The Vibe Coding Stack Is Finally Getting a Design Layer
Taste, Rams, Lovable aesthetics, QuiverAI in Paper, and Figma Make custom skills — five unconnected announcements pointing at the same thing.
Something shifted this week. A handful of announcements, all unconnected, all pointing at the same direction: the tools that generate code are starting to take design seriously.
Start with Taste (buildwithtaste.com). It's a menu-bar app that watches you screenshot and bookmark things you like — UI patterns, typography choices, reference sites — and synthesises them into a DESIGN.md file you can drop into Cursor↗, Claude Code↗, or Codex↗. Your aesthetic preferences, made legible to an AI. Instead of fighting the model's defaults on every project, you set the terms once. That's a small workflow change with a disproportionate effect on output quality.
Then there's Rams (rams.ai), which tackles the other end of the same problem. Once the AI has built something, Rams reviews it — accessibility errors, spacing inconsistencies, WCAG violations — and files structured feedback on every PR. It's framed as a design engineer, not a linter, and that framing matters: it catches things that are technically valid code but visually wrong. It ships as a skill file for Claude Code↗ and Cursor↗, so the setup friction is low.
Lovable↗ shipped an aesthetics update this week that fits the same pattern. Before you start building, you can specify typography, layout, and colour preferences in plain language and see mockups reflecting those choices. It delays the first build by thirty seconds and saves twenty minutes of iteration on the output. For solo builders without a design background, this is the most useful thing they've shipped.
The most technically interesting news was QuiverAI's Arrow 1.0 landing inside Paper↗. Arrow is a purpose-built SVG generation model — a16z-backed, trained specifically on vector graphics rather than adapted from a general image model — and it hit #1 on SVG Arena the day it launched. The integration means Paper↗ users can generate and edit real, editable SVG assets in context. For a tool already betting that real code is the design substrate, native vector generation is a logical step.
Figma↗ Make shipped custom skills: reusable slash commands that package a workflow into a stable set of instructions for an AI agent. /insert-sample-data, /apply-component-spec — things you'd otherwise re-explain every session. Team-shareable skills are coming next, which is where this gets actually interesting for larger orgs.
None of these individually changes how most designers work day-to-day. But the direction is consistent. The parts of the AI coding stack that previously treated visual quality as an afterthought are each, separately, reaching toward it. Whether they reach far enough is a different question.
— Keats