First, read the purpose and timing of this cross-role interview compilation
Product managers, engineers, and designers can use AI to participate in more implementation work, but who defines the problem, who verifies quality, and who is responsible for launch must still be made clear in each task. This article is suitable for comparing the three guests’ experiences by role and then choosing one small task to test whether it fits your own team. It is a compilation of historical interviews, not a description of Anthropic’s or OpenAI’s current organizational structure, nor a universal conclusion that role divisions should be eliminated.
Source recheck: 2026-10-25. The opening wording in the original draft, “three episodes compiled in 2026-05,” should be understood as the compilation context and cannot be treated as all three episodes being released in the same month. The release dates on Lenny’s official episode pages are respectively: Alexander Embiricos, 2025-12-14; Jenny Wen, 2026-03-01; Cat Wu, 2026-04-23. References in the text to “next year,” release time required, team headcount, and role descriptions must all be placed back at the time of each interview; predictions cannot be rewritten as facts already realized today.
This recheck covered the dates, guest introductions, and public topics on the official episode pages; it did not involve listening to the full episodes again or checking the paid full text. The existing details and quotations below are retained as the original compiled notes and are not falsely presented as verbatim transcripts from this recheck.
Read according to the work you own
- Product requirements, priorities, and prototypes, go directly to Cat Wu interview notes.
- Coding agents, work breakdown, and evaluation, go directly to Alexander Embiricos interview notes.
- Design exploration, interactive implementation, and quality judgment, go directly to Jenny Wen interview notes.
- To compare perspectives, go directly to Cross-episode comparison table; note that it is an editorial summary, not a joint statement by the three guests.
Turn inspiration into a checkable collaborative task
This site recommends: take one reversible small change; the product owner first writes the user problem and completion conditions clearly; the design owner checks normal, empty, loading, and failure states; the engineering owner checks permissions, data, and regressions; finally, the designated owner verifies the real page and final result. Record production, inspection, and rework time, and then decide whether to expand usage. Faster prototypes do not mean the work is already live, and individual release-day figures from the episodes cannot be used to promise your own team’s delivery speed; no new team-efficiency experiment is added here.
Continue reading:General division of labor for AI product team collaboration,How to accept an agent.
Official episode entry points:Cat Wu,Alexander Embiricos,Jenny Wen.
2026-05 Compilation of three Lenny Rachitsky interview episodes.Leaders of three core roles inside Anthropic + OpenAI (PM / Eng / Design)Share how AI companies really operate in 2026.
| Episode | Guests | Company / Role |
|---|---|---|
| 1. Product | Cat Wu | Anthropic, Head of Product, Claude Code |
| 2. Engineering / Codex | Alexander Embiricos | OpenAI, Head of Codex |
| 3. Design | Jenny Wen | Anthropic, Head of Design, Claude |
I. Cross-cutting theme: 3 shared understandings in 2026 AI companies
1. Role boundaries are disappearing
| Company | Manifestation |
|---|---|
| Anthropic | "PMs, engineers, and designers are merging."(Cat Wu)<br>Translation: PMs, engineers, and designers are converging<br>“Designers are now becoming more and more like PMs and engineers” (Jenny Wen) |
| OpenAI | “Boundaries blurring. PM feels more empowered — can use Codex themselves to answer questions / prototype / analyze data” (Embiricos)<br>Translation: Boundaries are blurring. PMs feel more empowered — they can use Codex themselves to answer questions / build prototypes / run data analysis |
2. Scarcity in the AI era ≠ coding ability
| Quote | Source |
|---|---|
| "As code becomes much cheaper to write, the thing that becomes more valuable is deciding what to write."<br>Translation: The cheaper writing code becomes, the more valuable “deciding what to write” becomes | Cat Wu |
| “Product taste > engineering ability”(Hires many PMs with engineering backgrounds because of taste; taste = aesthetics / taste / judgment) | Cat Wu |
| “Execution + understanding customers > the idea itself”(Execution + understanding customers > the idea itself) | Embiricos |
| "AI can weigh in, but it can't necessarily solve disputes between people about what should go into a feature."<br>Translation: AI can offer opinions, but when people disagree with each other over “what a feature should do,” AI cannot resolve it | Jenny Wen |
3. The agreed AGI bottleneck = not compute power
"The current underappreciated limiting factor is literally human typing speed or human multitasking speed." — Alexander Embiricos
Translation: The currently severely underestimated limiting factor is human typing speed, or rather, human multitasking speed.
