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

How Anthropic's product team turns model research into products

Anthropic's product work runs on a tight loop between models, products and evals. Claude Code and Opus 4.5 lifted each other, and Labs places discontinuous bets that may sit off the core roadmap. Product managers make research usable by turning vague feedback into specific evals, which also matters for safety when capabilities arrive suddenly.

Takeaways

• Model and product improve each other. Penn says Opus 4.5 would not have had its breakout moment without a product like Claude Code, and Claude Code would not have grown as fast without Opus 4.5. Opus 3 had earlier given Anthropic its edge in coding.
• Judge token spend by experimentation, not volume. Penn treats tokens as the input and experimentation as the real output. Teams should sweat tokens as much as pixels, and work in public, as Anthropic did when nearly the whole company tested early Claude versions in one Slack channel.
• Labs makes discontinuous bets off the core roadmap. Labs has incubated Claude Code, skills, Claude Design and MCP. Ben Mann sets its vision, pods stay small, and a thesis that is not working yet gets revisited after one or two model generations.
• Product managers turn vague feedback into evals. About 80% of early instruction-following complaints meant Claude would not write the right JSON. Thirty to forty examples became a standing test that now passes at 100% or 99.9%. Evals are becoming the new PRDs, though Anthropic still writes a PRD for every model.
• Sudden capabilities make safety harder without evals. Penn argues that emerging capabilities that appear suddenly are harder to make safe unless teams have evals. She sees product and user overhang on current models, and Anthropic plans to keep improving its model safeguards in the coming weeks and months.
• AI should push back, humans should sign off. A thinking partner that simply agrees does not help, and the goal is better ideas rather than ideas 10% better. In writing, who verifies and signs off matters more than who writes, and AI writing lags because capability is jagged-edged.
• Hiring favors first-principles, low-ego, hands-on people. Penn's team has not changed its hiring loop in three years. It looks for first-principles thinkers and low-ego team players, and wants managers who ship hands-on. She expects tone and character, proactivity, judgment and persistence to stay valuable.

How Penn works with AI herself

• Use a Claude skill built on the book Crucial Conversations to prepare for hard talks with her team.
• Have Claude draft monthly business reviews while she reviews them.