Cat Wu

Cat Wu leads product work on Claude Code, collaborating with its creator, Boris Cherny, on feature priorities, testing, and rollout. Anthropic listed her as Head of Product, Claude Code at its May 6, 2026 developer conference. Her remit concerns the product rather than oversight of the underlying models; her strategy rests on keeping the product adaptable as those models improve.[2][4]

In a May 2025 Latent Space interview, Wu described her approach as “I think I PM with a pretty light touch.” Engineers supplied much of the roadmap by building features they wanted themselves; she cleared obstacles involving legal, marketing, and other functions. Longer-term planning meant asking what models might soon accomplish: finding relevant information, completing every part of a complex task, and composing tools effectively.[1]

Faster implementation did not, in Wu’s account, lower the bar for new features. They still needed intuitive interactions and a clear fit with the product vision. Working prototypes helped the team judge abstractions and interaction patterns sooner than documents alone. In her 2026 essay on product management, she advocated short experiments, demos, and evaluations. She credited Noah’s plugins prototype with helping validate its user experience, and Conner’s agent-team evaluations with identifying where coordinated Claude Code instances worked and where they needed improvement.[1][3]

Wu used a recurring task to observe changing capabilities: beginning in October 2024, she asked each new model to add a table tool to Excalidraw through Claude Code. She recalled successive failures, then occasional success with Opus 4 sufficient for a prerecorded Claude 4 launch demonstration. By Opus 4.6, she described enough reliability to attempt feature requests live before professional developers. This was her practical test of one feature-building task, rather than a general benchmark.[3]

Wu favors a lean harness: the tools and instructions around the model should help it act without unnecessarily steering it. In a May 2026 Ars Technica interview, she said the team’s evaluations had not shown a measurable improvement from semantic-navigation plugins, so these remained optional rather than core tools. Her essay supplied a concrete example of removing scaffolding: early todo lists needed periodic reminders to mark completed items; when a newer model handled that behavior, the team removed the reminders.[4][3]

Her advice for prototypes is to establish capability before optimizing token cost. Cutting usage too early, she argued, can prevent a team from discovering whether a feature is possible at all. That willingness to experiment also required a personal adjustment: describing herself as a perfectionist, Wu said product managers needed to identify a handful of true non-negotiables and relinquish some control over the rest.[3]

In May 2025, Wu emphasized developers’ control over which actions Claude Code could take, crediting teammate Robert with leading permission-system work. She also described working closely with legal on web fetching, then limited to user-provided URLs or links mentioned in previously fetched material. For non-interactive automation, her recommendation was incremental: try one test, refine the prompt, expand to ten, and examine failures before attempting a large run.[1]

Wu argued that the developer merging AI-generated code remains responsible for its maintenance, documentation, and abstractions. She wanted productivity assessment to capture reduced feature cycle time—such as the interval between first commit and merge—and useful work that otherwise would remain unbuilt. She described customer-reported bugs receiving quick fixes that engineers said would otherwise have stayed in the backlog; these were examples motivating further measurement, rather than established productivity gains.[1]

Usage transparency became another product concern. Wu told Ars that the team debugged complaints live with users and found expensive resumptions after session caches expired, alongside plugins that launched many subagents without users realizing it. She described adding cache warnings and usage information to expose these patterns, while acknowledging that clearer explanations required understanding where the tokens were going.[4]

Wu’s involvement with Claude Code began as an internal user providing feedback. Her earlier engineering and research-product work supplies context for that hands-on role.[1][3]

Wu began her career as a product engineer at Scale AI and Dagster, then became a venture capitalist. In her retrospective account, she continued writing automation to discover new companies and detect growing interest in open-source projects. She joined Anthropic in August 2024 as a product manager on the Research PM team, which connects researchers with real-world customers.[3]

When Claude Code became available internally that fall, Wu used it to build Streamlit applications for analyzing large-scale user feedback, run evaluations to explore useful benchmarks, and create reinforcement-learning environments to understand training better. She recalled spending hundreds of hours prompting these projects without writing their code by hand.[3]

Her visualization work generated extensive feedback for Cherny. Wu credited Boris, Sid, and Ben as Claude Code’s original team; Cherny separately recalled asking for her as product manager while the team was recruiting. She thus joined the product after using it in her own work, rather than founding its original engineering effort.[1]