Human speed of inputting prompts + manually verifying AI work,not GPU compute / model intelligence,that is the real AGI bottleneck.
4. The jointly recommended “product philosophy”
| Philosophy | Proposer |
|---|---|
| Just do things(Just do it) — first principles + crossing team boundaries + rejecting rigid job roles (rigid position definitions) | Cat Wu |
| Be kind and candid(Kind and candid) — directness is a form of kindness | Embiricos |
| It is what it is(That is how things are; accept it) — stay calm in the face of what cannot be controlled, without becoming dispirited | Jenny Wen |
II. Cat Wu (Anthropic Head of Product, Claude Code)
1. Evolution of the PM role (before AI vs. after)
| Dimension | Before AI | AI era (2026) |
|---|---|---|
| Time horizon | Long-term planning (6–12 months) + cross-team coordination | Ship features every week / in extreme cases, 1 day |
| Core skills | Roadmap / cross-team / PRD | Quick access to users + Taste + Eval |
| Writing PRDs | Required | Occasionally written |
| Product convergence speed | 6 months → launch | 1 month → 1-day launch |
2. Anthropic PM team structure (30–40 people)
| Team | Function |
|---|---|
| Research PM(Research PM) | Led by Diane, customer feedback + model releases |
| Cloud Developer Platform(Cloud developer platform) | API |
| Cloud Code team | Core code product |
| Enterprise(Enterprise edition) | Enterprise adoption |
| Growth(Growth) | Growth of the overall product suite |
3. Five key abilities for PMs in the AI era
- Set clear goals: Who is the core user? What is the most important problem? (Example: reduce permission prompts / permission pop-ups)
- Build repeatable shipping (release) processes: Cloud Code uses the research preview(research preview) model to lower the commitment threshold, releasing new features in 1–2 weeks
- Cross-team framework(Engineering / Marketing / Docs are tightly coupled; when a feature is ready, marketing + docs can respond immediately)
- Eval capability(model evaluation) — witness weekly changes in models
- First-principles thinking(first principles) + crossing boundaries to fill gaps
4. Anthropic’s product matrix + use cases
| Product | Applicable scenarios | Key features |
|---|---|---|
| Cloud Code terminal edition | One-off coding tasks | New features land first |
| Cloud Code desktop edition | Frontend work + task overview + non-technical users | Real-time preview of web apps |
| Cloud Code Web / Mobile | Kicking off tasks while out(can start tasks even when outside) | Can start tasks even when away from the computer |
| Cowork(Claude with Hands — Claude with hands, able to perform non-code actions) | Non-code tasks(Slack 0 / Inbox 0 / writing docs / slide decks) | Connect Google Calendar / Slack / Gmail / Drive, synthesize information → a 20-page deck (presentation) |
5. Cowork case: 20-page slide deck
Input: Connect data sources (Google Calendar + Slack + Gmail + Drive) + prompt + a narrative passage Output: Generate a 20-page draft in several hours Note: The PM still makes the final decision. Give Cowork a standardized deck so it follows the design system.
6. Impact of new models on products (2 dimensions)
| Impact | Example |
|---|---|
| Remove crutches (assistive features) | The early to-do list(task list) feature in Cloud Code (to help models do large refactors) → new models no longer need it |
| Unlock new capabilities | Code review feature — new models make it reliable enough that engineering teams rely on it before merging PRs |
7. Two success factors for Anthropic
Limited funding and limited distribution since founding—why did it succeed?
