Joerg Hiller
Mar 19, 2026 22:46
Claude Code’s Head of Product shares how exponentially improving AI models are forcing product teams to abandon traditional roadmaps for rapid experimentation.
Cat Wu, Head of Product for Claude Code at Anthropic, dropped a revealing look at how rapidly improving AI models are fundamentally breaking traditional product management approaches. The key insight? What’s technologically possible at the start of a project no longer predicts what’s possible at the end.
The numbers back this up. According to METR research cited by Wu, Opus 4.6 can now complete software tasks that would take humans nearly 12 hours—roughly 41 times more capable than Sonnet 3.5 (new) was just 16 months ago when it handled 21-minute tasks.
The Old Playbook Is Dead
Product managers traditionally gathered requirements upfront, locked in a roadmap, then executed over months. That’s not working anymore when model constraints you designed around can vanish mid-project.
“You’re building on ground that’s rising underneath you,” Wu writes. Her team has responded by ditching long-term roadmaps entirely in favor of what she calls “side quests”—short, self-directed experiments where anyone on the team (engineers, designers, PMs) can prototype ideas in an afternoon.
Several popular Anthropic features emerged this way: Claude Code on Desktop, the AskUserQuestion tool, and todo lists all started as informal experiments rather than planned roadmap items.
Three Tools, One Workflow
Wu’s daily workflow now spans three distinct AI products. Claude.ai handles strategic thinking and quick answers. Claude Code builds prototypes and evals. Cowork manages everything else—email, todo lists, slide decks, Slack research, travel booking.
External PMs are finding similar patterns. Bihan Jiang, Director of Product at Decagon, told Wu that what used to take weeks of building to get in front of customers now happens in “a couple of hours.” Kai Xin Tai at Datadog described the shift as moving “from defining certainty upfront to accelerating discovery.”
Practical Shifts for Product Teams
Wu outlined four concrete changes her team has adopted:
Prototype before documenting. After writing a spec, send it to Claude Code and see what comes back. “Even a rough prototype changes the conversation,” she notes. When a team member shared a plugins spec, the AI-generated prototype came back nearly production-ready.
Revisit features with each model release. Claude Code with Chrome happened because users were manually copying instructions between tools. The hack worked well enough that it became a built-in feature.
Optimize for capability first, costs later. Use more tokens than you think you need during prototyping. “You can always bring costs down later as cheaper models catch up.”
Keep implementations simple. Complex workarounds for model limitations become unnecessary baggage when the next model drops. Anthropic cut 20% of their system prompting with Opus 4.6 alone.
What This Means for AI Product Teams
The broader industry context matters here. AI product management has emerged as a distinct discipline requiring both traditional PM skills and deep technical understanding of model capabilities. With regulations like GDPR and emerging AI governance frameworks adding compliance layers, the role has grown more complex even as the tools have grown more powerful.
Wu’s core message for fellow PMs: track two things simultaneously—how AI changes your workflow and how it changes what’s possible in your product. The teams that do this well won’t be caught off guard when capabilities leap forward.
For enterprise software teams watching AI development costs and timelines, the implications are significant. If prototyping cycles compress from weeks to hours, competitive advantages built on execution speed may erode faster than expected.
Image source: Shutterstock
