Stephen Haney announces QuiverAI’s SVG generation inside Paper – vector output from text or references, aimed at logos, illustrations, and animations. QuiverAI​’s models have been on my radar since Dann Petty joined the team as a Founding Product Designer.

Paul Bakaus has packaged 23 design commands into a single agent skill that teaches Claude Code, Cursor, Codex, and Gemini CLI how to actually design. Type vocabulary, color systems, motion, spatial logic — the foundations most prompts miss. The Live mode that writes accepted variants back to source is where it gets genuinely interesting.

Figma Make now supports custom skills — markdown files that capture conventions and workflows you use repeatedly, callable from any prompt with a slash command. Pair /build-from-prd with a Notion connector and any PRD becomes a prototype that matches your standards.

A useful baseline study on how people actually use AI well. The most uncomfortable finding for designers: in conversations that produce artifacts (code, UI, documents), users are less likely to question the model’s reasoning. Polished output suppresses critical evaluation, even though that’s exactly when it’s most needed.

Brian Lovin shows the Notion design team’s Prototype Playground, a single Next.js repo on Vercel where every designer gets a namespaced folder and a small set of shared Notion-flavored components. The interesting parts are not the scaffolding but the slash commands and skills layered on top: /figma runs a three-phase loop with the Figma MCP and Chrome DevTools MCP until the build matches the source (~80% on the first try), and a find-icon skill writes its own TypeScript search script after the team got tired of Claude hallucinating ​“search-icon.” See also the Stripe and Vercel pieces below for the same pattern at other companies, as well as another interview with Brian at Dive Club.

Owen Williams, design manager at Stripe, walks Claire Vo through Protodash, the internal prototyping platform he’s been building for 18 months. The V1 was a bundle of Cursor rules plus an MCP server wired to Sail, Stripe’s design system, so any designer can open the repo and build a page without ever touching React or routing. V2 wraps the whole thing in a browser UI running on internal dev boxes, with embedded LLM chat, click-to-annotate feedback, a design review mode, and fidelity toggles (monospace, grayscale) to signal work-in-progress. The same pattern as the Notion and Vercel pieces in this section: production design system plus a thin internal harness, calibrated to one team’s review culture in a way no off-the-shelf tool can be.

Hannah Hearth runs a tooling Show and Tell with her team at Vercel and writes up the results: Codex + Claude pair programming, Conductor for parallel agent threads, UI Fork for in-browser variant exploration, and Cleanshot’s Pin tool still earning its place.

“OpenAI’s newest image model, ChatGPT Images 2.0, is now available to use with Make Image and Edit Image in Figma Design, Draw, Slides, Buzz, and FigJam, and also in Figma Weave. ChatGPT Images 2.0 reliably generates high-quality images, and excels at producing smarter visuals like infographics or multilingual generations, executing on better editing and aesthetics, and preserving faces across generations.”

Figma adds granular AI-credit governance scoped to billing groups on the Enterprise plan.

Figma now lets you add reference images by clicking the new ​“Add reference” button, copy/​paste, or drag-and-drop when generating or editing with AI. Closer to how designers actually work — reach for a moodboard first, prompt second.

Grace Li lists 7 common design smells in GPT 5.5: huge typefaces with tight tracking, lack of textures, bento boxes with unrelated icons, pill-with-dot in hero sections, gradient keywords, grid background, and rounded cards with three nested cards. ​“TLDR: if you’re vibe coding with GPT 5.5, the easiest wins are: loosen the tracking on your headlines, delete the status pill, and pick one accent color instead of a gradient. That’ll get you out of the uncanny valley before the next model release closes the gap on its own.”

GPT 5.5 is strong at programming, but not great at visuals: ​“Across 5,000+ preference pairs, GPT‑5.5 ranks #13 in Website Arena, #16 in UI Component Arena, and #19 in 3D Design Arena. It loses to Claude Opus 4.7 and Sonnet 4.6, to Gemini 3.1, to GLM 5.1, and to Kimi K2.6.”

“Claude Design can read a design system carefully when the prompt is about the system. When the prompt is about a composition that uses the system, it stops respecting the components and just generates lookalikes.” TJ Pitre spent a few hours testing Claude Design against two real design systems and concludes that the tool references your system, but doesn’t consume it. Claude would happily inline HTML tags with style props instead of importing them from your component library.

