Did Anthropic just kill Figma?
TL;DR
Claude Design feels like Anthropic’s first genuinely differentiated product — Theo starts skeptical but ends up calling it “the best software Anthropic’s ever shipped,” because the comment-driven UI iteration and Claude Code handoff are actually useful for real landing-page work.
This is aimed more at Figma-style prototyping than direct in-repo editing — even with a codebase mounted, Claude Design mostly builds mockups around your system, then exports HTML/CSS/JSX context for Claude Code to implement, which Theo says mirrors how big companies keep design systems in sync.
The collaborative workflow is the killer feature — Theo loves being able to click elements, leave multiple comments, batch them into one prompt, draw on the UI, and hand the result to Claude Code, calling that flow “magical” for designers and engineers working together.
The product is promising, but the bugs and limits are brutal — designs vanished mid-session, previews froze, response tags leaked into output, and Theo burned through 65% of his Claude Design weekly quota on a $200/month Max 20x plan in one demo.
Anthropic’s UI taste shows up, but AI design still makes weird mistakes — the tool produced bad word wrap, fake or wrong logos, inaccurate mock UIs, and awkward scroll hints, yet still got to a landing-page direction Theo said he’d happily keep refining off-stream.
Theo’s deeper point is that great design tools empower bridge-builders, not just designers — he tells a long story about a Twitch designer named Iris, whose ability to absorb product, user, and engineering constraints convinced him that tools like this could massively amplify people who connect those worlds.
The Breakdown
Theo opens skeptical, but very ready to believe
Theo says he was “excited to stop talking about Anthropic,” then immediately gets pulled back in by Claude Design. He’s unusually bullish because he already uses Claude Opus for UI work on projects like Lawn and Shoe, and he’s seen firsthand how even Anthropic’s simple markdown “design skill” noticeably improves interface generation.
First look: very Figma-like, with some smart AI-native ideas
Jumping straight into the product, he notices features like importing a design system, mounting a local folder, live CSS “knobs,” comments that batch into one message, and export to Claude Code. He immediately points out one core prompting trick the product itself seems to understand: asking for multiple varied options instead of regenerating the same prompt over and over.
Redesigning the T3 Code landing page in public
Theo uses the real T3 Code marketing site as the test case and gives Claude Design a concrete brief: highlight support for Claude Code, Codex, OpenCode, and Cursor; emphasize open source, Git/GitHub workflows, parallel work, and dark mode for “users sensitive to light mode.” The early result is mixed: he likes the overall interface and collaborative comment model, but the actual design has bad word wrap, wrong logos, fake UI screenshots, and a lot of classic “AI doing design things.”
Why this matters: his Twitch story about Iris
The most human part of the video is Theo’s story about Iris, a Twitch designer who changed how he thought about design. She took his technical pushback on impossible CSS interactions, came back hours later with a revised design that solved the underlying implementation issue, and later even built an interactive prototype for what became Mod View using vanilla HTML, CSS, and jQuery despite having no coding background.
The bigger thesis: design is about bridging worlds
Theo uses that story to make his real point: the best people in product are the ones who bridge gaps between users, PMs, designers, frontend, and backend. He says his career at Twitch accelerated because he was a frontend engineer who also talked to backend, product, and users — and he sees Claude Design as potentially giving people like Iris much more power to prototype, test, and collaborate without waiting on engineering.
The product starts wobbling under real use
As he spins up more variants, Claude Design starts showing Anthropic’s usual rough edges: frozen sessions, missing files, previews disappearing, weird response tags bleeding into the UI, and quota burn that feels scary fast. He notes that while he’d barely touched Claude that day, Claude Design alone had already eaten 18%, then 31%, then nearly 50% of his weekly allowance — with some Pro users reportedly getting only two prompts before hitting limits.
The handoff to Claude Code is the moment that wins him over
Once he gets a version he likes, Theo exports it to Claude Code and is impressed by the handoff: Anthropic bundles the design files, assets, pasted screenshots, and even chat history into a context package Claude Code can use. He likes the philosophy here — not a fancy proprietary bridge, just “zip all the context and throw it at the agent with a link” — because it feels robust and practical.
Final verdict: buggy, expensive, but absolutely a real product
By the end, Theo lands in a very different place than where he started. He still sees sharp edges everywhere, says the usage looks expensive, and jokes that if he were Figma he’d be “scared as [__]” — but he also says the result is good enough that he plans to keep refining the T3 Code page off-stream. His closing read is simple: this is useful, different, and finally plays to Anthropic’s strengths in design taste and high-resolution visual understanding.