At a glance
WHAT IT’S REALLY ABOUT
How Anthropic Labs built Claude Design fast through prototypes, feedback loops
- Claude Design is a new Anthropic Labs product that helps people collaborate with Claude to create polished visual artifacts like prototypes, slides, and one-pagers.
- Anthropic Labs operates as a “bet factory,” running many small experiments, doubling down on what shows promise, and shipping to users at very high cadence rather than predicting long-term futures.
- The team intentionally avoided traditional upfront processes (PRDs, vision docs, OKR cycles) and instead used concrete prototypes—often generated from recorded conversations and transcripts—to align quickly and iterate.
- Claude was embedded into the development loop to accelerate user research synthesis, feature design, shipping workflows (handoff to Claude Code), and feedback triage via clustering and suggested fixes.
- A key lesson was that speed doesn’t prevent mistakes, but short cycles make mistakes cheap—e.g., advanced “power user” controls were removed within a week after broader users found them confusing.
IDEAS WORTH REMEMBERING
5 ideasOptimize for learning speed, not perfect upfront plans.
The team ran a talk→design→ship→feedback loop 50–100 times in 10 weeks, treating product shaping as an empirical process rather than something to be decided in documents.
Prototypes are higher-fidelity alignment tools than PRDs.
Docs are easy to misinterpret, while a working prototype makes the experience concrete; their prototyping cycle dropped to minutes by using transcripts and Claude to generate options.
Keep teams tiny to reduce coordination overhead.
Claude Design was built by ~3 people for most of development, with everyone doing everything (engineers interviewing users, PMs coding, designers analyzing data), minimizing handoffs and meetings.
Instrument user feedback so “talking to users” becomes the easiest default action.
They created shared Slack channels with users and used Claude to find cross-user themes, ensuring insights weren’t lost across parallel conversations.
Build internal tooling immediately when it unblocks repeated loops.
They built a feedback clustering tool “in an afternoon,” then automated follow-ons (bug suspicion, suggested fixes, and one-click transfer into dev workflows) because these steps repeat constantly at high iteration cadence.
WORDS WORTH SAVING
5 quotesMy favorite thing to talk to users about is, "Please complain at me."
— Dan Carey
We do not try to predict the future. A lot of labs groups out there do try to predict the future. They say, "In ten years' time, we're gonna have this amazing technology, and it's gonna do these amazing things." We don't do that. Uh, instead, we try to ship. We watch people use the stuff. We learn what they do. We ship, we watch, we learn. We ship, we watch, we learn.
— Dan Carey
We started working on Claude Design because Claude Code made engineers really, really fast, and the rest of us had to keep up.
— Dan Carey
I do think it's worth calling out the things that we did not do. So we did not write a PRD in advance. We had zero vision docs. We had zero OKR meetings.
— Dan Carey
Before we close, I wanna share one of the more counterintuitive things we've learned working in labs, and that is that you do not wanna work on the thing that already works. You often wanna prototype the thing that almost works.
— Dan Carey
High quality AI-generated summary created from speaker-labeled transcript.
