Joerg Hiller
Mar 19, 2026 22:46
Claude Code’s Head of Product shares how exponentially enhancing AI fashions are forcing product groups to desert conventional roadmaps for speedy experimentation.
Cat Wu, Head of Product for Claude Code at Anthropic, dropped a revealing have a look at how quickly enhancing AI fashions are basically breaking conventional product administration approaches. The important thing perception? What’s technologically potential initially of a mission now not predicts what’s potential on the finish.
The numbers again this up. In response to METR analysis cited by Wu, Opus 4.6 can now full software program duties that might take people practically 12 hours—roughly 41 occasions extra succesful than Sonnet 3.5 (new) was simply 16 months in the past when it dealt with 21-minute duties.
The Outdated Playbook Is Useless
Product managers historically gathered necessities upfront, locked in a roadmap, then executed over months. That is not working anymore when mannequin constraints you designed round can vanish mid-project.
“You are constructing on floor that is rising beneath you,” Wu writes. Her group has responded by ditching long-term roadmaps completely in favor of what she calls “facet quests”—quick, self-directed experiments the place anybody on the group (engineers, designers, PMs) can prototype concepts in a day.
A number of standard Anthropic options emerged this manner: Claude Code on Desktop, the AskUserQuestion instrument, and todo lists all began as casual experiments fairly than deliberate roadmap objects.
Three Instruments, One Workflow
Wu’s every day workflow now spans three distinct AI merchandise. Claude.ai handles strategic considering and fast solutions. Claude Code builds prototypes and evals. Cowork manages all the things else—electronic mail, todo lists, slide decks, Slack analysis, journey reserving.
Exterior PMs are discovering comparable patterns. Bihan Jiang, Director of Product at Decagon, instructed Wu that what used to take weeks of constructing to get in entrance of shoppers now occurs in “a few hours.” Kai Xin Tai at Datadog described the shift as shifting “from defining certainty upfront to accelerating discovery.”
Sensible Shifts for Product Groups
Wu outlined 4 concrete adjustments her group has adopted:
Prototype earlier than documenting. After writing a spec, ship it to Claude Code and see what comes again. “Even a tough prototype adjustments the dialog,” she notes. When a group member shared a plugins spec, the AI-generated prototype got here again practically production-ready.
Revisit options with every mannequin launch. Claude Code with Chrome occurred as a result of customers had been manually copying directions between instruments. The hack labored nicely sufficient that it turned a built-in characteristic.
Optimize for functionality first, prices later. Use extra tokens than you suppose you want throughout prototyping. “You possibly can all the time deliver prices down later as cheaper fashions catch up.”
Preserve implementations easy. Complicated workarounds for mannequin limitations change into pointless baggage when the subsequent mannequin drops. Anthropic reduce 20% of their system prompting with Opus 4.6 alone.
What This Means for AI Product Groups
The broader business context issues right here. AI product administration has emerged as a definite self-discipline requiring each conventional PM abilities and deep technical understanding of mannequin capabilities. With rules like GDPR and rising AI governance frameworks including compliance layers, the function has grown extra complicated even because the instruments have grown extra highly effective.
Wu’s core message for fellow PMs: monitor two issues concurrently—how AI adjustments your workflow and the way it adjustments what’s potential in your product. The groups that do that nicely will not be caught off guard when capabilities leap ahead.
For enterprise software program groups watching AI improvement prices and timelines, the implications are vital. If prototyping cycles compress from weeks to hours, aggressive benefits constructed on execution velocity could erode quicker than anticipated.
Picture supply: Shutterstock
