The company building the models is shipping product faster than the companies building on top of them.
The Summary
- Cat Wu, Head of Product for Claude Code at Anthropic, details how the company's product team operates at a speed that outpaces traditional product development, including building features before the underlying models are ready
- Anthropic uses "launch rooms" where cross-functional teams ship products in compressed timeframes, fundamentally rewriting the relationship between product strategy and execution speed
- The AI product manager role is evolving: less time on specs and roadmaps, more time understanding model capabilities and constraints that don't exist in traditional software
The Signal
Anthropic ships Claude Code updates in weeks, not quarters. That matters because the model labs were supposed to be infrastructure companies. They'd build the brains, and everyone else would build the products. That's not what's happening.
Wu describes a product process that inverts the normal order. Typically, you spec the feature, build the tech, then ship. At Anthropic, product teams often start building the interface and user experience before the model can actually do the thing. They're building the steering wheel while the engine is still being assembled. This works because model capabilities improve on a predictable curve. If you know GPT-5 or Claude 4 will be able to handle multi-file refactoring in three months, you start building the UI today.
"Speed now matters more than strategy."
The "launch room" process is where this gets concrete. Cross-functional teams lock in for compressed cycles, sometimes just days, to ship a feature end-to-end. No six-month roadmap. No strategic planning deck. Just: model capability emerging, user need identified, team assembled, feature shipped. This isn't lean startup iteration. It's faster. It's building in public, but with the model itself as the uncertain variable.
This has direct implications for everyone building on top of these models:
- Model labs are now your product competitors, not just your infrastructure providers
- The cadence of foundation model releases now sets the tempo for product work across the entire stack
- Traditional product planning horizons (quarterly roadmaps, annual strategies) are mismatched to the rate of capability improvement
Wu's description of the evolving PM role is where the human work question gets interesting. Product managers at AI companies spend less time writing specs and more time understanding what the model can and can't do. That's a different skill. It's closer to working with a talented but unpredictable contractor than managing a deterministic software system. You need to know the edge cases, the failure modes, the contexts where the model hallucinates or refuses or produces garbage.
The interview reveals that Anthropic's product team treats model updates almost like a new platform launch every few months. Each major model release reshuffles what's possible. Features that were impossible in January are trivial in March. Roadmaps decay faster than the models improve.
The Implication
If you're building on foundation models, your competitors aren't just other startups. They're OpenAI, Anthropic, Google, and whoever else controls the base layer. They have the models, the compute, and now, increasingly, the product velocity. The defensibility question for AI application companies just got harder.
For product people, the lesson is clear: your planning horizon needs to compress. Quarterly roadmaps are already outdated by the time you write them. Build for the model capabilities coming in three months, not the ones you have today. And get comfortable shipping before you're ready, because the models themselves are shipping before they're ready.