Anthropic's product chief coaches her managers with Claude: she wants an AI that pushes back, not one that agrees

By Zendoric · July 28, 2026.

Dianne Penn, head of product management for research and labs at Anthropic since 2023, has said on an episode of Lenny's Podcast released this weekend that she uses Claude, the company's flagship model, as a coaching tool to manage her team. As she told it, she has created a custom Claude "skill" —a saved function the assistant can invoke on demand for a recurring task— inspired by the book "Crucial Conversations: Tools for Talking When Stakes Are High", which helps her prepare difficult conversations at work and calibrate "the right level of detail" before having them.

Penn describes the use as a kind of one-on-one coaching: it helps her brainstorm how the other person is going to react, build trust faster or be more direct without losing nuance. But she sets an explicit limit: she does not let Claude think for her. She first forms her own point of view and only then turns to the model to polish it, because —she says— "what you don't want is an AI that just agrees with you"; what she is looking for is a "thought partner" that disagrees when necessary, just as she would demand of a human colleague.

That phrase points to a problem the AI industry itself has spent more than a year trying to solve: sycophancy, the tendency of models trained with human reinforcement to tell the user what they want to hear rather than what they need to hear. In general, the most visible case was that of OpenAI, which in 2025 had to withdraw a GPT-4o update for being excessively flattering after user complaints. Anthropic has made Claude's "personality" —including its willingness to push back— a declared training priority, and Penn's testimony is, at bottom, the internal-facing version of that same product argument.

Our reading is that the value of this anecdote lies not in whether Claude is a good coach —Penn is an Anthropic executive talking about Anthropic's tool, and it should be read with that caveat— but in where AI is making inroads without our noticing: people management, a terrain we assumed was protected precisely because it demands judgment and human relationships, two capabilities that our sector analysis on employment has identified as the most resistant to automation. What Penn describes does not replace that judgment, it trains it: Claude does not decide how to deliver bad news, but it helps rehearse it beforehand.

That distinction matters for the short, hard transition we anticipate: if AI reduces administrative and writing work, what remains as a good manager's differentiator is precisely the difficult conversation, the one that cannot be automated. That the tools are beginning to serve as a gym for that skill —instead of replacing it— is a signal in favor of our underlying thesis: an AI that frees up time can also be used to improve the parts of work that remain irreducibly human, provided that whoever uses it, as Penn recommends, keeps their own judgment as the starting point and not as an afterthought.