How PM work functions in Anthropic's research: who they hire, how they onboard them and how they retain them

This analysis is based on the text of an episode of The Skip podcast, hosted by Nikhyl Singhal, in which he interviews Dianne Na Penn, head of product for Anthropic's research and Labs teams. The downloaded article is a written summary of the episode (which runs about 59 minutes). It is not the full transcript, so what follows reflects only what that summary covers.

According to the text, Penn joined Anthropic in 2023 as the company's first technical product manager, before the Claude app existed. Her team has managed every model as a product from Claude 2 through the Claude 5.5 family, and has incubated Claude Code, MCP and Claude Design. The host argues that the way PMs work at a frontier lab is a preview of what the role will look like at other companies.

The first two months should be hard. Every new hire is told in their onboarding plan that those months should feel like some of the most difficult of their career, and that this means they are doing well. The goal is not to see who breaks, but to show quickly what that pace demands and to build habits. The new PM is given a real area, usually a high priority, but with no scaffolding or chain of approvals. The answers are out there: most channels are public, and Penn advises going back two months and reading what happened, so as to arrive with one's own opinion rather than asking everyone to stop and bring them up to speed. There is an “onboarding buddy,” but what matters most is the initiative to figure things out. Those who adapt fastest embrace this; those who struggle come with reasons why they couldn't do something, often waiting for a structure that will never arrive. She can tell which way a person is heading within the first or second month. Outside the labs, according to the author, the “listening tour” is getting shorter and impact is expected within a few weeks.

Newcomers are expected to push back. The text notes that PMs from large companies struggle with hierarchical deference: if the vice president says X, X gets done. A research lab is more like academia, where a first-year PhD student can challenge their advisor. Those who succeed debate with sound judgment: they bring important user feedback to researchers and leaders, flag gaps and question assumptions. There is a right way to do it. A newcomer's first ideas were probably considered two or three years ago, so the first step is to read what was tried and why it didn't pan out. Then they press on with curiosity while building trust. Penn rejects the “quick take” along the lines of “you're wrong” that doesn't engage with the details, and values those who have thought two or three steps ahead and persist if their first idea doesn't land. Her phrase for someone who does this well is that they “join the discourse.” The text adds that the culture makes this easier: most channels are public and it works on a “yes, and…” logic, where someone you didn't know was following the conversation chimes in with a “have you thought about X?”, kindly.

As for whom they hire, her favorite profile is a startup's first PM who was also its first customer success lead, did business development and perhaps IT support. Founders and people who have achieved something difficult inside a large company, where inertia was the obstacle, are also a good fit. According to the text, what is valued is that track record of overcoming inertia, not a ladder of ever-bigger titles. That is why many career paths on her team have been, in her words, “meandering” rather than climbing one level after another in the same role. Her advice for interviewees is to talk about what they achieved when the company was the obstacle, not the scope they were given.

Planning is done month by month. In its early years Anthropic ran a full annual OKR process, which was soon dropped because nearly every assumption about usage and adoption turned out to fall short. No one wanted to spend three months on an annual plan that becomes obsolete in two weeks. The team now works with a short set of assumptions about where the world will be in three to six months and a one-month plan to make progress. The long-term vision comes from the research organization and the mission, which avoids reacting to whatever a competitor launched last. The current example is life sciences, with the stated ambition of compressing decades of biological progress into a few years; that is not a quarterly goal, but giving scientists access to the models and deciding which training experiments to run is monthly work.

Before committing to a project, Penn asks what the models will be able to do in three and six months. Her example is a fine-tuning API: the team debated offering one, but by the time a customer learned to post-train and deploy a model, the base models would have advanced two or three generations and they would have to start over. The underlying habit is to stay “stubborn” about where the models are heading, even when a reasonable request pulls in another direction. The lesson the author proposes: write down what you are stubborn about for the next six to twelve months and treat everything else as a plan that gets rewritten every month.

Finally, the text highlights two groups that perform especially well. The first is early-career PMs, described as highly effective and with a high ceiling, who have no processes to unlearn: they don't wait for approval for the next step, they decide they own a capability, talk to customers and to whichever researcher they need, and move forward. The second is forward-deployed engineers who moved into product; they know the users and the systems they use, and bring that knowledge to the research team. Penn did something similar in 2023, when she wrote the system prompt for claude.ai because there was no one else to do it. The author sees this as a kind of modern APM program that starts with time spent alongside customers. His advice: if you want to get into product and there is no opening, take a customer-facing role at an interesting company and make the jump from there; and if you are already senior, be prepared to unlearn some of what you know.

The text mentions that the full episode covers further topics that this summary does not develop: the bootcamp the team runs for all of Anthropic's PMs, how it decides whether a problem calls for changing the model or fixing the product, her shift from strategy to the details of execution, and her document on how to sustain this pace without burning out. Since the only material available is this written summary, no further details on those parts can be provided.