyway. The designs are characterized by being text-heavy, editorial, and often having poor contrast or similar structures across iterations, suggesting a limited set of underlying templates. Some designs are deemed unworkable due to readability issues.
The Gemini 3 Pro designs are presented, but not without significant frustration due to the Gemini CLI's instability and breakage. Despite the technical difficulties, Gemini's initial designs are considered the 'coolest so far,' showing good variety and interesting elements like placeholder images and unique bar designs. However, they also exhibit outright mistakes, such as text cropping issues and poor detail execution, despite good layout concepts. One design even crashed.
Revisiting GPT 5.2, the designs generated with the skill (or at least, where the system likely applied it) are shown. While some variety is present, many still lean towards an 'editorial brutalist aesthetic' and are not considered exceptional. The presenter notes that it took many attempts to get a design that didn't look like all the others.
The presenter explicitly demonstrates Gemini 3 Pro's designs without the frontend design skill, noting a distinct aesthetic often seen from Gemini. These designs are praised for being 'really cool and nice,' featuring unique elements like fake drop shadows. However, some designs are criticized for being overly complex, having poor blur effects, or outright mistakes. When Gemini is then shown with the skill, the designs improve, appearing more like 'generic templates' but with better execution, though still needing additional tuning. Some designs are still deemed terrible, like a 'glass skeuomorphic one.'
This section highlights the dramatic improvement in Opus 4.5's designs when the frontend design skill is applied. The designs are described as 'awesome,' 'minimal with a cool thing,' and 'really cool,' showcasing significant variety and attention to UX (e.g., a navigation bar for switching designs). The contrast between Opus's designs with and without the skill is emphasized as 'insane,' demonstrating the skill's transformative effect. While some designs are not to the presenter's personal aesthetic, their execution and quality are far superior.
The Kimmy K2.5 model is tested, but the presenter notes significant difficulties in getting it to work due to its poor handling of harnesses and code structuring. Without the skill, Kimmy's designs are described as 'cool neo retro vapor wave aesthetic' but often 'too extra' with conflicting animations and design elements, feeling like a 'designer's first day on the job.' With the skill, Kimmy's designs show some improvement, but still exhibit issues like incorrect color choices and poor grouping of text, and are generally not considered usable.
The presenter conducts a live poll with his chat to gauge preferences between the best Gemini and Opus designs. He then demonstrates his iterative design process: taking favored designs and instructing the model to create more iterations based on them. Gemini's iterations are largely disappointing, failing to incorporate the liked elements and producing 'garbage.' In contrast, Opus, when given the same task, produces 'meaningful iteration,' clearly inspired by the preferred designs, demonstrating its superior ability to understand and build upon user preferences. The chat overwhelmingly agrees that Opus with the design skill is the better choice for malleability and adherence to instructions.
The video concludes by explaining how to set up and use the frontend design skill. It's available through Vercel's agent skills directory (skills.sh). Users can copy the skill, paste it into their terminal, and select which tools (like Cursor) they want it to work with, either globally or per project. The presenter reiterates that this simple markdown file unlocks powerful design capabilities in Opus, making it his preferred model for frontend design despite his general preference for Codeex for other tasks.
"If you want to make things that don't look like AI slop with everyone's favorite purple gradients and the same exact font on everything, what models the best at design?" — Presenter
"This whole document is effectively built to steer the model towards better design." — Presenter
"Never use generic AI generated aesthetics like overused font families like robboto intererial system fonts, cliche color schemes, particularly purple gradients on white backgrounds, predictable layouts and component patterns, and cookie cutter designs that lack context specific character." — Frontend Design Skill (markdown file)
"Opus' ability to understand what you like about a design and then do follow-ups based on that is significantly stronger overall." — Presenter
"Gemini randomly can put out good designs, but Opus actually does mostly what you tell it to." — Darth Mini Mall (chat user, quoted by presenter)
WorldofAI
5 min read
IndyDevDan
7 min read
IndyDevDan
6 min read
Kenny Liao
7 min read
IndyDevDan
8 min read
Ben AI
5 min read