“The pitch for Claude Design’s workflow is roughly: I have a design system, I want to generate new product surfaces from it, and I want AI to do most of the lift. That workflow exists today. You can pair Figma with an MCP server like our Figma Console MCP, or with Figma’s own MCP server, or with Code Connect, and then point an AI app generator at it. Lovable, v0, Replit, Figma Make, Claude Code working inside your repo. Your Figma file stays the canonical source. Your codebase stays the production surface. AI does the generation in between.That flow is more linear, more honest about where source of truth lives, and it produces output that actually uses your component library, because the AI is operating inside the repo where the components live.”

Kris Puckett, Design Manager at Stripe, spent months building Epilogue, a real iOS app with 14,000 lines of Swift, entirely through conversation with Claude. This essay is a specific and honest account of what the designer-building-with-AI experience actually looks like: what broke, what he learned about asking precise questions, what ​“vague frustration keeps you stuck, specific confusion gets you answers” actually looks like in practice. ​“I realized the bottleneck was never coding ability. It was articulation. The ability to describe what I wanted clearly enough that something else could build it.”

Karri Saarinen, CEO of Linear, writes one of the more grounded takes on AI’s current state. Linear’s cloud agent now fixes more than 1,000 issues per month, but Karri is clear that hard problems remain hard and design tools are still challenging to use. On having a design tool operate directly on the production codebase: ​“A lot of the design work I do is not production design. I am not trying to implement the final version or test every edge case. Most design work is about making decisions, understanding the problem, and finding the fit. That process generates many variations and messy ideas.”

The expertise paradox section is the most useful: ​“AI often feels most impressive in domains where you know the least.” Expertise makes AI harder to use but also more valuable, because experts know how to steer, constrain, and evaluate the output.

“Designers have always (and will always) answer the question ​‘What’s worth making?’ ” Joel Lewenstein, Head of Design at Anthropic, argues that as the cost of software drops, the most important decisions shift from ​“can we make this?” to ​“should we?” Design, in his framing, is what narrows the possibility space fast enough to keep up with the speed of delivery. He describes Claude Design as a tool for getting ideas ​“good enough to move discussion forward,” cutting idea-to-internal-feedback time from days to hours.

“A state-of-the-art image model that can take on complex visual tasks and produce precise, immediately usable visuals, with sharper editing, richer layouts, and thinking-level intelligence.” OpenAI’s second-generation image model promises a step change in instruction following, precise object placement, dense text rendering, and cross-aspect-ratio generation.

DESIGN​.md encodes your design rules, colors, typography, and component preferences in a plain Markdown file that AI agents can parse and validate against. Google moves DESIGN​.md from a Stitch-specific feature to open-source format specification. If the format gains traction across tools, it could become the missing link between design systems and AI generation pipelines.

A useful companion to Google’s announcement above. Meng To shares 15 takeaways from actually using the format: when to Remix vs. Iterate, how to treat DESIGN​.md as ​“reusable project memory,” and why curation is part of the design process. The most actionable takeaway: ​“Start with DESIGN​.md, generate the first design, remix and expand it, create section variations, move into a builder, then assemble the full site.” Don’t miss his video tutorial on turning a DESIGN​.md into landing pages, mobile screens, and motion design.

“I’m increasingly sure that 2026 signals the end of product design as a full-fledged stand-alone function within companies.” Gokul Rajaram makes the controversial maximalist case: startups will outsource design systems to consultants, AI will generate the UI, and design headcount will shrink to 20% of its current size at larger companies.

It’s an argument that reduces design to UI mockup production, which is where most of the responses below push back. The reactions to Claude Design launch in the previous issue have already discussed confusion in terminology, the difference between design practice vs. design production, and made the case that ​“real design is the search for fit between form and context, not the generation of form itself.”

“The role of someone who figures out what needs to exist, why, how it should work, how it should be positioned, differentiated and made memorable has never been more in demand.” Josh Puckett (also the maker of Pica included below) separates the deliverable from the practice: he agrees the mockup-maker role is going away, but argues the ​“what should this be and why?” role is more in demand than ever.