Shipping in Days: How Claude Code's Head of Product Runs a Team at AI Speed

Wu describes her role on the Claude Code team as complementary to Boris Cherny, the tech lead who created the product. Cherny is the product visionary who defines what the "AGI-pilled" version of the product should look like three to six months out. Wu's job is to map the path from today to that vision, make sure marketing, sales, finance and capacity planning are aligned, and ensure nothing blocks a feature once it is ready; the line between them is remarkably blurry.

Wu says she interviews hundreds of PMs who want to work in AI, and many approach the job the old way. Before AI, technology shifts were slow enough to plan on six-to-twelve-month horizons, and code was expensive, so PMs invested heavily in coordinating with partner teams. Now, she says, feature timelines at Anthropic have fallen from six months to one month and sometimes to a week or a day.

The implication is less emphasis on aligning multi-quarter roadmaps and more on finding the fastest route to put something in users' hands. Wu wants a corner of the product where an engineer's or PM's idea can reach users by the end of the week. The PMs who do best on AI-native products, she says, are those who shorten the time from idea to user and who define which tasks must work out of the box.

Wu lists three things PMs can do to help a team move fast. First, set clear goals, because LLMs are so general that they create ambiguity about who the product is for. Her example is a goal of letting professional developers at enterprises safely reach zero permission prompts, which rules out many approaches. Second, build a repeatable shipping process, and at Claude Code that means launching almost every feature as a clearly branded research preview.

The third element is a framework for pulling in partners. When an engineer has a dogfooded feature, they post it in an evergreen launch room, and the docs, product marketing and developer relations leads can turn around an announcement the next day. Instead of long PRDs, the team runs weekly metrics readouts and keeps a list of team principles covering key users and trade-offs, so people can decide without waiting on a PM. Wu says PRDs still exist for particularly ambiguous or infrastructure-heavy projects.

Asked whether the powerful Mythos model explains the pace, Wu says it helps a little but not the bulk of it. The team has been moving quickly for several quarters, and she attributes most of it to low process and an expectation that anyone can take an idea into the world in under a week.

Lenny asks about the recent leak of Claude Code's source. Wu says the team looked into it immediately and found human error: someone working with Claude on a PR that changed how packages are released, and it passed two layers of human review. She calls it a process failure, says the person is still at Anthropic, and notes that the team has hardened its processes with most of the safeguards already shipped.

On the move to stop people using Claude subscriptions with OpenClaw, Wu explains that demand for Claude has been very high and the company is working to scale infrastructure and make its harness more token-efficient. Subscriptions were not designed for third-party products, which have different usage patterns. Anthropic chose to prioritize first-party products and the API, while giving everyone some credits alongside their subscription to ease the transition.

Wu says Anthropic has roughly 30 to 40 PMs across several teams: research PMs who channel customer feedback on models to research, the Claude developer platform (which includes managed agents), Claude Code and Cowork, enterprise, and growth. On whether the PM role will shrink, she says all roles are merging: PMs do some engineering, engineers do PM work, and designers land code. Her team chooses to hire engineers with great product taste, some of whom can go from a tweet of user feedback to a shipped product within a week with almost no PM involvement.

Nearly all of her PMs have been engineers or have shipped code on Claude Code, which she says builds trust and speed. When asked which background matters most, she returns to taste: as code becomes cheaper to write, deciding what to write becomes more valuable. With tens of thousands of GitHub issues asking for everything, it takes care and taste to choose what is worth building.

An engineering background helps for now because it gives a sense of how hard something should be, which feeds prioritization. But Wu cautions that the valuable skill set changes every few months with jumps in coding capability. The durable ability is first-principles thinking, spotting the gaps the team has, and being low-ego enough to wear many hats.

Asked where human brains remain useful, Wu points to the common sense and EQ needed for a launch, which has a thousand moving pieces. Models do not always know who the stakeholders are, how they relate, or the right venues to keep them on board. She expects models to improve at this but sees gaps today.

On staying sane amid the chaos, she says the team hires people who lean into it, face challenges with a smile, and accept they will not be perfect. She jokes that a Sunday-night P0 becomes a P00 by Monday and a P000 by afternoon. The approach is to sleep well, prioritize brutally, and accept that some products ship less polished than she would like, because quick feedback lets the team fix them in the next release.

The main thing sacrificed is product consistency. Overlapping features exist because the team wants users to reveal which of two form factors is better, so new users may not know the best path to a task, and existing users feel pressure to check Twitter daily. That is why Anthropic added a `/powerup` command, departing from its earlier "no onboarding flow" principle, to walk people through the 10 features they most need among 100.

Lenny notes that Anthropic began far behind, with less funding and distribution, and is now growing extremely fast. Wu names two ingredients. The first is a unifying mission: the company hires people who care about bringing safe AGI to humanity, and when priorities compete, they ask which better serves that mission. That lets the company make fast decisions that cut across the org and have everyone stand behind them.

The second ingredient, in her view, is focus. She describes a culture where teams will sacrifice their own goals and KRs for Anthropic's goals, and says she would be extremely happy if Claude Code failed but Anthropic succeeded. On the OpenClaw decision, she frames it as doubling down on first-party subscriptions, since growing the number of users reached matters most, even at the expense of third-party products.

