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

Lenny's Podcast · video · 2026-04-23 Reader 🖥️

practical · mixed (expository/argumentative) — A conversational interview mixing practical advice on PM practices, product strategy, and org design at Anthropic with implicit arguments about how AI changes the nature of product work.

Structure

Key terms

• product taste Taste & Judgment
  The ability to decide what to build, how to build it delightfully, and which of thousands of requests are worth pursuing — includes judging model output quality
• AGI pilled Riding Model Improvements
  The degree to which you build your product assuming future superintelligent models vs. optimizing for current model capabilities
• research preview
  A branded early release mechanism that reduces commitment — features ship in 1-2 weeks with explicit framing that they may not persist
• harness Agent Harness
  The product scaffolding (prompts, tools, UI) built around the model to elicit maximum capability from its current level
• golden path Golden Path
  The optimal interaction pattern that aligns with the model's strengths and compensates for its weaknesses
• thinking words
  The verbs displayed in the UI when Claude is processing (e.g., 'manifesting')
• the model will eat your harness for breakfast Riding Model Improvements
  As models improve, product scaffolding built to compensate for model weaknesses becomes unnecessary and should be removed

Propositions

• explicitcorroboratedFeature timelines at Anthropic have compressed from 6 months to 1 month, and sometimes to one week or one day.conf: high
  supports ←Lenny Rachitsky's Post · p1 — Both report the same six-months-to-a-day timeline compression at Anthropic; this post adds the mechanism detail (research-preview labeling, the launch-room pipeline).
• explicitcorroboratedThe PM role should shift emphasis from multi-quarter roadmap alignment to minimizing the time from idea to user hands.conf: high
  supports ←Lenny Rachitsky's Post · p2 — Same claim about the PM role shifting from roadmap coordination to shipping-speed enablement, quoting the same source (Catherine Wu).
• explicitcorroboratedAs code becomes much cheaper to write, deciding what to write (product taste) becomes the most valuable skill.conf: high
  supports ←Adam Mosseri: AI is a tailwind for authenticity · p1 — Near-identical claim: as building/code becomes cheap, deciding what to build (product taste) becomes the most valuable skill.
• explicitAn automation that works 95% of the time is not really an automation; the last 5-10% is where the real value is.conf: high
• explicitdevelopedWith each new model generation, the team removes prompting interventions and product crutches that are no longer needed.conf: high
  extends ←Lenny Rachitsky's Post · p4 — Where the primary extraction describes removing now-unnecessary prompting crutches as models improve, this post describes the complementary proactive practice: building and shipping a product ahead of model readiness specifically so each new model release can be tested against an existing prototype immediately.
• explicitAnthropic's unifying safety mission allows faster cross-org decision-making because competing priorities can be resolved by reference to the mission.conf: high
• explicitIf Claude Code failed but Anthropic succeeded, the whole team would be extremely happy.conf: high
• explicitdevelopedAll roles are merging: PMs do engineering, engineers do PM work, designers land code.conf: high
  extends ←Lenny Rachitsky's Post · p3 — Adds concrete team-composition detail (most PMs have engineering backgrounds, designers are front-end engineers, engineers ship end-to-end without a PM) to the primary extraction's more general 'all roles are merging' claim.
• explicitThe valued skill set changes every few months due to large jumps in model coding capability.conf: medium
• explicitdevelopedAsking the model to introspect on its own mistakes reveals what misled it, enabling you to fix the harness.conf: high
  refinesHarness design for long-running application development · p15 — Wu adds model self-introspection on mistakes as a harness-debugging technique alongside the Labs post's human trace-reading.
  extends ←Lenny Rachitsky's Post · p5 — Adds a concrete failure-mode example (a subagent that delegated verification without actually checking its work) to the primary extraction's more general claim about model self-introspection as a debugging technique.
• explicitThe big shift from 2024 to the Claude Code generation is from chat-based to action-based products.conf: high
• explicitcontestedSimple setups actually work better than heavily customized ones.conf: medium
  contradictsHarness engineering for coding agent users · p1 — Wu reports simple setups outperform heavily customized ones, cutting against Böckeler's recommendation that users build an elaborate outer harness on top of the built-in one.
• explicitAnthropic prioritized first-party products and API over third-party usage (Open Claude), even though it upset parts of the community.conf: high
• explicitHumans still provide a level of common sense, stakeholder awareness, and EQ that models lack.conf: high
• explicitThe product consistency cost of rapid shipping is that new users may not know the best path to accomplish a goal.conf: high
• explicitcontestedBuilding 10 great evals is more valuable than building hundreds; evals are underappreciated by PMs and engineers.conf: high
  supportsPodcast: Tiger Teams, Evals and Agents: The New AI Engineering Playbook · p3 — Wu's few-great-evals-underappreciated claim corroborates Bhagwat's claim that evals carry outsized importance in AI engineering.
  contradicts ←Kaggle is making AI benchmark creation effortless · p7 — Kaggle's bet that quantity and diversity of evals outweigh quality risks directly opposes Wu's claim that 10 great evals beat hundreds.
• explicitCloud Code's source code leak was caused by human error that passed through two layers of human review.conf: high
• implicitToken cost per knowledge worker increases with every model jump or substantial product improvement.conf: medium
• explicitThe progression of Claude Code's vision is: single task success, multi-tasking (6 at a time), then orchestrating 50-100+ remote agents with self-improving feedback loops.conf: high