- Unified mission: Bring safe AGI to all humanity — this is cited in every decision
- Focus: Overall goals > individual product lines / teams sacrifice individual interests for the company as a whole / prioritize first-party + API (owned channels + API, even if it harms the open-source community)
8. Cloud Code source code leak incident + Open Claw decision
| Event | Explanation |
|---|---|
| Source code leak | When the package was released Human error(Passed 2 layers of manual review). The process has been hardened. The person involved is still at Anthropic |
| Open Claw tightens subscriptions | Due to high Quad demand, infra expansion + token efficiency optimization are needed.Prioritize first-party + API (own channels + API), even if it affects the open-source community |
9. PM Tool Stack
- Cloud Code(Primary tool, pushing boundaries)
- Cowork(Summarizing customer meetings / creating deck presentations)
- Slack(Company OS, operating-system-level communication tool)
- Internal self-built web app(Custom sales decks / sales demos)
- Vanta(Compliance automation platform)
10. Advice for Professionals
| Advice | Notes |
|---|---|
| Automate repetitive tasks | Let yourself focus on creativity |
| Take automation to 100% precision(Automation must be 100% accurate) | 95% has no value,the final 5–10% is the key |
| Build practical apps(daily use, use every day) | More valuable than building a prototype |
| Don’t over-customize workflows(workflows) | It will pull you away from the core task |
11. Recommended Books + Favorite Products
| Recommendation | Notes |
|---|---|
| How Asia Works(How Asia Works) | Economic development |
| The Technology Trap(The Technology Trap) | Impact of past technological revolutions on workers |
| Paper Menagerie(The Paper Zoo, a short-story collection by Ken Liu) | Short stories on AI + self-discovery |
| Drive to Survive / Free Solo(Documentaries Drive to Survive / Free Solo) | Extreme pursuit of a single point |
| Waymo(Google self-driving car, willing to pay 2× premium — willing to pay double the price) | Top choice for many people inside Anthropic |
III. Alexander Embiricos (OpenAI, Head of Codex)
1. Current State of Codex (2026-05)
| Metric | Data |
|---|---|
| Growth after GPT-5 (since August) | 20×(20×) |
| Weekly tokens served | Trillions |
| API ranking | No. 1 (most served coding model, the programming model with the largest API call volume) |
2. Codex vs Cloud Code
- Claude Code (popular since 2024) → then Codex emerged(Early Codex Cloud was difficult to set up)
- Most users now use Codex via IDE extensions / CLI (command line), with sandbox security and an interactive agent
3. Key Acceleration Factors for Codex
| Factor | Explanation |
|---|---|
| GPT 5.1 Codex Max | Compared with before30% faster at completing tasks + smarter |
| Compaction(Compression, a key innovation) | Lets the model work continuously beyond the context window(context window) — three-layer coordination across model + API + harness (execution framework) |
| Shell + Sandbox(Command line + sandbox) | Safely use the shell |
| Tightly coupled product research team | parallel build (parallel development) / internal dogfooding (using your own product first) |
4. Codex Vision — Software engineering teammate(Software engineering teammate)
"It's a bit like this really smart intern that refuses to read Slack, doesn't check data dog unless you ask it to."
Translation: It’s a bit like a very smart intern who refuses to check Slack messages and won’t look at monitoring unless you tell them to.
Now: smart intern, needs guidance Future:Proactive / collaborative partner across the entire development cycle (ideation → maintenance, from brainstorming ideas to maintenance)
5. Real Acceleration Cases (Organizational Level)
| Project | Data |
|---|---|
| Sora Android App(Android version of OpenAI’s video-generation app) | 18 days to complete the internal version,28 days public release, launched on Google Play |
| Atlas Browser(OpenAI browser) | Originally 2–3 weeks / 2–3 engineers → 1 week / 1 engineer |
6. Impact on 3 Roles
| Role | Impact |
|---|---|
| Engineers | Improved efficiency in writing complex code + reduced burden from Code review/validation + from manual coding → higher-level thinking(from hand-writing code → higher-level thinking) |
| PM | Crossing communication boundaries, answering questions yourself / prototypes / data analysis / compressed Talent stack (talent structure) |
| Designer | Directly build animations + app prototypes → quickly go from idea to product |
7. Evolution of the 3 Abstraction Layers of AI Coding
Current: Coding (writing code)
↓
Mid-term: Spec-driven / Chatter-driven (specification-driven / conversation-driven) development
↓
Future: Higher-level + Auto-completion + Contextual help (higher-level + autocomplete + context-sensitive help, embedded in IDE/browser)
But:not everyone is used to writing specs (specification documents),The transition will be gradual.
