Arguably, Anthropic's product release cadence is truly one of a kind.
If you plotted Anthropic's recent product launches on a calendar, you'd find that nearly every day brings a new feature.
Recently, Lenny Rachitsky hosted Kat Wu, Anthropic's product lead for Claude Code and Cowork, for a podcast episode. The conversation was dense with insight, covering the evolving PM role, Anthropic's internal processes, a source code leak, and the OpenClaw decision.
I've distilled the key points from that conversation:
1. Who She Is
Kat Wu spent years as an engineer, did a brief stint in VC, then joined Anthropic to lead product for Claude Code and Cowork.
She partners with Boris (the creator and tech lead for Claude Code). Boris defines the product vision—what it should look like three to six months out. Kat breaks that vision into executable paths and coordinates across go-to-market, sales, finance, and other teams.
"We agree on about 80% of ideas. For the remaining 20%, whoever cares more pushes it through."
One distinctive trait of their team: nearly all PMs have engineering backgrounds, and designers were previously front-end engineers.
That wasn't intentional, Kat says. But having an engineering background helps you quickly gauge how hard something is to build—a judgment that's absolutely critical at their current pace.
2. How Fast?
Anthropic's product iteration cycle has compressed from six months to one month, with some features shipping in a single day.
The secret behind this isn't a complex methodology. Kat highlights three things:
Set clear goals. LLMs are too general. Without locking in a user and scenario, teams easily lose focus. For example, Claude Code's target user is the professional developer; a specific feature might aim to solve "too many permission dialogs causing fatigue" so enterprise developers can achieve zero-permission prompts securely. This single goal eliminates countless irrelevant solutions.
Research Preview mechanism. Almost every feature launches first as a research preview. Users know it's an early version that might not stick around. The benefit: the team doesn't need to reach perfection before shipping. They can push something out in a week or two.
Launch Room process. When an engineer thinks a feature is ready, they drop it into an "evergreen launch room" channel. Docs lead Sarah, PMM lead Alex, and DevRel's Tarek and Lydia quickly pick it up, and it can be live the next day.
"We want to remove every barrier to shipping. Everyone on the team should be able to turn their idea into a product within a week—or even a day."
Lenny couldn't resist asking: You use internal models like Mythos… is that why you're so fast?
Kat replied:
"We do use those models internally, and they speed things up a bit. But most of the acceleration comes from process and team culture."
3. The Model Will Eat Your Product
Kat discussed a notable phenomenon: every time a new model comes out, the first thing they do is delete features.
Early on, Claude Code had a to-do list feature. When the model refactored large codebases, it would stop after changing 5 call sites, even when 20 needed changing. The team came up with a workaround: have the model list all the tasks first, then complete them one by one.
That worked brilliantly. But after Opus 4, the model started proactively listing and completing tasks on its own. The to-do list went from a necessary crutch to a nice-to-have UI.
"With each new model release, we read through the entire system prompt, paragraph by paragraph, and reconsider: does the model still need this hint? If not, we delete it."
Lenny gave a metaphor: "The model will eat your harness for breakfast."
Kat agreed.
But what excites her more are the new capabilities unlocked by advanced models. Take code review: they tried several versions, but accuracy wasn't sufficient until Opus 4.5 and 4.6, which finally made the engineering team rely on it. Now, before merging a PR internally at Anthropic, Claude's code review is mandatory.
"It's important to build prototypes of things that aren't quite ready yet. That way, when a new model drops, you can plug it in and see if the gap has closed."
This echoes what Mike Krieger said in a previous podcast: build products for the future model.
4. Source Code Leak
Regarding last month's Claude Code source code leak, Kat responded:
"We investigated immediately. It was a human error—someone made a mistake while using Claude to write a PR that updated the release process. Even though it went through two layers of human review, it slipped through."
When Lenny asked if that person still worked there, Kat's answer reflected Anthropic's culture:
"It's a process issue. The most important thing is to learn from it and add more guardrails. Most improvements are already live."
5. The OpenClaw Decision
Lenny also brought up OpenClaw. Recently, Anthropic restricted third-party tools (like OpenClaw🦞) from using Claude subscription quota, sparking community outcry.
Kat explained: demand for Claude has grown rapidly, and subscription quota was originally designed for first-party products. Third-party tools have different usage patterns, putting extra pressure on infrastructure.
"We spent a lot of time figuring out the smoothest transition. I'm relieved we could give each user some credits as a buffer. But we do need to prioritize first-party products and the API."
Lenny sided with Anthropic:
"You're offering nearly unlimited usage at $200 per month, which itself is subsidizing users. The company also needs to be profitable."
