Claude Design

Sources#

• A Field Guide to Fable: Finding Your Unknowns
• Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next
• Designing with Claude: From prompt to production
• How the product designer who built Claude Design uses it

Summary#

A product from Anthropic Labs that lets you collaborate with Claude to create polished visual artifacts — designs, prototypes, slides, one-pagers, and more — by describing what you want in plain language. Built by a team that was "three people for most of the development" in "about ten weeks from idea to launching," it went live as a research preview roughly a month before Dan Carey's May 2026 talk (≈April 2026). It is the worked example behind several wiki concepts: Prototype Over PRD, Compounding Loop Optimization, and Build for the Next Model.

Origin#

A designer on the team, Nate Parrott (previously a designer on Claude Code), hacked together the first prototype over a weekend while working on a different bet. It was deliberately thin: the Agent SDK, a very thin IDE wrapper, and an existing skill he'd already been using in Claude Code. The first prototype was "a terminal and a browser and that's about it." Posted to Slack, the team's reactions ("here's what's promising / here's what's totally broken, please fix") became the roadmap for the first couple of weeks. Carey's framing: 99% of the value came not from that prototype but from "ten weeks of iterating and shipping and talking to users every single day."

Key features (and the friction each removed)#

Most features began as fixes to the team's own workflow (Compounding Loop Optimization):

• Multiplayer — real-time co-editing of the same design. Built because teammates were serially handing prototypes back and forth; became users' first request after launch.
• Handoff to Claude Code — push a design into production without re-exporting files and re-typing the conversation's context. Users' second request after launch ("now how do I get this into production?").
• Codebase exploration / GitHub linking — ship in your brand and on top of your existing code.
• Open export — designs export as HTML, CSS, JavaScript (not a closed format), reflecting the "be an open tool" philosophy.
• Self-hosting — the team used "Claude Design to design Claude Design," the purest form of Dogfooding as Product Discipline.

Product philosophy#

• Lift the floor, not the ceiling. The team built advanced pixel-level "power user" controls that a few vocal testers loved, but "everybody else hated them" — confusing and "actively harmful." They ripped them out in a week. The lesson: be "a tool that lifts the level of craft for everybody, not just the ceiling on power users" (cf. Vibe Coding vs. Agentic Engineering's "vibe coding raises the floor").
• Openness for needs you'll never meet. Because some power users want specifics the tool won't support, the bet is maximal openness — HTML/CSS/JS export plus integrations — so specialized users can take designs into their own tools.
• Forward-looking integration claim (verify). As of the May 2026 talk, Carey said that "this week or next week" the team would ship the ability for any design tool to integrate with Claude Design via their existing MCPs (see MCP and Computer Use). Stated as imminent, not confirmed shipped — treat as a dated forward claim pending verification.

Launch and iteration metrics#

• 62 improvements shipped between a Friday launch and the following Monday, all from launch-day user feedback (token efficiency, image handling, codebase exploration, near-instant exports). Carey: not an all-nighter effort, but "very natural" because the team had drilled the daily loop for ten weeks.
• Doubled token limits about one month in, responding to users who liked it and wanted to use it more.
• The capability gaps in the early prototype were closed not by engineering but by Opus 4.7 shipping — the canonical Build for the Next Model payoff.

What it became — the builder's account (July 2026)#

Nate Parrott's own first-person post, a year after the first prototype, is the second detailed source and revises the picture in three ways.

The origin motive was a velocity gap, not a bet assignment. Parrott was the sole product designer on Claude Code for VS Code; when Opus 4.5 landed in November 2025 "the engineers were shipping far more than before, while I was still delivering at the pace I always had. I needed to find a way to catch up." The tool exists because AI accelerated one role in a three-person team and not the other. See Nate Parrott for the full path (failed screenshots-into-chat attempt → a month of side-project search → HTML + split-view → distilling Anthropic's brand into prompts).