8. AGI Limiting Factor (Most Important Insight in This Podcast)
"The current underappreciated limiting factor is literally human typing speed or human multitasking speed, unlike writing prompts."
Translation: The currently severely underrated limiting factor is human typing speed, or multitasking speed, not the part about writing prompts.
Humans need to:
- write prompts
- Manually verify AI work(code review)
- Multitask management
Outlook:
- Next year(2026) — early adopters see agents become self-sufficient
- In a few years — large companies follow
- Reach AGI — when productivity growth feeds back to AI labs (when productivity growth feeds back into AI labs, triggering recursive acceleration)
9. Practical Advice for Using Codex
| Advice | Details |
|---|---|
| Give it the hardest tasks | Don’t start with simple tasks; instead,give it the hardest bug |
| Build trust | First let it understand the codebase + write a plan → gradually increase tasks |
| Try in parallel | Run multiple tasks at the same time |
| Treat it like a new employee | Build trust gradually |
10. Codex Team Recruiting
Hiring: Engineers / Salespeople / Product Requirements: technical ability + opinions on the future of coding agents
11. Recommended Books + Favorite Products
| Recommendation | Notes |
|---|---|
| The Culture (Ian Banks)(Culture series, Iain Banks) | Optimistic AI future / "gay space communist Utopia" (the author’s semi-joking label) |
| Fire Upon the Deep(A Fire Upon the Deep) | Superintelligent sci-fi space opera |
| Jujutsu Kaisen anime | Righteous protagonist + helping others |
| Tesla self-driving | Mixed-initiative design(Mixed-initiative design, human empowerment + assistance, both empowering people and providing assistance) |
12. Key Product Element — Mixed-Initiative
"Tesla self-driving is a good example of mixed-initiative software — empowering humans while providing assistance."
Translation: Tesla self-driving is a good example of "mixed-initiative software" — it both empowers humans and provides assistance, rather than fully replacing them.
Codex’s design is moving in the same direction — not autonomy max (maximum full autonomy), but humans + AI synergy.
IV. Jenny Wen (Anthropic, Head of Design, Claude)
1. Core Claim:The Design Process is Dead(The design process is dead)
"The design process that designers have been taught and treated as gospel is basically dead. In the age of AI, designers have to let go of the old process."
Translation: The “design process” that designers have been taught and treated as gospel is basically dead. In the AI era, designers must let go of the old process.
3 Characteristics of the Old Process:
- Extensive research / discovery
- Fixed divergence-convergence cycle
- Beautiful long-term decks
Why must it be abandoned?