6. Cowork: The Underrated Product
The episode also featured a segment on Cowork.
Kat used Cowork to prepare a 20-page keynote for the upcoming Code with Claude conference. Her approach: connect Google Calendar, Slack, Gmail, and Google Drive, then tell Cowork the topic and narrative direction.
Cowork spent about an hour scanning X for product launch records, the internal launch room channel, and demo channel, then compiled a 20-page presentation.
"I read it in the morning—it was pretty good. The text was a bit dense, so I gave feedback and iterated once. But visually, it looked like an Anthropic designer made it, because Cowork can read our design system templates."
She splits her product usage: if the output is code, use Claude Code. If not (PPT, docs, emails), use Cowork.
The Applied AI team—a technical go-to-market team that helps customers adopt the Claude API—is probably the biggest internal token consumer after engineers. They use Cowork to auto-generate briefings before client meetings: who they're meeting tomorrow, what they've asked before, action items, and the latest launch date for a feature.
These are custom workflows they built and shared with the team.
7. The Future of PMs
Kat's take on the evolving PM role is one of the most valuable parts of the episode.
"Code is getting cheaper. So what becomes more valuable? Deciding what to write."
She says Anthropic currently prefers hiring engineers with product taste, rather than traditional PMs. Many engineers on the team can go from seeing user feedback on X to shipping a feature over the weekend, with little PM involvement.
"The roles of PM and engineer are converging. You can either hire more engineers with product taste, or more PMs to guide engineering direction—the effect is similar."
But what's the trade-off?
Product consistency. Sometimes the same requirement gets two features built because the team has two equally good approaches, so they ship both and let users vote. For new users, this means more time figuring out what to use.
Kat admits the /powerup feature—a built-in tutorial guide—goes against their original principle that "a product should be intuitive enough not to need a tutorial." But with so many features, users need someone to tell them which ten out of a hundred are must-haves.
8. Ask the Model Why It Erred
Kat shared a unique technique for building AI products: have the model reflect on its own behavior.
For instance, she noticed that after modifying front-end code, the model would run tests—but never actually open the page to check the UI. She asked it: why don't you check the UI?
The model's answers can be surprising:
"Sometimes the model says a certain paragraph in the system prompt confused it. Sometimes it says it delegated the verification task to a sub-agent, which didn't do it, and it didn't check the sub-agent's work."
These insights directly point to harness improvements.
"Stay curious about the model's decisions. Ask it why it made that choice. You'll see what misled it, and then fix the harness to fill that gap."
9. 50 Claudes in Parallel
Discussing the roadmap, Kat broke down product evolution into stages, like building blocks:
Step 1: Single task success. With a clear prompt, can the model consistently output mergeable code or shareable documents?
Step 2: Multi-task parallelism. Multi-coding was already trending by end of 2025. Users currently run about 6 Claudes simultaneously.
Step 3: 50 to hundreds of Claudes running concurrently. At that scale, local machine memory won't suffice—tasks will need to run remotely. The interface will need to tell you which tasks require your attention, and the model should be able to verify its work so you can trust it when it says "done."
"And this process needs to self-improve. You give feedback once, and the model never makes the same mistake in future runs."
10. Just Do Things
Kat's life motto: Just do things.
"Work is fake anyway. If you understand the constraints, you can figure out what to do and just do it. Move fast, learn from mistakes, apologize and fix it if you're wrong."
In Anthropic's context, this makes sense: many companies have strict role boundaries—PMs do PM things, engineers do engineer things. But Anthropic encourages crossing boundaries: if you see a problem, solve it, regardless of whose territory it is.
11. Don't Stop at 95%
The last takeaway worth noting is about automation.
Kat says she's seen two extremes: people who never automate, and people who obsess over tooling—adding MCPs, setting up skills—spending more time configuring than completing tasks.
But the more common pitfall? Many people automate to 95% and give up.
"If an automation doesn't work 100% of the time, it's not real automation. That last 5% does take more effort, but you need to invest it—teach Claude your preferences, give it feedback—until it runs reliably."
She admits she hasn't reached 100% on Cowork-based Gmail inbox zero yet either.
But that's the direction: find the repetitive, unpleasant tasks, hand them to AI, and polish them until they're fully reliable. The time you save is for the things you truly want to do.
That's AI's real leverage for everyone.
◇ ◆ ◇
Related links:
• Lenny's Podcast article: https://www.lennysnewsletter.com/p/how-anthropics-product-team-moves
• Full YouTube video: https://www.youtube.com/watch?v=PplmzlgE0kg
• Kat Wu on X: https://x.com/_catwu