Wu gives a mental map of the products. She uses the Claude Code CLI for one-off coding tasks because features land there first and it is the most powerful surface. Desktop is better for front-end work with a live preview pane, for people uncomfortable in a terminal, and as a control plane showing CLI, desktop, web and mobile sessions. Web and mobile let you kick off tasks without tethering a laptop to your phone.

Cowork covers work whose output is not code, such as clearing Slack or inbox, building slide decks, or writing docs. Her advice is to first connect the data sources relevant to your role, such as Google Calendar, Slack, Gmail and Google Drive, since quality depends on context. She then walked through a recent example: for a Code with Claude talk, she gave Cowork a draft from her product marketer plus her own rejected draft and asked for a proposed outline.

After she chose what belonged in the final talk, Cowork worked for hours, searched Twitter, the launch room and the demo channel, and produced a 20-page deck she found pretty good. She needed one round of feedback because it was too wordy, and it matched Anthropic's design system because she gave it the standard external deck template. Wu sees this as the PM's enduring role: Claude brainstorms and synthesizes, but the human makes the final decisions.

Wu says her stack is mostly Claude Code, Cowork and Slack, which she calls the core OS of Anthropic, and she spends about 30% of her time pushing the limits of the tools to learn what they are bad at. Claude Code has lowered the barrier to building custom apps, so people build bespoke software instead of using tools that do not fit. One salesperson built a web app that pulls customer context from Salesforce, Gong and notes to tailor Claude Code decks to each customer, including cloud platform, security needs and the right feature slides, in seconds rather than 20-30 minutes.

Beyond engineering, she names Applied AI, a very technical go-to-market team that helps customers adopt the API, as a heavy user of Cowork and Claude Code. They prototype for customers and, the night before meetings, ask Cowork to build a dossier of action items and open asks, and to research answers in Slack. Wu says token cost per engineer or knowledge worker rises after each model jump but remains well below salary, and people are trusted to use tokens responsibly even though wasting them is frowned upon.

The skill Wu considers hardest is defining what the product should look like a month from now, given uncertainty about model capabilities and user behavior. She says it is very hard to be the right amount of AGI-pilled. It is easy to design for a hypothetical super-strong model that needs only a text box; the harder task is eliciting the most from the current one, guiding users onto the golden path and patching weaknesses.

She names three practices. One is to ask the model to introspect on unexpected behavior, for example why it ran tests but never used the UI, which can reveal confusing prompts or an unchecked sub-agent. Another is to find the handful of users who articulate model quality accurately. The third is to build evals: even ten great ones help quantify the goal and progress, and a small pod works with research on Claude Code behaviors.

Wu singles out Amanda, who shapes Claude's character, and the Claude team's lunchtime vibe checks on new models as trusted feedback. She says personality is central to Claude's success: light-hearted but confident, low-ego, positive and willing to give earnest feedback rather than agree with everything. Those traits, she argues, make a great coworker.

Wu says many changes with each new model are removals of features that were crutches. The classic case is the to-do list: early Claude Code would change five of twenty call sites and stop, so a teammate gave it a list tool to track them. With Opus 4 and later models it uses the list on its own, and today it is a de-emphasized part of the product. The team rereads its system prompt at every model launch and removes reminders the model no longer needs.

The more exciting result is features that become viable. Code review is her example: after several attempts, recent models made it reliable enough that engineers depend on it before merging, with multiple review agents traversing the codebase and synthesizing real issues. That is why Wu says it is important to build products that do not fully work yet and swap in the newest model to see whether it closes the gap.

On the long-term vision, she describes building blocks. The task is the core unit; as models improve, people run several at once, and Wu expects dozens or hundreds, likely remotely. That raises needs for interfaces that show which tasks need attention, agents that fully verify their work, and a self-improving loop in which feedback prevents repeat mistakes.

Wu's advice for thriving through the transition is to notice manual tasks you repeat and hand them to Claude Code or Cowork. AI can take over the tedious parts of a job so you can focus on the creative parts and on the pet projects nobody had bandwidth for. She also urges people to iterate until an automation works every time, because an automation that works 95% of the time is not really an automation.

Both she and Lenny admit to giving up on automations too early, such as inbox triage that is mostly right. Wu says her team is working to simplify customization, since users currently need to know about skills, feedback and updating them. She also recommends building apps you actually use daily and warns against the opposite trap of endlessly tuning setups of skills and MCPs instead of shipping, saying simple setups often work better.

Discussing the divide between people who dismissed early chatbots and those who code with AI, she says the shift is from chat-based products of 2024 to action-based ones. The big aha moment is when an agent does things for you instead of only telling you what to do.

Asked for a life motto, Wu answers "Just do things." She values first-principles thinking: know what you are optimizing for, deduce the right action, articulate it to stakeholders, and do it quickly, learning from mistakes and apologizing or fixing them if needed. She believes it is liberating to tell people this, since many companies define roles and code-ownership boundaries rigidly.

She credits an early job at a 20-person company, with no process and big problems, with teaching her to work without boundaries. Lenny adds that many people must unlearn school's habit of waiting to be told what to do. Wu closes by asking listeners to send reproducible examples where Claude Code or Cowork fail, since edge cases and errors help improve future models and harnesses.