Arguments

Narratives

• An engineer noticed that early Claude Code would start a 20-call-site refactor but only complete 5. Sid on the team added a to-do list tool inspired by VS Code's 'find all references' list. This forced Claude to track and complete all sites. With Opus 4+, the model naturally uses the to-do list without prompting, making the crutch optional.
  → Illustrates how product scaffolding compensates for model weaknesses and becomes unnecessary as models improve
• Cat Wu used Co-work to prepare a slide deck for the Code with Claude conference. She connected Google Drive, Slack, and other sources, gave Co-work a prompt describing the narrative she wanted, and it worked for an hour — searching Twitter, internal launch channels, and demo channels — producing a 20-page deck in Anthropic's design system. She reviewed it the next morning, gave one round of feedback (too wordy), and had a near-final product.
  → Demonstrates Co-work's capability for non-code knowledge work and the PM's role as decision-maker on top of AI-generated options
• A sales team member built a custom web app that pulls from Salesforce, Gong, and internal notes to auto-generate tailored customer decks. It customizes slides based on the customer's deployment (Bedrock, Vertex, Claude for Enterprise), their concerns (code review, HIPAA compliance), and removes irrelevant slides — reducing 20-30 minutes of manual work to seconds.
  → Illustrates the 'surge in personalized work software' enabled by Claude Code lowering the barrier to building custom apps
• Applied AI team members use Co-work to prepare for 5-10 customer meetings per day. The night before, they ask Co-work to summarize upcoming meetings, pull customer history, surface action items from past meetings, and research answers to open questions — producing a dossier they review before each call.
  → Illustrates Co-work adoption beyond engineering, in a technical go-to-market role
• The Claude Code source code leaked because a human working with Claude to write a PR made an error that passed through two layers of human review. Anthropic kept the person on the team, treated it as a process failure, and hardened safeguards.
  → Illustrates Anthropic's blameless culture and process-improvement orientation
• Code review was attempted multiple times with earlier models but accuracy was insufficient. Only with Opus 4.5/4.6 and Sonnet 4.6 did the team feel confident enough that multiple code review agents could traverse the entire codebase and synthesize real issues — to the point where the engineering team now relies on it before merging PRs.
  → Illustrates the strategy of building products at the edge of current capability and waiting for models to catch up

Questions addressed

• ✅ How does Anthropic's product team ship so fast?
  Clear goals, research preview as a low-commitment launch mechanism, tight engineering-marketing-docs process, and a culture that empowers any individual to go from idea to shipped product in under a week.
• ✅ What is the future of the PM role in an AI-native world?
  Roles are merging. Product taste — deciding what to build and judging quality — is the durable skill. Engineering background helps short-term for difficulty estimation. The PM who can wear many hats and has low ego about role boundaries will thrive.
• ✅ How should product teams handle the tension between building for current model capabilities vs. future AGI-level models?
  Be 'the right amount of AGI pilled' — build harnesses that elicit maximum value from current models while prototyping features that don't fully work yet. With each model upgrade, systematically remove scaffolding and test new capabilities.
• ✅ Why has Anthropic succeeded despite starting behind OpenAI?
  Two factors: (1) a unifying safety mission that accelerates cross-org decision-making and enables teams to sacrifice local goals for company goals, and (2) intense organizational focus.
• ✅ When should you use Claude Code vs. Desktop vs. Mobile vs. Co-work?
  Claude Code CLI for most powerful coding features; Desktop for front-end work with visual preview and a graphical interface; Mobile for kicking off tasks on the go; Co-work for non-code outputs (slides, docs, Slack, email).
• ✅ What are the costs of shipping this fast?
  Product consistency suffers — overlapping features, new users confused about best paths, users feel pressure to check Twitter daily for updates. Acknowledged as a real cost they're trying to mitigate (e.g., /powerup onboarding).
• ✅ How should knowledge workers adapt to AI tools?
  Automate repetitive tasks, invest in getting automations to 100% reliability, build tools you actually use daily rather than one-off prototypes, and avoid over-customization.

