Summary
Hassan El Mghari, head of developer experience at Together AI, argues that design quality has become the sharpest differentiator for AI-generated applications — and that most vibe-coded products share a recognizable set of visual failures that can be systematically avoided. Drawing on five years of building apps that have reached millions of users, El Mghari catalogs the tells of AI-generated UIs (purple gradients, italic headers, excessive emojis, inconsistent spacing) and explains how he codified them into a system called Hallmark.
Hallmark operates as a design skill for AI coding agents: it encodes anti-patterns as explicit guardrails while supplying curated theme libraries that push generated UIs toward cleaner, more distinctive aesthetics. The talk includes side-by-side comparisons of outputs with and without Hallmark across landing pages for a learning app, an invoicing tool, and other use cases — all generated in a single prompt. El Mghari also runs a live comparison between GLM 5.2, an open-source model released roughly two weeks before the talk, and Claude Opus 4.8 for design iteration tasks, finding the smaller and faster GLM competitive on output quality at significantly lower cost and latency.
The session is practically oriented toward developers and product builders who want their AI-generated products to stand out, and introduces Together AI’s inference API as the infrastructure layer for running models like GLM 5.2, Flux, and others in these pipelines.
📺 Source: AI Engineer · Published August 21, 2026
🏷️ Format: Keynote Launch
