Anthropic’s first technical PM on token maxing, the jagged edge, and living in the future | Dianne Penn

When Diane Penn joined Anthropic in 2023, the company was still small: the product group had about five engineers, and a single engineer supported the whole API business. Penn says the team’s strong mission-driven culture and startup-like energy mattered from the beginning, as it explored how language-model technology might benefit users and society. The early product was a chatbot, but the organization was also investigating tools and other capabilities.

Penn describes Golden Gate Claude as a small, fast experiment that helped the company discover a distinctive product voice. After interpretability research surfaced a model feature associated with the Golden Gate Bridge, Anthropic turned it into a quirky Claude.ai experience within about 24 hours. When prompted for a spaghetti recipe, Claude would also comment on the bridge’s International Orange color. The experiment reached only about 2,000 people, by Penn’s estimate, but brought research to users through a product that felt different. She credits engineers, product, design, and research working together in a bottom-up effort, and sees the project as an early sign of the culture that continued at Anthropic.

Penn identifies Opus 3 as an important organizational milestone. At a company of fewer than 200 people, teams across inference, research, fine-tuning, and pre-training rallied over many months to build a frontier model and give users a reason to choose Claude. Working closely through that effort built trust between product and research; several research leaders who collaborated then later led areas including reinforcement learning and model alignment work.

Coding became an early differentiator after Penn noticed people using models to write substantial code, rather than only autocomplete. A relatively small training change to improve Opus 3 for that behavior helped attract early developers and Claude enthusiasts. Later, Opus 4.5 and Claude Code reinforced each other: the model enabled broader agentic coding tasks, while Claude Code gave people a product experience through which they could feel the model’s capabilities. Penn frames the inflection as the combination of a frontier model and a frontier product, not either one alone.