Gaps

• incompleteThe conversation is entirely from Anthropic's perspective. No competing viewpoints from OpenAI, Google, or independent practitioners are represented. The interview format means Lenny largely agrees with and amplifies Cat Wu's positions rather than stress-testing them.
  Read next: Read critical analyses of Anthropic's approach from outside observers; compare with product practices at other frontier labs.
• incompleteThe claim that 'all roles are merging' and that Anthropic hires engineers with product taste is presented without discussing the failure modes: loss of deep specialization, design consistency degradation, or accessibility problems that arise when specialists are replaced by generalists.
  Read next: Read organizational design literature on the specialist vs. generalist trade-off; look at empirical evidence from companies that eliminated specialist roles.
  closed byAdam Mosseri: AI is a tailwind for authenticity — Addresses the specialization-loss leg of that gap (not the design-consistency or accessibility legs): Mosseri names loss of deep specialization as the real risk of merging roles and gives Instagram's concrete mitigation — a generalist product-staff core with senior/creative specialists pulled in on demand, keeping designers because they carry taste, and deliberately investing in growing tomorrow's senior specialists so the pull-in model does not starve.
• incompleteThe 'just do things' philosophy and extreme speed of shipping are presented as unqualified goods. The costs (product consistency, user confusion, bugs shipped) are acknowledged but treated as acceptable trade-offs without exploring when they become unacceptable — particularly for enterprise customers who need stability.
  Read next: Look at enterprise SaaS literature on release management and the cost of change fatigue for large organizations.
• uninformedThe discussion of Co-work replacing tedious PM tasks (decks, meeting prep, email triage) lacks any engagement with the quality ceiling problem: are these AI-generated outputs actually good enough for high-stakes contexts, or do they create a mediocrity trap?
  Read next: Seek empirical studies on AI-generated content quality in professional settings; read critiques of 'good enough' AI output.
• incompleteToken spend is discussed but actual numbers are not shared. The claim that token cost per worker is 'still much lower than salary' is unverifiable and the trajectory toward salary-equivalent spend is acknowledged but not analyzed for its implications.
  Read next: Look for public data on AI compute costs per developer from companies willing to share; model the economics of AI-augmented knowledge work.
  closed bySWLW #710: Revised rules of Eng Leadership, When AI costs more than the engineer, and more. — Tunguz supplies the actual AI-spend-vs-salary numbers that gap called unverifiable (680x spread; top-1% at ~40% of an engineer's salary; median ~$137/yr) and analyzes the trajectory toward salary-equivalent spend (bull-case parity with a median employee's revenue contribution, breakeven ~2029) — with the mild unit shift from per-worker token cost to per-engineer AI bill vs. revenue contribution.
• incompleteThe Open Claude decision is framed purely as a resource allocation choice. The community trust cost and the broader ecosystem implications of an AI company restricting how its model can be accessed through third-party tools are not explored.
  Read next: Read platform economics literature on ecosystem lock-in; review community responses and long-term effects of similar decisions by other platform companies.

Closes gaps elsewhere

• Anthropic's Designs Three-Agent Harness Supports Long-Running Full-Stack AI Development · g4 — Supplies the missing when-to-change-the-harness signal: at each new model generation Anthropic removes prompting interventions and crutches no longer needed, using model introspection on its own mistakes to find what to fix (p5, p10).

Related documents

• Podcast: Tiger Teams, Evals and Agents: The New AI Engineering Playbook
• Harness engineering for coding agent users
• Harness design for long-running application development
  Agent Harness Evals Riding Model Improvements
• Don't train the model, evolve the harness
  Agent Harness Evals Riding Model Improvements
• How Anthropic Builds And How Engineering Will Change Soon | Thariq Shihipar
  Agent Harness Claude Code Taste & Judgment
• The AI paradox: More automation, more humans, more work | Dan Shipper
  Evals Near-Zero-Cost Code Riding Model Improvements
• Why cultivating agency matters more than cultivating skills in the AI era | Max Schoening (Head of Product, Notion)
  Riding Model Improvements Taste & Judgment
• ChatGPT Just Became A Work Agent
  Agent Harness Evals
• AI Is Changing the Rules of Entrepreneurship
  Near-Zero-Cost Code Taste & Judgment
• AI models can disappear overnight. Is your engineering team built to survive it?
  Agent Harness Evals