Design is the main reason Hassan El Mghari's simple AI apps reach users, so he spends the final 10 to 20 percent of the work improving their UI.
2
Hallmark improves generated websites by blocking familiar AI-generated patterns and supplying themes that give models better visual references.
3
Agents produce a starting point, and better results come from screenshots, detailed prompts, small feature batches, and repeated iteration.
Summary
Hassan El Mghari builds around ten apps a year without being a designer. He says design and UX have helped some of those apps reach millions of users. His concern is the recognizable look of AI-generated interfaces, including purple gradients, italic headings, all-caps pills, gradient logos, excess emoji, and poor spacing. Hallmark, a design skill he built, encodes these patterns as gates and gives models a library of themes. He also recommends using a fast, smaller model such as GLM 5.2 for repeated app changes after creating the first version in Codex or Claude Code. In his demonstration, people struggled to distinguish a page made with GLM 5.2 from one made with Opus 4.8. His practical advice is to keep an inspiration vault, provide screenshots, write longer prompts, save recurring preferences in a skill file or AGENTS.md, and handle one or two features at a time. The generated app is only a base. The rest comes from iteration.
Hassan El Mghari says he has built about ten apps a year for five years, with some reaching millions of users. His apps are usually simple one-page products with attention paid to design and animation. He credits design and UX as the main reason they found users. Examples include a logo creator used by about 85,000 people, a comic generator that can print a physical comic, a subtitle tool used by about 8,000 people, and a cloud agent that creates a pull request from a GitHub repository. His goal is not to make the world's best-designed apps. It is to make products that feel much better than raw AI output.
AI-generated interfaces have recognizable visual tells
El Mghari says people can often identify a vibe-coded app within seconds, even when they cannot explain the reason. He lists repeated patterns: purple gradient backgrounds, italic headings, a "scroll to explore" prompt, all-caps pills with widely spaced letters, gradient logos, too many emojis, random graphics, and spacing or padding problems. He argues that these patterns can be named and turned into explicit instructions. Once a builder knows that a logo, heading, or background repeatedly looks wrong, the preference can be given directly to an agent rather than left as an unspoken intuition.
Hallmark blocks slop patterns and supplies design themes
Hallmark is El Mghari's design skill for reducing the familiar look of AI-generated websites. It codifies the patterns he calls AI slop and turns them into gates, such as telling a model not to use a purple gradient or italics in a title. Hallmark also provides a library of themes that models use as context when generating a site. El Mghari says strong inspiration improves model output, and he had more than 10,000 people try Hallmark about a month and a half after launch. In his examples, Hallmark-generated pages have fewer familiar tells and provide a better starting point, although they still need further work.
Small open source models can handle repeated design changes
El Mghari usually starts an app in Codex or Claude Code, then uses a smaller open source model for iteration. He recommends GLM 5.2 because he considers it the first open source model he had seen that was very good at design. In his comparison, one landing page came from GLM 5.2 and another from Opus 4.8, and the audience had difficulty telling them apart. The GLM 5.2 version was created faster, while the Opus version cost five times as much and took longer. He says the cheaper model still needs to be good enough, but iteration does not always require the most expensive model.
A few follow-up prompts can improve an otherwise usable page
El Mghari shows an image playground made with GLM 5.2 in one prompt. The first version works, but it has several familiar AI design tells. After one or two follow-up prompts, the page has a better logo, improved loading states, better animations, and cleaner spacing. The overall site remains roughly the same, but it feels more finished. He uses this example to argue that the first generated page should not be treated as the final result. A fast model can make these smaller changes quickly, which makes repeated improvement practical.
El Mghari calls references and screenshots his most important practical recommendation. He keeps an inspiration vault containing websites and apps he admires. When starting a new app, he might ask for a mix of the look of Duolingo and other saved examples, then paste many screenshots into the prompt. He says the output is always much better when the model has this visual material. The references help turn a vague preference into something the agent can inspect. This is also how he thinks about Hallmark's themes: better examples give the model a stronger basis for making design choices.
Longer prompts and smaller feature batches improve control
El Mghari records voice notes that last one to three minutes, then turns them into prompts describing the app, its users, its controls, its behavior, and its visual inspiration. His prompts are often two or three paragraphs long. He also advises splitting work into one or two features per prompt instead of asking for seven features at once. After the initial prompt, he queues more detailed requests and lets the agent work through them. He saves accumulated preferences in a skill file or AGENTS.md, including recurring instructions such as how to handle a logo. These files preserve decisions that would otherwise have to be repeated.
El Mghari objects to treating a one-shot app as finished. Whatever the agent creates is a base that needs more context, more inspiration, and further back-and-forth. The builder has to simplify the app and improve how it looks and feels through iteration. His process combines a strong initial prompt, screenshots, smaller follow-up tasks, model selection based on the task, and a growing file of design preferences. The agent can produce a functioning starting point, but the final quality depends on the human deciding what to keep, what to remove, and what to ask for next.
"You can kind of look at it and within a second or two you can kind of tell that this kind of looks AI generated."02:49
Who should watch
You are shipping AI-generated interfaces and people can recognize the visual style before they understand the product.
You want a repeatable way to give an agent design preferences instead of correcting the same logo, spacing, or typography problems every time.
You are paying for a large model on every iteration and want to test whether a smaller open source model is sufficient for interface changes.