- Engineers can now quickly build features(quickly build features, relying on Claude Code)
- Designers blocking engineers = unnecessary bottleneck
- The vision timeline has shortened to 3–6 months(prototypes rather than long-term decks)
2. Dramatic Shift in Designers’ Time Allocation
| Task | A few years ago | Now (2026) |
|---|---|---|
| Mocking / Prototyping | 60-70% | 30-40% |
| Collaborate directly with engineers | 20% | 30-40% |
| Implement / Polish in code | Almost 0 | 20-30% |
| Cross-team coordination | 10% | 10% |
3. Two Layers of New Design Work
| Type | Description |
|---|---|
| Support implementation + execution | No longer making polished mocks (engineers can quickly produce them) — directly help engineers polish + guide direction |
| Create short-term vision(short-term vision) | 3–6 months ahead — in the form of a prototype, not a long-term deck |
4. Three Types of Valuable Designer Prototypes
| Prototype | Characteristics |
|---|---|
| Strong Generalist(Strong Generalist) | Strong across multiple core skills(similar to T-shape but more "block-shaped") — easily adapts as the boundaries of PM + Eng expand |
| Deep Specialist(Deep Specialist) | Deeply focused on one domain (software engineering / visual design / icon design) —differentiated product |
| Craft New Grad(Craft New Grad) | Early-career,blank slate + eager to learn — not constrained by traditional processes, brings fresh perspectives |
5. Priority Traits for Designer Hiring
- Resilience(Resilience) + Adaptability
- Build actual things and share them(make tangible things and share them — has a documented portfolio of projects, track record of projects)
- Excited by technology + willing to work at the frontier of design | Build tangible things within 4–6 months
6. Advice for New Designers
| Advice | Notes |
|---|---|
| Build actual things and share them(Build tangible things and share them) | More important than a portfolio |
| Learn coding tools(Learn programming tools) | No need to write code from scratch,but must know how to use coding tools |
7. Limitations of AI in Design (Honest Assessment)
| Capability | Current state (2026) |
|---|---|
| First-pass ideas(Initial draft ideas) | ✅ AI can provide them |
| Taste + Judgment(Taste + judgment) | 🟡 Improving |
| Replacing Strong Generalist / Deep Specialist / Craft New Grad(replacing the three types of designers above) | ❌ Not yet possible |
| Decision-making + responsibility | ❌ Humans must sign off(make the final call) |
8. The Future of AI UI
"Chatbots and terminals are not the final UI. It will be a combination of chatbots + tactile UIs + model-generated UIs."
Translation: Neither chatbots nor command-line terminals are the final user interface form. The final form will be a combination of "chat + tactile interfaces," plus interfaces generated in real time by models.
- Chatbots(Chatbots) — flexible
- Tactile UIs(Tactile traditional interfaces) — efficient for specific tasks
- Model-Generated UIs(Model-generated interfaces) — will increase in the future (replacing hand-coded interfaces)
9. Cowork (Claude with Hands) Product Story
| Stage | Duration |
|---|---|
| Multiple versions of internal prototypes | Long-term exploration |
| Internal → external launch | 10 days |
Killer use cases for Cowork (Jenny personally):
- Introspection(Introspection) — analyze years of personal notes to gain self-insight
- Hiring Rubric(Interview scorecard) — read past design craft notes →generate a rubric for evaluating interview candidates
10. Design Management Philosophy
Switching between IC and Management (IC = Individual Contributor, frontline practitioner / Management = manager)
Over the past year, Jenny moved from manager → IC work, gaining hard skills + empathy(empathy).May return to management in the future.
Advice for design managers:Should periodically return to an IC role — to regain the feel for the work.
Low-leverage tasks = high-leverage tasks (counterintuitive)
"Some of the best traits of leaders and managers are that they choose low-leverage tasks, which actually end up being very high-leverage because it's them who's doing it."
Translation: One of the best traits of good leaders and good managers is — they choose to do "low-leverage" tasks, and precisely because they are the ones doing them, those tasks become "high-leverage."
Examples(Examples):
- Testing the product
- Reproducing bugs
- Making personalized cards for team members personalized cards(commemoration days / anniversaries)
Why it is high leverage: it shows the boss genuinely cares + genuinely uses the product.
Roasting culture (psychological safety + high standards)
| Element | Description |
|---|---|
| Roasting(Roasting / mutual teasing) | Team members can "roast" one another + roast the manager → psychological safety |
| Balance | + High standards(high standards, similar to Radical Candor — frank and direct) |
| Core | Care deeply + Challenge directly |
11. Legibility Framework (understandability — whether an idea is "immediately understandable," 2×2 matrix)
VC (venture capitalist) perspective:
- Eligible (qualified / worth investing in) = at the frontier,hard to understand(illegible) →high value for VC + Designer(because others dismiss it, others overlook it)
Example: Cloud Studio prototype = an illegible idea, but it led to the birth of useful concepts.
Three patterns applicable to early successful companies: focus on exciting but unclear ideas.
