更少的 PM、更多的重复次数：Whatnot CPO 为何让资深产品人回去干活

Tom Verrilli, chief product officer at Whatnot and formerly at Twitch and Twitter, says Whatnot's product team was founded on the premise that "we regret that product management exists." He traces the logic back to history: early on, founders and CEOs talked directly to engineers and designers about what to build. As internet businesses scaled faster than any before them, delegating those decisions to a dedicated class of people became necessary.

Verrilli's argument is that PM is a trade rather than a qualification. It is something you get good at by doing, and as every personal trainer says, muscles are built by reps. The flip side is that the more you abstract engineers and designers away from decision-making, the weaker their own muscles become. Ideally they would be so close to users that they could make the same calls a PM would.

He also blames the ratios that crept in as companies grew. Somewhere along the way a "pod" formula appeared in which every six engineers earn a designer, a PM and an engineering manager, so a billion-user product ends up with a great many PMs. In his view you probably do not need a PM for notification infrastructure, and hiring that many of them "infantilizes" capable engineers and designers who never had to make decisions.

Lenny Rachitsky points out that engineers and designers often do not want the less glamorous parts of PM work, such as alignment meetings and docs. Verrilli agrees that specialization works both ways, and an infrastructure lead may reasonably not want to spend time debating minutiae. The regret statement, he says, is a reminder that you hire a PM against a really specific need, not for the sake of having one.

That requires a matching culture. Anyone at Whatnot can be the directly responsible individual for new product work, including an engineer or designer, but everyone goes through the same product review and does the same real work. Because of this, PMs are mapped to problems or core projects rather than to engineering teams, which means a team can go a year without a PM even though plenty of product work is going on.

Verrilli says he has also learned that you no longer need to scale layers of management to get leverage. It is easier than ever to talk directly to the code base, and AI gives individual contributors enormous reach. He notes that top-down is not a dirty word either: senior people with real ground truth can make quick decisions and remove alignment debate, so one very senior PM can cover more than three entry-level ones.

Whatnot currently has about 21 or 22 PMs, loosely grouped into buyer, seller, and trust and risk. Every six months the CEO, Verrilli and senior leads define what must be true over the next half, assign a DRI to each priority, and hand any orphaned item to a PM. Assignments are usually outcomes, such as working out what needs to be true, not a spec for a particular feature.

Over two years, 31,832 people applied to be a PM at Whatnot, and the company hired one. Verrilli says what is trending down is candidates who spend interviews talking about alignment meetings and stakeholder management. He admits he used to be one of those PMs, whose specialty was politics rather than technical or customer insight, and he treats stories about failing to keep the CEO updated as an anti-pattern.

What jumps out in an interview is a candidate who shows both macro and micro thinking: describes the system and the end state, then shows impatience about how to validate it quickly. He is wary of people from large companies who babysat existing products and polished their edges, rather than owning a novel problem and iterating through a hard change. In large organizations it is easy to go along with inertia instead of being an agent of change.

To build systems thinking, Verrilli recommends mental exercises. In reviews he asks what the team will do if an experiment comes back green, and what if it comes back red. He also asks them to imagine a thousand times more usage than expected and to trace the knock-on effects before touching code. The internal motto is "know then go": think through everything that could go wrong at scale, then move anyway, because you will pre-empt many problems.

Later he adds that every hire in any role does a hands-on case study. Many candidates present beautifully, but when given a prompt and data and made to defend a PRD verbally, their thinking decays quickly. He calls that the difference between people good at theater and people good at specifics.

Verrilli's scar tissue comes from a system where a successful PM was promoted into a director role and told to stop being hands-on. "We took all of our A players and then promoted them out of doing things," he says, leaving them to spend their time in alignment and a yo-yo of reviews. At Whatnot the four or five managers spend over 90 percent of their time on IC work, and he does about half.

The benefit is speed and coherence. A VP with a decade or more of experience has well-honed instincts and can decide quickly, while seeing more of the board than a junior PM. He cites Twitch, where discovery and ads were always fighting over feed impressions; once one PM owned both, the trade-off was made organically. Putting one person across multiple things removes months of politics that turn a company-first decision into a career-first one.

For him, IC work means whatever it takes to ship: reading support tickets, pulling the data yourself, querying the code base, writing the spec and running the stand-up. He has shipped some production code but says it is not the best use of his time. On pay, he argues the arithmetic favors the model, since the compensation of five junior PMs, four senior ones and a VP could fund three people on top-level pay.

Staying in the weeds improves the system too. In a growth meeting someone says a spike was fraud; Verrilli asks how it was labeled, who labels it and by what procedure. A good director will go and find out, which strengthens the agents and processes used for data labeling. He calls this the discipline of doing fewer things and going all in on them, rather than letting growth hide all sins.

For PMs struggling in a tough market, Verrilli's advice is to start doing IC work in the role you are in and keep your muscles honed. He also wants people to get back to scoping the right thing, understanding the problem and defining what good looks like, and to coach juniors toward an answer instead of endless review rounds. On the idea that PMs should ship code, he is cautious: learning the code base is valuable, but git and the minutiae take far longer than describing the problem.

The top AI gain, he says, is data science. Using Hex threads at Whatnot, a PM can pull nuanced cohorts, inspect a single user's logs, and build sensitivity models or forecasts quickly. He has spent less time talking to data scientists in the past year than ever, while spending ten times more time in data. AI tooling also spots regressions and odd interactions between products faster, which helps shipping.

