Jenny Wen, Design Lead at Anthropic, joined Double Diamond for a conversation about designing Claude, one of the most widely-used AI assistants in the world. From her early work on Dropbox Paper to bringing FigJam from concept to launch at Figma, Jenny has spent her career building collaboration tools that feel both powerful and deeply human. Today she’s applying that experience to designing interfaces for intelligence itself, leading design for Claude Cowork and helping shape how millions of people interact with AI.
At a glance…
Jenny describes designing AI products as fundamentally different from traditional software because you can’t map out all possible user flows when dealing with large language models. Where traditional products have deterministic states, AI products offer core primitives with essentially endless interaction possibilities. Her team discovered that knowledge workers are rapidly becoming more comfortable with AI-specific concepts like memory, context, and agents, leading to decisions about using technical terminology rather than abstracting it away. She demonstrates how she uses Claude extensively in her own design process, calling it especially effective at “garbage in, treasure out” analysis of large datasets from user interviews to social media feedback. The rapid advancement of AI coding capabilities has fundamentally changed how product teams operate, with single engineers now capable of shipping entire features end-to-end. As models become more capable at longer-running tasks, Anthropic has started hiding more technical details because users increasingly trust Claude enough not to need visibility into every step. Jenny emphasizes that while shipping code has become accessible to everyone through AI assistance, craft and the ability to execute well remain the most valuable designer skills.
The fundamental difference between designing traditional software and AI products
Jenny explains that designing AI products requires a completely different mental model than traditional software. “When you are designing now with these LLMs, you basically can’t map out all those flows. You basically just have these core primitives and these specific states that you want the model to get to, but the possibilities in which the user can actually interact with the model are just basically endless.” She compares this to her work at Figma, where despite having millions of possible design outcomes, the team still designed around specific use cases while remaining open to discovering new ones through user behavior. The challenge becomes designing the variables rather than controlling specific outcomes.
Building Claude Cowork through rapid iteration and dogfooding
The development of Claude Cowork reveals Anthropic’s approach to AI product development through continuous iteration and internal testing. Jenny describes their policy of putting experimental features in front of employees immediately: “Anything goes when it comes to dogfooding internally. We have the most janky stuff just running internally as long as it doesn’t disrupt people’s workflows.” The team experimented with multiple technical approaches and agent architectures before recognizing product-market fit when users started applying Claude Code to non-coding tasks. What became the “10-day ship” was actually the culmination of months of prototyping and internal feedback, with the final decision to ship happening when they saw clear user demand.
Knowledge workers are rapidly adopting AI-native concepts
Jenny has observed a significant shift in her target users’ AI literacy over just the past year. “Even like a year ago I was uncertain about whether we should reveal words like memory and context and agents or should we abstract them in a way where they’re sort of fluffy. But now I think people know those words.” She explains that knowledge workers at AI-adopting organizations have become much more comfortable with technical AI concepts than expected. Rather than simplifying terminology, Anthropic increasingly uses precise technical language because their users understand it. This evolution in user sophistication has influenced design decisions throughout Claude’s interface.
[Demo] Using Claude Cowork to synthesize user feedback and generate design concepts
Jenny demonstrates her actual design process using Claude to research and iterate on Claude itself. She shows how she can point Claude at folders of user interview transcripts or ask it to search social media for product feedback, then synthesize insights into actionable product improvements. “It’s really good at this thing that I call garbage in treasure out where you can just throw it a stack of papers essentially and then pull useful insights out of that.” She then uses Claude to generate initial wireframes and ask clarifying questions, emphasizing that she’s not having Claude make final designs but using it to overcome the blank page problem and generate starting points for further iteration in Figma.
[Demo] Trust, transparency, and the evolution of showing users what AI is doing
Anthropic has gradually shifted toward showing users less of the technical details as trust in Claude has grown. “When we first started, we actually showed all of this code and underlying stuff that Claude is doing, but over time we’ve actually started to obscure it more and hide it.” Jenny explains this happens because people are becoming more trusting and the models are getting better simultaneously. However, for longer-running tasks, users still want visibility into what the AI is planning to do, leading to the plan-approval interface in Cowork. The team continuously calibrates how much detail to show based on task complexity, duration, and user trust levels.
[Demo] How model improvements make design decisions obsolete
Jenny describes the unique challenge of designing for rapidly evolving AI capabilities. “We’ll do things where we’re like ‘we actually might not need this selector or something because you know in a month from now Claude will be able to just choose whether it wants blue or red.’” Her team regularly receives snapshots of new model capabilities and must constantly evaluate which interface elements will become unnecessary as the underlying AI improves. The key skill becomes understanding model trajectory and designing accordingly rather than building for current limitations. This requires designers to think in terms of what will be automated away versus what needs human oversight.
The collapse of traditional product roles and the rise of single-engineer, single-designer features
Jenny describes a fundamental shift in how product teams operate when engineers can build entire features independently using AI. “For a given sub-feature, I’m actually just working with one engineer as opposed to five to 10. We launched scheduled tasks last week and that was just me and one engineer.” This new dynamic often eliminates the need for traditional project management, freeing PMs to work on higher-leverage activities like enterprise partnerships while designers focus on polish and user experience refinement. The process becomes much more iterative and collaborative, with engineers implementing first passes that designers then refine based on user feedback.
The future of human-AI interaction beyond chat
While Jenny believes chat will remain important as the most flexible interaction paradigm, she expects to see more dynamic UI generation. “I think what we’re already seeing is that the models are getting better at generating UI on the fly. In moments where UI is much more helpful and faster and direct, there should be UIs.” She envisions a future where AI can conjure specialized interfaces when needed while maintaining the fallback of natural language conversation for anything not covered by the generated UI. The challenge becomes training models to understand when UI is more effective than conversation and ensuring generated interfaces follow familiar patterns.
What designers should focus on learning in an AI-native world
Jenny argues that shipping code has become accessible to anyone who can chat with an AI, making it less of a specialized skill. “I think getting to ship to production is not that hard anymore. If you can chat with an LLM you can ship to production now.” Instead, she emphasizes execution and craft as the most valuable capabilities: “It’s just the ability to execute and make something good. I think shipping is a skill and right now, in the way that we work, people either have it or they don’t.” She believes designers need to develop speed in getting good work into production rather than focusing on learning technical implementation details that AI can now handle.
---
Thanks for reading. Stay in the loop on new episodes and upcoming events by subscribing.
