scripod.com

How Anthropic’s product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)

In this episode, Cat Wu, Head of Product for Claude Code and Cowork at Anthropic, shares insights from building AI-native products at breakneck speed—where models evolve faster than roadmaps, roles blur, and shipping imperfect tools is not a compromise but a strategic necessity.
Cat Wu outlines how AI is reshaping product management: shipping cycles have accelerated from months to days, demanding PMs who prioritize rapid iteration, product taste, and deep model fluency over traditional planning. She emphasizes building products that anticipate future model capabilities—not just current ones—and highlights the underrated power of prompting models to introspect on their own errors. At Anthropic, mission alignment replaces bureaucratic friction, enabling teams to deprioritize individual projects for company-wide goals. Roles are converging—engineers take on product decisions, designers come from engineering backgrounds—and success hinges on first-principles thinking, judgment, and emotional intelligence—areas where humans remain irreplaceable. Tools like Cowork automate high-value workflows (e.g., generating slide decks in an hour), but reliability requires 100% accuracy, not 95%. Crucially, Cat stresses building daily-use applications—not prototypes—and adopting a 'just do things' mindset to thrive amid constant change.
00:00
00:00
The PM role is changing to emphasize product taste and quick iteration for AI-native products
01:32
01:32
Boris is the tech lead and product visionary; Cat focuses on the path to the product vision and cross-functional alignment
04:30
04:30
AI accelerates engineering and model capabilities, shortening product timelines
06:19
06:19
Shortening the time from idea to product in users' hands is essential for AI-native PMs
08:59
08:59
PRDs are written for projects requiring heavy infrastructure and long-term work
10:28
10:28
The main reason for Anthropic's fast shipping rate is the low-process environment that empowers team members to turn ideas into products quickly
11:54
11:54
The Claude Code source code leak resulted from human error during a package release, not a security breach or malicious act.
12:54
12:54
Third-party access to Claude subscriptions was restricted to prioritize first-party products and the API
14:19
14:19
Anthropic has around 30–40 product managers across five key teams
15:42
15:42
Most PMs on their team have engineering backgrounds, and designers were also front-end engineers
17:54
17:54
Product taste is the most important skill for those coming from engineering, product, or design backgrounds in the AI field
20:10
20:10
Human brains remain essential for picking work, prioritizing, and providing common sense and EQ knowledge that models currently lack
22:24
22:24
Ship imperfect products as long as they don't block the core use case
27:05
27:05
With agentic tools in the ecosystem, people feel the need to check X daily
27:48
27:48
/powerup shows best practices because users needed guidance despite the product's intuitive design goals
31:25
31:25
Teams are willing to sacrifice for Anthropic's goals—even celebrating if Claude Code fails but Anthropic succeeds
34:57
34:57
Cowork is growing fast but many still don't understand its purpose
35:59
35:59
Cowork synthesized a 20-page conference deck in one hour from integrated data sources
38:44
38:44
Cowork generates a slide deck outline via Claude after connecting communication and data-storage tools, with the PM retaining final content authority.
41:49
41:49
Claude Code powers internal tools like a web app for customizing sales decks, replacing manual work
49:55
49:55
Token costs per knowledge worker increase with model improvements, though still lower than average engineer salaries
51:17
51:17
Good PMs spot patterns, set directions, and adjust paths amid model and user behavior ambiguity
57:54
57:54
Individual feedback helps identify issues like lack of explanation, abruptness, memory quality, and self-testing levels
58:45
58:45
Claude's personality is light-hearted, fun, low-ego, positive, and action-biased
1:00:45
1:00:45
Opus 4 can naturally perform tasks without prompting, making earlier prompting crutches unnecessary
1:05:12
1:05:12
The core building block is making individual tasks successful — as models get smarter, task success rate increases, and users move toward doing multiple tasks simultaneously
1:07:23
1:07:23
Iterate on AI automations until success rates are high, then redirect focus to creative projects previously out of reach
1:09:19
1:09:19
95% accuracy isn't reliable—AI automations must reach 100% accuracy to be trustworthy
1:11:58
1:11:58
Build apps that are used daily to get real value from AI
1:13:41
1:13:41
The ability of AI agents to perform tasks is an eye-opening moment
1:18:18
1:18:18
Waymo allows productive, hands-free commuting—worth paying a premium for