AI Agent Design Needs Taste, Not Just Better Models

Why visual judgment, sharper prompts and better references may matter more than another model upgrade

AI Agent Design has an awkward new problem: an application can be technically impressive and still look as if it came from the same digital assembly line as everything else. The interface works, the page loads and the copy is polished enough. Yet within seconds, something feels generic.

That feeling is not mystical. In his AI Engineer talk, Together AI developer experience lead Hassan El Mghari argues that people are already learning to recognise the visual fingerprints of an AI-generated product. The challenge for founders is no longer simply getting an agent to build something. It is teaching the agent what good judgment looks like.

Why AI-generated interfaces feel so familiar

El Mghari points to a growing collection of familiar tells: purple gradient backgrounds, italic headlines, unnecessary “scroll to explore” prompts, widely spaced all-caps labels and an enthusiastic surplus of emoji. None of these choices is automatically wrong. The problem is that they appear so often, and with so little connection to the product, that they stop feeling intentional.

This is what makes “vibe coded” products easy to spot. The code may have been produced quickly through natural-language instructions, but the design often reflects the model’s safest average rather than the company’s character. A founder can end up with a competent interface that says almost nothing about why the product exists or whom it serves.

A generated interface can be functional in minutes and forgettable just as quickly.

The same tension appears in the wider world of marketing agent stacks. Automation increases output, but more output does not automatically create distinction. When every team has access to similar tools, judgment becomes the competitive layer.

Taste can become a practical system

El Mghari’s answer is Hallmark, a design skill he released roughly six weeks before the talk and which, according to the episode description, had reached more than 10,000 users. Hallmark turns recurring design mistakes into gates and gives the model a library of themes. In plain English, it creates rules for what to avoid and supplies stronger starting directions.

That matters because “make it beautiful” is not useful creative direction. A model needs signals it can act on: which patterns feel tired, how typography should behave, what kind of visual rhythm suits the product and which references capture the intended mood.

Codifying taste does not mean forcing every product into one template. It means making preferences visible enough that an agent can use them. A strong system creates boundaries while leaving room for different identities inside them.

The biggest model is not the whole answer

El Mghari has shipped around ten applications a year for five years, with some reaching millions of users. He credits design with helping those products travel beyond a technical audience. His workflow often begins in Codex or Claude Code and continues through iteration with a smaller open-source model.

During the talk, he compares two landing pages, one produced with GLM 5.2 and another with Opus 4.8, and asks the audience to identify which is which. Almost nobody can. The demonstration challenges a comforting assumption: that choosing the most capable model will automatically produce the strongest product experience.

Model capability sets the ceiling, but references, constraints and iteration shape what users actually see.

The lesson is not that model choice is irrelevant. It is that teams may be overestimating the value of switching models while underinvesting in the creative context supplied to them. Whatever the agent produces should be treated as a base for judgment, not as the finished application.

Give the agent something worth looking at

The most practical habit in the talk is wonderfully simple: keep a personal vault of screenshots you admire. When beginning a page or feature, show the agent relevant examples instead of asking it to invent a visual direction from a blank prompt.

References do more than communicate colors or layouts. They help express density, hierarchy, atmosphere and interaction patterns that are difficult to describe precisely. El Mghari’s broader point is that strong inspiration tends to produce stronger work. His personal site also provides a route into the range of products behind that perspective.

This does not mean copying another interface. A good reference answers questions such as “How calm should this dashboard feel?” or “How prominently should the primary action appear?” The team still needs to decide what belongs to its own product, audience and brand.

Prompting works better as a focused conversation

El Mghari recommends giving an agent one or two features per prompt rather than seven. That smaller scope makes it easier to inspect the result, correct its direction and preserve coherence. A long wishlist may feel efficient, but it can bury the decisions that matter most.

He also suggests recording a voice note and turning it into a detailed prompt of several paragraphs. That approach fits a more postliterate working environment, where speaking can capture context and nuance faster than typing a perfectly structured brief from scratch.

Preferences should accumulate instead of disappearing between sessions. Teams can write recurring choices into a skill file or an AGENTS.md file so that the agent gradually receives a clearer design constitution. The benefit is not merely speed. It is continuity across pages, features and contributors.

What founders can learn from AI Agent Design

The central lesson is that AI does not remove the need for taste. It makes taste more valuable because implementation is becoming easier to access. Founders do not need to become professional designers, but they do need to notice patterns, articulate preferences and recognise when a generated solution is only adequate.

• Name the clichés. If your team can identify recurring AI design habits, it can tell the agent to avoid them.
• Build a reference vault. Save interfaces that demonstrate the mood, hierarchy or interaction quality you want.
• Reduce the scope of each prompt. Develop one or two features, review them and then continue.
• Preserve your decisions. Record established preferences in reusable instructions rather than rediscovering them in every session.

AI Agent Design becomes stronger when the agent is treated as a fast collaborator, not an autonomous creative director. The real edge belongs to teams that can combine machine speed with human discernment—and keep refining after the first impressive result appears. For more conversations at the intersection of technology and company building, Innohub’s guide to English-language YouTube channels on leadership, growth and business offers a useful next stop.

▶

Sources

• The Missing Layer: Design Taste in AI Agents — Hassan El Mghari, Together AI — AI Engineer talk published on August 21, 2026; primary source for the article.
• Hassan El Mghari — Speaker's personal website.
• Hassan El Mghari on LinkedIn — Speaker profile and professional background.
• Hassan El Mghari on X — Speaker's public social profile.