Shortcast

Today’s guest is Kat Wu, head of product for Claude Code and Co‑Work at Anthropic. I’ve never seen a team ship this fast, and Kat sits at the center of how AI is changing product building. Kat, welcome—let’s dig into how you operate, your split with Boris, and what PMs need to do to keep up.

Thanks for having me. It’s easy to design for a future, super‑smart model; the real craft is getting today’s models to do the most they can and building products that bring out that capability.

You’re releasing at a wild pace. What changed to move from six‑month timelines to shipping in days?

We remove every barrier to launch, keep process light, and set the expectation that an idea can be in users’ hands within a week—sometimes a day.

Help us understand your partnership with Boris and your lane versus his.

Boris drives long‑range product vision. I map the path from now to that horizon and make sure marketing, sales, finance, and capacity are aligned so nothing blocks shipping; we mind‑meld on most decisions, then each of us owns the parts we care most about.

You interview a ton of PMs. What are they getting wrong about AI product work?

The cadence is much faster now. Great PMs tighten the loop from idea to live feature, carve out sandboxes for rapid launch, and define the few core tasks their product must nail.

What enables that speed besides having strong models?

Clarity, cadence, and a crisp launch pipeline. We set specific goals and users, ship in clearly labeled research preview to lower commitment, then run a tight handoff between eng, docs, PMM, and DevRel so an approved feature can announce the next day.

Do PRDs still matter?

We run weekly metrics readouts and team principles so decisions don’t bottleneck on PM. For ambiguous or infra‑heavy work, we still write short PRDs to lock goals, delightful use cases, and failure modes.

Is your new internal model the secret to speed?

It helps, but process and expectations explain most of the acceleration.

Quick hits. The leaked Claude Code source—what happened?

It was human error during a release PR that even passed two reviews. We hardened safeguards and focused on learning, not blame.

And the OpenClaude change blocking third‑party use of Claude subscriptions?

Demand spiked and we had to prioritize first‑party products and the API. We offered credits for transition, but usage patterns differ and we had to choose.

What does the PM org look like?

Roughly thirty to forty PMs across research PM, the developer platform, Claude Code and Co‑Work, enterprise adoption, and growth.

Are PMs still needed if engineers can ship so fast?

Roles are blending. We hire engineers with strong product taste so many features can go end‑to‑end without PM. That said, real product taste is rare and we’ll hire it in any background.

Which skills matter most now?

Taste and time horizons. As code gets cheaper, deciding what to build is the premium skill. An engineering background helps estimate effort, but the valued mix shifts every few months.

Where do humans still shine?

Common sense, stakeholder mapping, and EQ. Launches have a thousand moving parts, and models still miss the social and organizational nuance.

How do you stay sane in constant change?

We hire people who lean into chaos, stay calm, and relentlessly prioritize. It’s okay to ship imperfectly if the core use case is unblocked and you’ll iterate fast.

What trade‑offs come with moving this quickly?

Less product consistency. We sometimes ship overlapping ideas to test form factors, so we added an in‑product onboarding flow to guide people to best practices.

Anthropic’s rise has been stunning. What are the key ingredients?

A unifying mission—safe AGI for everyone—and a willingness to subordinate local goals to company priorities. That focus lets us make fast, org‑wide calls.

When should people use CLI, Desktop, Web or Mobile, and Co‑Work?

I use CLI for the newest code features and quick tasks. Desktop shines for front‑end work and a visual overview of sessions. Web and mobile let you kick things off on the go. If the output isn’t code, I reach for Co‑Work.

Give us a real Co‑Work example from your PM work.

I connected Calendar, Slack, Gmail, and Drive, then asked for a conference deck with a specific narrative. Co‑Work pulled launches, team demos, and public posts, then produced a polished, on‑brand draft I refined in minutes.

What’s your daily stack and any cool internal builds?

Claude Code, Co‑Work, and Slack are my core. People now whip up tailored internal tools fast; one sales app assembles customer‑specific decks by pulling Salesforce, Gong, and notes, saving half an hour per deck.

Are token costs getting wild internally?

Usage rises with every model jump, but it’s still well below salary. We give people room to build while expecting responsible spend.

Which team is second to engineering in usage?

Applied AI. They prototype with customers and juggle a lot of comms, so they lean hard on both Co‑Work and Code, including daily briefs before customer calls.

What emerging PM skill is hardest to learn?

Being the right amount of AGI‑pilled. Don’t build for a fantasy model; learn to coax maximum capability from today’s model and guide users to its strengths.

How do you build that muscle?

Ask the model to reflect on mistakes, curate a small set of power users whose feedback is high signal, and write a handful of focused evals to define and measure success.

Why does Claude’s personality matter so much?

Great teammates are competent, low‑ego, and encouraging. We aim for that vibe—light but serious about the work, action‑oriented, and willing to admit errors and fix them.

What happens when new models land?

We delete old crutches the model no longer needs and unlock new features. Code review, for example, crossed the reliability bar only with the latest models, and now we lean on it before merges.

What’s the longer‑term vision for Code and Co‑Work?

Make single tasks rock‑solid, then scale to many in parallel, then to hundreds. That requires remote orchestration, strong verification so “done” is trustworthy, and self‑improving loops from your feedback.

Advice for PMs and founders who want to thrive through this shift?

Automate the repetitive parts of your job until they’re truly hands‑off, then spend your new time on higher‑leverage work. Aim for one hundred percent reliability—ninety‑five percent isn’t automation.

Any final product counsel?

Build tools you actually use daily and avoid disappearing into over‑customization. The big unlock is moving from chat to action—watch the agent do the work for you.

Lightning round. Favorite books or shows and a product you love?

How Asia Works, The Technology Trap, and Paper Menagerie; Drive to Survive and Free Solo. Outside of Anthropic tools, Waymo changed my routine by giving me quiet commute time back.

A motto you live by, and your favorite Claude thinking word?

Operate from first principles and just do the thing. Favorite thinking word: manifesting.

If AGI freed up your time, what would you do?

Help the world through the transition, then climb a lot—probably in Fontainebleau—and finally catch up on a very long reading list.

Where can people find you, and how can listeners help?

Find me on Twitter; my DMs are open. What helps most are clear repros of edge cases where Code or Co‑Work fails; we route love notes to a team channel and issues to our feedback channel so we can fix them.

Thank you, Kat. If you enjoyed this, subscribe on your podcast app, consider leaving a rating or review, and find past episodes at lennys podcast dot com. See you next time.