The second gain is asking Claude how the code base works instead of bothering engineers; he says there is now no excuse not to understand the system's boxes and lines. The third is specific to live products: while watching a user struggle, you can inspect the code base in real time to tell whether it is a bug or a comprehension gap. He calls it a feedback loop on steroids, and tells a story of engineers changing an auction timer config from seven to three seconds mid-show at a seller's request.

Lenny relays a data scientist friend's complaint that their job has become reviewing half-right analysis from non-specialists. Verrilli sympathizes, and sees it as another argument for fewer, more senior PMs with more reps. He adds that organizations have long underinvested in data engineering, structures and labeling, and Whatnot's best data scientists are now pushing on tracking and attribution so that misunderstanding is harder. Using an AI tool does not remove responsibility for good analysis or code.

Fewer people for the same output, he notes, does not mean fewer people overall: well-used systems let you grow faster and take on more. He expects roles like a light-weight engineering manager or tech lead running a small team to trend up, because the cost of trying hard things has fallen and small incubation teams can take swings at problems once thought too difficult.

On the future team shape, he thinks Whatnot will still use specialist designers, engineers and PMs for high-conviction projects, but leave more free space around the edges. Anyone, including a PM on a Friday afternoon who cannot focus on a PRD, can fix something if they know the customer problems and the code base. Asked about Lenny's post arguing PMs are best positioned for AI, he agrees, but says those core skills are not what PMs were rewarded for over the last five years, when storytelling and framework communication won.

Verrilli says nuance is required, but "hire people and get out of the way" is wrong. The better the hire, the more you can trust them, yet Whatnot lives in a "verify then trust" mode. He is better placed than his directs to see how buyer, seller and trust fit together, and they are better placed to know the individual features. Because he approves what they build, he has to learn the details too.

He describes founder-CEO Grant sitting in a review, saying a proposal does not seem right, then clearing the rest of his day to pull up tickets, code and data line by line with the team. That sets a culture of seeking truth, rather than a review culture of PMs trying to extract a green light to gain credibility. Verrilli tries to be T-shaped, going deep when required and broad otherwise, because you cannot make good macro decisions without the micro.

On working with product-minded founders, he has worked for several in a row and now has two. He avoids doubling up: if founders Grant or Logan are on something, he checks that they are accountable and steps out, to avoid what one PM called "the two dads problem." He also reminds product leaders in founder-led companies that it is not their company, and that their job is translating the founder's vision into reality, not competing with it.

When he disagrees with an executive, he starts from curiosity and asks whether the person has context he lacks. Then he asks whether their mind is made up or whether they are open to input. If nobody has data and it is just two opinions, the CEO's opinion wins and that is fine; if you have data, bring it. Top-down works, he adds, only when leaders are in the weeds and specifically correct.

Colleagues told Lenny that Verrilli uses the phrase "play the accordion." He explains it as a mental model, not a framework. Code lets you ship and iterate quickly, but there are two failure modes: throwing spaghetti at the wall without direction, or writing long multi-year roadmaps that waste the advantage tech has over other industries, which is learning. A/B tests mean you can update your understanding and change direction as you go.

In the analogy, before you can play a note you must pull the accordion all the way out and take in air, which means working out what you are trying to get done. Music happens only when you press the keys and push it back in, shipping something. Then you pull out again to ask what you just learned and what to do next. Verrilli says pulling out creates no value on its own, but without doing it constantly you are probably not playing the right thing.

His example is live commerce listings. A seller can simply hold up AirPods on stream, so listings seem unnecessary. Zooming out, new buyers will expect search to work, and search needs to know what is for sale. Forcing every seller to make listings, at about three and a half minutes each, would sharply cut how much they can sell per hour, so that fails too. He says you keep stretching through the longer-term implications.

On agentic commerce, Verrilli says he welcomes agents for programmatic purchases like light bulbs, and for high-intent searches such as a cable or black shoes by Thursday. But he notes that e-commerce, after 30 years, has never exceeded 20 percent of US retail spend, and much shopping is low intent. People wander the mall because a store owner has taste and curation. Retail is a seven-and-a-half trillion dollar industry in the US, so he does not see a winner-takes-all outcome.

He sees live commerce as the first format to combine the scale of the internet with the social experience of physical shopping. At Twitch, streams under about a thousand viewers were considered non-economic because of CPMs, but commerce has different economics: a stream with 30 to 50 people is like a store you would never close. Whatnot, he says, serves a different customer need from agents. He closes with the story of buying a live California spiny lobster from a San Diego seller's dockside stream.

From Twitter in the 2015–16 era he drew two lessons. If you truly find product-market fit, a level of fervor you can feel, it does not matter how badly you screw up the organization. And when people say something is complex, often leadership is just weak: everyone knew the 140-character limit would be lifted, but no one made the call for almost two years after he left. Editing tweets took even longer.

His own failure theme is trusting averages. A feature used by only 3 percent of users may be 100 percent of what a particular group does, and deprecating it for convenience can blow up their use case, and in e-commerce, their business. He compares it to a mall turning off the power in the run-up to Christmas, and agrees with Lenny's Bezos paraphrase that when you have data and an anecdote, you trust the anecdote.