Anthropic 首位技术 PM：Evals、Token Maxing 与不可替代的人类判断

A frontier model reaches users most powerfully when its new capability arrives together with a product vehicle designed to expose that capability.

Anthropic's early coding advantage grew from noticing users write long-form code and making targeted training changes before coding became an obvious category.

Anthropic Labs uses small pods to pull on promising threads before the user need, product shape, and technical path are fully defined.

Product teams must translate vague complaints into specific failure modes that researchers can reproduce, measure, and improve.

For measurable model behavior, an evaluation set can define the work more directly than a prose product requirements document.

PRDs remain valuable for ambiguous opportunities and for aligning engineering, product, legal, and safety around a shared direction.

Managers of AI product teams need hands-on shipping experience to develop judgment about what the technology can do and what good looks like.

Claude becomes a better thinking partner when it can recognize weak reasoning and push back instead of agreeing reflexively.

As AI contributes more to writing, accountability may shift from who drafted the text to who verified it and signed off on the result.

Hard-earned human judgment remains valuable because it compresses nuance, experience, persistence, and context that models have not fully lived through.