12. Recommended Books + Favorite Products
| Recommendation | Notes |
|---|---|
| The Power Broker (Robert Caro)(The Power Broker by Robert Caro) | Long but worth finishing — insights on long-term careers + getting things done |
| Insomniac City (Bill Hayes)(Insomniac City by Bill Hayes) | A memoir of Oliver Sacks’s final days — death / love / life |
| Sentimental Value (Norwegian film, Chinese title The Value of Emotion) | Same director as The Worst Person in the World |
| The Pit Season(TV series The Kingdom) | Watching highly competitive people do their work |
| Retro (community photo-sharing app) | Used for 2 years / no social media counts (likes) / no ads /look back on past weeks |
| It is what it is(Motto: "It is what it is; accept it") | The current world / industry is uncontrollable,move forward with calm(understood its depth more after a 10-day meditation retreat) |
V. Cross-Episode Comparison Table (Who Said What)
| Topic | Cat Wu | Embiricos | Jenny Wen |
|---|---|---|---|
| Role boundaries | PM/Eng/Designer are converging | Boundaries blurring | Designers spend 30–40% more time in code |
| Decision-making core | Taste + Goals | Execution > Ideas | Human accountability |
| AGI timeline | Not stated explicitly | Next year early adopters + a few years for large companies | Not stated explicitly |
| Progress of AI writing code | Claude Code ships in 1–2 weeks | Sora 18 days / Atlas 1 week | Cowork 10 days |
| Advice for new PMs/Designers | Automate repetitive tasks + practical apps | Build trust for the hardest tasks | Build real things + Share |
| Life philosophy | Just do things | Be kind and candid | It is what it is |
VI. Lenny’s Own Core Observations (Running Through the Questions)
- "Where will human brains stay valuable as AI gets smarter?"(The smarter AI becomes, where is human value?) — all three guests emphasized:choosing projects / evaluating quality / common sense / EQ / decision-making and responsibility
- "How do you keep up with the speed?"(How to keep up with this speed?) — accept imperfection + hire senior people to maintain energy + prioritize
- "Career path with role blurring?"(After role boundaries blur, what should career paths look like?) — sacrifice consistency / user education costs rise / use "Slash Power Up" to guide users
VII. Actionable Takeaways for KB Users
- PM perspective: Use Claude Code + Cowork to build prototypes and do analysis yourself → no longer dependent on engineers
- Engineer perspective: Give the hardest bugs to Codex / build EVA (eval) capabilities / remove crutches after models become smarter
- Designer perspective:Enter the code layer + short-term 3–6 month vision + Build & Share
- Management perspective: Do "low-leverage" tasks + roasting culture + periodically switch between IC (frontline individual contributor) ↔ Manager
- Personal perspective:The hardest constraint is not AI intelligence, but your input speed → learn efficient input / choose mixed-initiative tools
See also
- Karpathy · Software in the era of AI(Software 3.0) — Karpathy Software 3.0 (common theme: AI rewrites software infrastructure)
- Claude Code Best Practices (Cal Rueb · Anthropic Internal Talk) — Cal Rueb of Anthropic’s engineer perspective on Claude Code
- AI Agent Technical Architecture Evolution 2025 (In-Depth Research Report) — In-depth look at Agent technical architecture
- Wang Kai’s Claude Code Terminal Workflow Methodology (Complete Version) — Wang Kai’s Complete Claude Code Workflow (Technical Approach)
- newtype · AI Coding in Practice and Toolstack Evolution — newtype AI Coding Toolstack Evolution
- newtype Observations on AI Business, Industry, and Investment — AI Business Trends + 5 Levels of AI + LLM=OS
- Xiaohongshu / Vibecoding Hands-On Collection (Games/Tools/Mini Programs) — Xiaohongshu Vibecoding in Practice (Comparison from an Independent Developer Perspective)
- Xiaohongshu / Vibecoding Aesthetics + Creator Economy Collection — “The Model Doesn’t Matter, Aesthetics Do” (Echoing Cat Wu’s Theory of Taste)
- Yitang / Wang Huiwen’s Tsinghua Product Course: Complete 66-Page Notes — Wang Huiwen’s Tsinghua Course (Comparison from a Traditional PM Perspective)