Anthropic's Claude Code lead on its ultra-fast product cycle: 'Jobs are fake'

Cat Wu said something that stuck with me: "Jobs are fake."

She's the product lead for Anthropic's Claude Code, and she recently appeared on Lenny's Podcast. There was so much in that episode that I tried to compress the core insights.

Anthropic's product velocity is ridiculous. Features go from idea to launch in as little as a day. It's not because they use the most powerful model (though that helps). It's about organization and process. Almost no approvals. Everything ships as a "research preview" to lower the bar for release. Engineers see user feedback on Twitter end-to-end, then build and ship a feature over the weekend without needing a PM. The PM's job isn't to gatekeep; it's to build the launchpad so engineers can ship whenever they want, with marketing and docs following the next day.

The most counterintuitive thing is their attitude toward new models. When others release a new model, they add features. When Anthropic releases a new model, the first thing they do is delete features — strip away all the patches they added for the previous model. To-do lists, forced reminders, prompting hacks — if the model is smarter, you don't need them. For example, early models would miss calls when refactoring 20 call sites, so the team added a to-do list tool to track them. By Opus 4, the model used it autonomously. Wu says every time a model upgrades, the team reads through the entire system prompt, asking paragraph by paragraph: "Does the model still need this reminder?" If not, delete. As the podcast put it: "The model will eat your harness for breakfast."

But more importantly, they build products that "can't work yet." They tried several versions of code review, but early models weren't accurate enough to ship. Only with Opus 4.5/4.6 did the team feel confident that engineers could trust it to catch most bugs before merge. Her advice: always build prototypes early, and when a new model comes out, plug it in and see if the gap is closed.

On the scarcest skill for PMs, Wu keeps coming back to "product taste." As code gets cheaper, taste becomes more valuable. GitHub has tens of thousands of issues with every feature request imaginable. Knowing which one to build and how to build it best — that judgment is what matters. Engineering background helps in the short term (to gauge implementation difficulty and prioritize), but she deliberately says "for the next few months" because model capabilities leap every few months, and required skills change. No one can predict farther. She believes what truly matters is first-principles thinking — understanding how the tech landscape has shifted, what the team is missing most, and then putting on whatever hat is needed without ego.

Another reason Anthropic moves so fast: unified mission. When two priorities conflict, ask which one better serves Anthropic's mission, and the answer is clear. The whole team executes instantly. Wu said something weighty: "If Claude Code fails but Anthropic succeeds, I'll be very happy." Everyone is willing to sacrifice their own product's key results for the company's goals. This culture enables lightning-fast decisions, and it's why a company that started late, raised less money, and had no distribution advantage can hit $11 billion ARR.

Some practical advice from Wu:

1. Delegate repetitive work to AI, but don't stop at 95%. 95% accurate automation isn't automation — that last 5% is where the real investment pays off.

2. Build apps you actually use every day. One-shotting a prototype, tweeting about it, and never opening it again teaches you nothing and gives you no leverage.

3. Don't get obsessed with showing off your setup. Her exact words: "Simple setups tend to work better." Some people spend tons of time stacking skills and MCP, and end up doing zero core work.

4. 2024's products were conversational. Claude Code's generation is action-oriented. The real aha moment is when AI can truly execute on your behalf, not just tell you what to do.

Her life philosophy circles back to that opening line: Just do things. Understand constraints, derive actions, then go. Don't wait for permission. Don't let your job title limit you. That probably explains why Anthropic can run this fast.