The scope reframe came from users, not the team. At an Anthropic Labs offsite idea-pitch session, "every person there threw together slides using it, often in the middle of the meeting before their turn to present. That session convinced the Labs team to staff it." The team then "stopped describing it as a tool for product mockups": Claude Design is for any visual communication — slide decks, landing pages, printable one-pagers, emails, animations, social visuals. Parrott's framing: "one click above product design… visuals whose main job is communication and ideation."

The boundaries are stated explicitly (unusual candor for a first-party post):

• No image model. "A poor fit for logo design — though that hasn't stopped people from trying." Bring existing assets instead.
• Not for production software. "If you're shipping production software, stick with Claude Code." The two round-trip in both directions: sync a Claude Code prototype into Claude Design for canvas iteration, or hand a Claude Design prototype to Claude Code to build from. Ideation vs. production, not a capability ranking.
• Model-bound range. "As models get better at vision, so does the range and quality of work Claude Design can do" — Opus 5 is called out as better than previous Opus models at reading charts, diagrams, and screenshots, which is what makes the deck/memo use case work. Build for the Next Model again, now on the vision axis rather than the coding one.

Status: in beta on Claude Pro, Max, Team, and Enterprise plans as of July 2026 (up from the research preview of the Carey account).

The daily working practice — think off-keyboard, ask for ten options and remix two, wireframe first, make the last mile manual — is compiled separately as Design by Selection.

Prior mention in the wiki#

Boris Cherny had flagged it as a vision example (recorded on Cowork): "Claude design I think is a really good example. It's pretty good today. It's going to get a lot [better]." Carey's talk is the first detailed account of how it was actually built.

Connections#

• Anthropic Labs — the incubator that produced it; the offsite where unprompted slide-making convinced Labs to staff it
• Nate Parrott — the designer who built the first prototype and gave the first-person account
• Design by Selection — the daily practice the tool is used through: think off-keyboard, generate ten options, remix two, hand-edit the last mile
• Claude Opus 5 — the vision gains (charts, diagrams, screenshots) that widen the deck-and-memo use case
• Dan Carey — PM who led it and gave the build account
• Claude Code — origin tech (Agent SDK, a Claude Code skill, thin IDE wrapper) and the handoff-to-production target
• Cowork — sibling Anthropic agent product (non-code knowledge work); Claude Design is the visual-artifact sibling
• Anthropic — vendor
• Claude Opus 4.7 — the model release that closed the early prototype's gaps
• Prototype Over PRD — the tool fast enough to make prototypes replace PRDs
• Compounding Loop Optimization — its features are crystallized loop-optimizations
• Build for the Next Model — its clearest case study
• Dogfooding as Product Discipline — "use Claude Design to design Claude Design"
• Living Design System — HTML/CSS/JS export + design-system output is the same portable-artifact move
• MCP and Computer Use — the integration surface for "any design tool via MCP"
• Vibe Coding vs. Agentic Engineering — "lift the floor for everyone" echoes vibe-coding raising the floor
• Unknowns as the Agentic Bottleneck — the reference technique productized: point it at a component you like and "it reads the underlying code, not just the screenshot," giving it markup and structure rather than appearance (Thariq Shihipar)

Open Questions#

• Did the "any design tool via MCP" integration actually ship on the stated timeline? (Forward claim from May 2026.) #oq/wait Partially answered: Nate Parrott's July 2026 post confirms "web search and MCP connections work in Claude Design too, whenever the design depends on outside information" — the client side of the claim is live. Whether specific design tools integrate through their own MCP servers is still unconfirmed.
• How does Claude Design's eval discipline work for visual/aesthetic output, where there's no compiler or test? (Same open question as Cowork for non-code artifacts; relates to character/taste evals.)

Sources#

• Designing with Claude: From prompt to production — Dan Carey, "Designing with Claude: From prompt to production," May 2026
• Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next — Boris Cherny's prior mention as a vision example
• How the product designer who built Claude Design uses it — Nate Parrott, July 2026 (case-study): builder's-eye origin, the Labs offsite, the scope reframe, stated boundaries, beta availability