{ "video": { "id": "7GMKdpLsxwU", "title": "The Missing Layer: Design Taste in AI Agents — Hassan El Mghari, Together AI", "duration": 850, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer" }, "analysis": { "video_id": "7GMKdpLsxwU", "title": "The Missing Layer: Design Taste in AI Agents — Hassan El Mghari, Together AI", "one_liner": "Hassan El Mghari (Together AI) argues that the 10–20% of extra effort spent on UI is the biggest competitive advantage for AI apps, and shows how to get it: codify the visual tells of \"AI slop\" into a skill, feed agents real references and screenshots, and iterate with cheap fast open-source models like GLM 5.2.", "summary": "The speaker builds ~10 apps a year and credits design and UX — not the models — as the number one reason some got millions of users. He catalogues the recognisable tells of vibe-coded UIs (purple gradients, italic headers, \"scroll to explore\", all-caps spaced pills, gradient logos, emoji), then demos Hallmark, a design skill he built that codifies those tells as \"AI slop gates\" and feeds the model a library of themes as context. The second half is a workflow argument: build the base in Codex or Claude Code, then iterate with a smaller, faster, cheaper model — he ran a live blind test where the audience could barely distinguish a GLM 5.2 landing page from an Opus 4.8 one that cost five times as much and was slower.", "key_points": [ "He's built roughly 10 apps a year for five years; some reached millions of users, and he attributes that primarily to design and UX, despite not being a designer.", "Demoed apps: a logo creator (~85,000 users), Make Comics (comic book starring the user as superhero, running on an iPad at the Together AI booth and printing physical comics), a video subtitle generator (~8,000 users), an open-source-model chat app, a landing-page variation generator (5–6 variants to choose from), and an AI cloud agent that takes a GitHub repo, spins up a sandbox and opens a PR.", "The concrete AI-slop tell list: purple gradient backgrounds, italics in headers, \"scroll to explore\", all-caps pills with wide letter spacing, gradient logos, heavy emoji use, and spacing/padding issues — he says you could enumerate 20–30 such patterns.", "Hallmark is a design skill he built that does two things: codifies the slop patterns as \"AI slop gates\" telling the model what not to do, and supplies a set of hand-built themes as context. Launched about a month and a half ago, over 10,000 people have tried it; it's indexed mainly on landing pages.", "Before/after examples (kids' learning app, invoicing app, indie podcast) are all one-shot single prompts; he's explicit that the Hallmark output isn't a perfect landing page, just a much better base to iterate from.", "Live blind test: two landing pages, one by GLM 5.2 and one by Opus 4.8. Only about four hands picked the GLM one correctly. The GLM page was generated much faster; the Opus page cost five times as much and was slower — and in a second pair, the Opus output arguably looked more AI-generated.", "GLM 5.2 came out two weeks before the talk and is, in his opinion, the first open-source model that's genuinely very good at design.", "An image-playground app one-shot with GLM 5.2 still had obvious AI tells, but one or two follow-up prompts produced a better logo, better loading states, animations and spacing.", "He cites Cursor's Composer 2.5 — an open-source model they post-trained — as proof that fast cheap models feel \"magical\" for iteration loops." ], "takeaways": [ "Learn to name the AI tells, not just sense them — most people spot an AI-generated page in two seconds but can't say why, and you can only tell the model \"don't do this\" once you can articulate it.", "Keep an inspiration vault: save every app or site you admire, then paste a pile of screenshots in when starting a new app (\"a mix of Duolingo and this app and this app\"). This is the single highest-leverage habit he names.", "Persist your accumulated design preferences into a skill or AGENTS.md — every time you have to regenerate the logo, write down the rule — or use something like Hallmark if you don't have your own yet.", "Write much longer, more specific prompts — he records 1–3 minute voice notes and ends up with 2–3 paragraph prompts covering the user, the layout and the inspiration — but split features across prompts rather than cramming seven features into one, queuing follow-ups after the big opening prompt.", "Split model tiers by task: build the base in Codex or Claude Code, then iterate with a cheaper, faster open-source model (GLM 5.2) that's still sufficiently good.", "Never ship the one-shot. Treat whatever the agent produces as the base and go back and forth with additional context and inspiration." ], "topics": [ "design", "ai-slop", "vibe-coding", "prompting", "open-source-models", "agent-skills", "ui-ux", "iteration" ], "tools": [ "Together AI", "Hallmark", "GLM 5.2", "Opus 4.8", "Claude Code", "Codex", "Cursor Composer 2.5", "nanobanana", "AGENTS.md", "GitHub", "NVIDIA H100", "NVIDIA B200", "Duolingo" ], "quotes": [ { "text": "vibe coded apps kind of all look the same. They have the same tells.", "at": "03:21", "url": "https://www.youtube.com/watch?v=7GMKdpLsxwU&t=201s" }, { "text": "just doing a little bit of extra effort like a little 10 to 20% after just focused on the UI uh is a really really big competitive advantage", "at": "03:13", "url": "https://www.youtube.com/watch?v=7GMKdpLsxwU&t=193s" }, { "text": "I show them a website and they're like oh that's AI generated right in two seconds but they can't tell me why", "at": "09:43", "url": "https://www.youtube.com/watch?v=7GMKdpLsxwU&t=583s" }, { "text": "If you take one thing away from this talk, please give your agents references and screenshots.", "at": "10:50", "url": "https://www.youtube.com/watch?v=7GMKdpLsxwU&t=650s" }, { "text": "whatever your agent creates is just the base, right? And it's on you to kind of like give it additional context, give it additional inspiration", "at": "13:30", "url": "https://www.youtube.com/watch?v=7GMKdpLsxwU&t=810s" } ], "words": 3540 }, "summary_url": "/#7GMKdpLsxwU", "transcript": { "html": "/transcripts/7GMKdpLsxwU.html", "txt": "/transcripts/7GMKdpLsxwU.txt", "vtt": "/transcripts/7GMKdpLsxwU.vtt" } }